PASS — the laptop's 5090, posture DYNAMIC-BOOST (the shipped posture) started 2026-09-21T20:23:53Z on this laptop harness /workshop/bench-laptop-5090-2026-09-21/harness-llm render /workshop/bench-laptop-5090-2026-09-21/harness-render log /workshop/bench-laptop-5090-2026-09-21/logs/rung-dynboost.log ⚠ this pass has NO CAP. The board refuses -pl, so there is nothing to set and nothing to wait for: nvidia-powerd floats the enforced limit between this board's default (95 W) and its maximum (175 W) against the CPU's draw. Every figure's cap cell is a RANGE out of the 2 Hz trace, and a single wattage typed into one would be a fabrication. The per-stage ranges land in results/*-cap-posture-dynamic-boost-*.json. ⚠ the clocks run UNLOCKED here, exactly as in the 95 W rung: the operator released the boot-time -lgc in Step 1 and re-applies it only in the runsheet's LAST step. The verdict is read back FROM THE CARD at this pass's open and written into its receipt. receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. envelope at this pass's open: 140 95 175 W (enforced, default, max) memory guard: memshed.timer was active at the window's open (remembered in /workshop/bench-laptop-5090-2026-09-21/.memshed-was) memory guard: memshed.timer is now inactive; MEMSHED_DRY_RUN=1 (STOPPED, never masked: the unit file is a symlink into ~/estate/systemd and 'mask --force' would destroy it. glimmer-check.timer is left alone — it next fires Tue 14:43Z, outside any daytime window, and masking a timer this bench does not need to touch is one more thing to forget to restore.) ================================================================================ PASS dynboost — posture dynamic-boost — LEG 1 (the eleven-arm language bank), then LEG 2 card GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff (NVIDIA GeForce RTX 5090 Laptop GPU) class gpu-5090-laptop-24g cap arg dynboost start 2026-09-21T20:23:53Z ================================================================================ receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. envelope at the pass's open: enforced/default/max = 140 95 175 W the enforced limit FLOATS in this posture. Its cell is a range from the 2 Hz trace. clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock-lock receipt -> /workshop/bench-laptop-5090-2026-09-21/logs/clock-lock-dynboost.txt ---- LEG 1: /workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh dynboost ---- card name read back: 'NVIDIA GeForce RTX 5090 Laptop GPU' (arm one; exact equality, not a substring) === the board, before anything is started (this is the dynamic-boost block) === 20:23:59Z index, uuid, name, power.limit [W], enforced.power.limit [W], power.default_limit [W], power.max_limit [W], persistence_mode, pcie.link.width.current, clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz], temperature.gpu, fan.speed [%] 0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 140.00 W, 95.00 W, 175.00 W, Enabled, 8, 180 MHz, 405 MHz, 3090 MHz, 32, [N/A] receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. posture [block dynboost open]: dynamic-boost — enforced/default/max = 140 95 175 W the enforced limit FLOATS here. Its cell is a RANGE from the 2 Hz trace, never a number. memory guard: memshed.timer is inactive; MEMSHED_DRY_RUN=1 === the UPS at the door (expected: absent on this box) === 20:23:59Z UPS UNREADABLE at block start -- whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason, and board watts are read normally: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (read 2026-09-21T14:52Z). Whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason; board watts, which is what every table in this ladder prints, are read normally from the card at 2 Hz. receipt -> results/gpu-5090-laptop-24g-ups-gate-dynboost.txt clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 172 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 172, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock-lock receipt -> results/gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-dynboost.txt === instance: one card, kv q8_0, parallel 1 === 20:24:07Z "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 8;" } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [card-mapping open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [card-mapping] open: 1612,14001,3090 (sm,mem,max.sm MHz) cap witness [card-mapping]: every 5s -> results/cap-witness-dynamic-boost-card-mapping.csv (pid 1243808) === card mapping: watch which physical card's memory rises === 20:24:49Z one: MATCHES: declared soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c, and that is the card whose memory grew every shape checked runs on the card it declares -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-card-mapping-dynboost.json clocks at [card-mapping] close: 1597,14001,3090 (sm,mem,max.sm MHz) posture [card-mapping close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=2 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-card-mapping.json posture [armA-ladder-gemma4 open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA-ladder-gemma4] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armA-ladder-gemma4]: every 5s -> results/cap-witness-dynamic-boost-armA-ladder-gemma4.csv (pid 1244285) === armA-ladder-gemma4 === 20:24:57Z contention gate: busy: busiest card mean 31.9% > 5.0% -- waiting (attempt 1/6) run 2: decode=156.554 tok/s ttft=366.82 ms W=144.67 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19460.0} contention gate: busy: busiest card mean 45.0% > 5.0% -- waiting (attempt 1/6) run 3: decode=155.899 tok/s ttft=359.07 ms W=143.65 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19460.0} --- num_ctx 98,304 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19960.0}; runner RSS 1.48 GiB run 1: decode=155.055 tok/s ttft=405.36 ms W=115.38 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19960.0} run 2: decode=155.493 tok/s ttft=361.42 ms W=134.94 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19960.0} run 3: decode=155.671 tok/s ttft=365.79 ms W=136.28 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19960.0} --- num_ctx 131,072 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20460.0}; runner RSS 1.52 GiB run 1: decode=155.244 tok/s ttft=367.5 ms W=118.63 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20460.0} run 2: decode=155.387 tok/s ttft=393.72 ms W=120.13 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20460.0} contention gate: busy: busiest card mean 6.9% > 5.0% -- waiting (attempt 1/6) run 3: decode=155.919 tok/s ttft=381.17 ms W=145.56 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20460.0} the biggest context this arm holds for gemma4:26b is 131,072 tokens -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-auto-ladder.json clocks at [armA-ladder-gemma4] close: 1845,14001,3090 (sm,mem,max.sm MHz) posture [armA-ladder-gemma4 close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=104 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=104) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA-ladder-gemma4.json arm A's headline rung at dynamic-boost, read from its own result file: num_ctx 131072 posture [armA2-kv-axis open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA2-kv-axis] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armA2-kv-axis]: every 5s -> results/cap-witness-dynamic-boost-armA2-kv-axis.csv (pid 1261262) === A2 instance kv=f16 === 20:33:40Z } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvf16-p1.json to every bencher so the result files carry it === armA2-kvf16-at131072 === 20:33:47Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=17477150965 size_vram=17477150965 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20693.0} contention gate: busy: busiest card mean 5.3% > 5.0% -- waiting (attempt 1/6) run 1: decode=165.893 tok/s (wall 164.854) ttft=458.97 ms W=149.12 J/1k=898.89 temp=44.0C fan=None% clk=2107.0-2257.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 106.5% of the 140 W cap, at 44 C -- well below any throttling temperature contention gate: busy: busiest card mean 18.8% > 5.0% -- waiting (attempt 1/6) run 2: decode=165.912 tok/s (wall 164.955) ttft=423.96 ms W=141.23 J/1k=851.23 temp=44.0C fan=None% clk=2145.0-2265.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 100.9% of the 140 W cap, at 44 C -- well below any throttling temperature contention gate: busy: busiest card mean 13.9% > 5.0% -- waiting (attempt 1/6) run 3: decode=165.702 tok/s (wall 164.784) ttft=419.35 ms W=142.02 J/1k=857.08 temp=44.0C fan=None% clk=2152.0-2265.0MHz mem=14001.0MHz no clock drop during decode (draw held at 101.4% of the 140 W cap) spread gate: spread 0.1% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-kvf16-at131072.json wrote 1 file(s) kv=f16 holds num_ctx 131072 at dynamic-boost -- that is this KV type's largest window === A2 instance kv=q4_0 === 20:35:14Z } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq4_0-p1.json to every bencher so the result files carry it === armA2-kvq4_0-at131072 === 20:35:18Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=17747851344 size_vram=17747851344 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19113.0} run 1: decode=156.783 tok/s (wall 155.886) ttft=354.28 ms W=137.79 J/1k=878.86 temp=45.0C fan=None% clk=2122.0-2362.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 91.9% of the 150 W cap, at 45 C -- well below any throttling temperature contention gate: busy: busiest card mean 32.2% > 5.0% -- waiting (attempt 1/6) run 2: decode=156.164 tok/s (wall 155.253) ttft=313.72 ms W=150.01 J/1k=960.59 temp=44.0C fan=None% clk=1275.0-2385.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 107.1% of the 140 W cap, at 44 C -- well below any throttling temperature run 3: decode=155.891 tok/s (wall 155.025) ttft=343.36 ms W=146.68 J/1k=940.91 temp=44.0C fan=None% clk=2145.0-2355.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 97.8% of the 150 W cap, at 44 C -- well below any throttling temperature spread gate: spread 0.6% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-kvq4_0-at131072.json wrote 1 file(s) kv=q4_0 holds num_ctx 131072 at dynamic-boost -- that is this KV type's largest window clocks at [armA2-kv-axis] close: 1762,14001,3090 (sm,mem,max.sm MHz) posture [armA2-kv-axis close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=34 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=34) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA2-kv-axis.json === restore the q8_0 instance (the seat posture) === 20:36:28Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [armA3-filled open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA3-filled] open: 1725,14001,3090 (sm,mem,max.sm MHz) cap witness [armA3-filled]: every 5s -> results/cap-witness-dynamic-boost-armA3-filled.csv (pid 1267598) === armA3-filled === 20:36:32Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. ### gpu-5090-laptop-24g m5 gemma4:26b -- FILLED WINDOW at num_ctx 131,072 (75% target = 98,304 tokens) ### fit at the window: fits: fully resident on the card per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20460.0} filler round 1: 183 copies -> 97992 tokens (target 98304) window filled: 97,992 tokens of 131,072 (74.8% occupied) COLD prefill: 2606.542 tok/s over 98025 tokens, ttft 38735.76 ms contention gate: busy: busiest card mean 28.2% > 5.0% -- waiting (attempt 1/6) run 1: prefill=1202887.165 tok/s (97992 tokens) ttft=793.31 ms decode=71.616 tok/s W=119.93 contention gate: busy: busiest card mean 14.1% > 5.0% -- waiting (attempt 1/6) run 2: prefill=1234638.209 tok/s (97992 tokens) ttft=770.02 ms decode=71.95 tok/s W=124.01 run 3: prefill=1397450.158 tok/s (97992 tokens) ttft=730.13 ms decode=72.002 tok/s W=114.3 prompt tokens: 97992 of 131072 (0.7476) prefill: 1234638.209 tok/s ttft: 770.02 ms decode at depth: 71.95 tok/s -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-filled.json clocks at [armA3-filled] close: 1597,14001,3090 (sm,mem,max.sm MHz) posture [armA3-filled close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=32 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=32) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA3-filled.json posture [armD-ladder-mistral open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armD-ladder-mistral] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armD-ladder-mistral]: every 5s -> results/cap-witness-dynamic-boost-armD-ladder-mistral.csv (pid 1279983) === armD-ladder-mistral === 20:39:12Z per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19480.0}; runner RSS 0.87 GiB run 1: decode=44.815 tok/s ttft=192.34 ms W=134.94 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19480.0} run 2: decode=44.786 tok/s ttft=212.06 ms W=135.42 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19480.0} contention gate: busy: busiest card mean 14.8% > 5.0% -- waiting (attempt 1/6) run 3: decode=44.784 tok/s ttft=213.73 ms W=141.18 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19480.0} --- num_ctx 65,536 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 1440.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20920.0}; runner RSS 0.89 GiB run 1: decode=44.769 tok/s ttft=187.04 ms W=130.42 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20920.0} contention gate: busy: busiest card mean 13.7% > 5.0% -- waiting (attempt 1/6) run 2: decode=44.737 tok/s ttft=195.63 ms W=141.04 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20920.0} contention gate: busy: busiest card mean 9.9% > 5.0% -- waiting (attempt 1/6) run 3: decode=44.79 tok/s ttft=178.63 ms W=140.74 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20920.0} --- num_ctx 98,304 --- does not fit: only part of the model is on the card (89.9% VRAM / 10.1% RAM); per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 2108.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 23028.0}; runner RSS 14.37 GiB ladder stops at num_ctx 98,304: does not fit: only part of the model is on the card (89.9% VRAM / 10.1% RAM) the biggest context this arm holds for mistral-small3.2:24b is 65,536 tokens -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-auto-ladder.json clocks at [armD-ladder-mistral] close: 180,9001,3090 (sm,mem,max.sm MHz) posture [armD-ladder-mistral close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=93 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=93) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armD-ladder-mistral.json posture [armD-forced-probe open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armD-forced-probe] open: 180,9001,3090 (sm,mem,max.sm MHz) cap witness [armD-forced-probe]: every 5s -> results/cap-witness-dynamic-boost-armD-forced-probe.csv (pid 1295874) === D-probe: the planner SPILLED at num_ctx 98304 -- re-running it forced === 20:46:59Z === armD-forced-at98304 === 20:46:59Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=23109042175 size_vram=23109042175 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 23167.0} run 1: decode=44.827 tok/s (wall 44.64) ttft=156.47 ms W=135.55 J/1k=3023.84 temp=48.0C fan=None% clk=1642.0-1747.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 90.4% of the 150 W cap, at 48 C -- well below any throttling temperature contention gate: busy: busiest card mean 9.9% > 5.0% -- waiting (attempt 1/6) run 2: decode=44.763 tok/s (wall 44.578) ttft=212.04 ms W=140.84 J/1k=3146.32 temp=47.0C fan=None% clk=1627.0-1657.0MHz mem=14001.0MHz no clock drop during decode (draw held at 100.6% of the 140 W cap) contention gate: busy: busiest card mean 19.8% > 5.0% -- waiting (attempt 1/6) run 3: decode=44.681 tok/s (wall 44.494) ttft=201.2 ms W=140.87 J/1k=3152.82 temp=46.0C fan=None% clk=1642.0-1657.0MHz mem=14001.0MHz no clock drop during decode (draw held at 100.6% of the 140 W cap) spread gate: spread 0.3% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-forced-at98304.json wrote 1 file(s) === D-probe: the forced rung FIT -- climbing one more to find the first that does not === 20:48:30Z === armD-forced-at131072 === 20:48:30Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=None size_vram=None -> refused: /api/ps reported no size for this model per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 0.0} -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-forced-at131072.json wrote 1 file(s) clocks at [armD-forced-probe] close: 1597,14001,3090 (sm,mem,max.sm MHz) posture [armD-forced-probe close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 150.0 W, n=21 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=21) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armD-forced-probe.json posture [armE-concurrency open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armE-concurrency] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armE-concurrency]: every 5s -> results/cap-witness-dynamic-boost-armE-concurrency.csv (pid 1299458) === E instance: parallel 4 === 20:48:40Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p4.json to every bencher so the result files carry it === E: concurrency 1 / 2 / 4 at dynamic-boost === 20:49:14Z per stream: [155.613] ### concurrency level 2 ### level 2: sum 255.022 tok/s, wall 203.219 tok/s over 2.5194s, W=144.35 temp=45.0C fan=None% per stream: [123.505, 131.517] level 2: sum 258.548 tok/s, wall 206.658 tok/s over 2.4775s, W=118.57 temp=45.0C fan=None% per stream: [129.244, 129.304] level 2: sum 258.685 tok/s, wall 205.79 tok/s over 2.488s, W=115.88 temp=46.0C fan=None% per stream: [129.375, 129.31] ### concurrency level 4 ### level 4: sum 348.495 tok/s, wall 287.911 tok/s over 3.5566s, W=115.34 temp=46.0C fan=None% per stream: [87.795, 87.758, 87.734, 85.208] level 4: sum 354.73 tok/s, wall 280.081 tok/s over 3.6561s, W=129.38 temp=47.0C fan=None% per stream: [88.726, 88.639, 88.696, 88.669] level 4: sum 351.71 tok/s, wall 283.478 tok/s over 3.6123s, W=136.15 temp=47.0C fan=None% per stream: [87.921, 87.944, 87.923, 87.922] one-dynboost takes 4 parallel streams at 351.7 tok/s aggregate (sum of streams) / 283.5 tok/s over wall -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-concurrency-one-dynboost.json clocks at [armE-concurrency] close: 1702,14001,3090 (sm,mem,max.sm MHz) posture [armE-concurrency close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=34 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=34) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armE-concurrency.json posture [g2-doorman open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g2-doorman] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [g2-doorman]: every 5s -> results/cap-witness-dynamic-boost-g2-doorman.csv (pid 1304711) === G2: the doorman on mistral at num_ctx 65536 (1 caller, then 4) === 20:51:30Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. fit: SPLIT and scored as one: 54.0% VRAM / 46.0% RAM contention gate: busy: busiest card mean 26.4% > 5.0% -- waiting (attempt 1/6) concurrency 1: 10 calls, median 470.05 ms, p95 488.97 ms, max 488.97 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 31.22} concurrency 4: 8 calls, median 922.97 ms, p95 4739.1 ms, max 4739.1 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 70.33} a doorman call on mistral-small3.2:24b costs 470.05 ms at the median and 488.97 ms at p95, one caller at a time -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g2-doorman-dynboost.json clocks at [g2-doorman] close: 1590,14001,3090 (sm,mem,max.sm MHz) posture [g2-doorman close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=12 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g2-doorman.json === G3/G4 instance: back to parallel 1 === 20:52:27Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [g3-vision open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g3-vision] open: 1695,14001,3090 (sm,mem,max.sm MHz) cap witness [g3-vision]: every 5s -> results/cap-witness-dynamic-boost-g3-vision.csv (pid 1308198) === G3: a vision seat (TEXT path only -- no image is read or sent by this bench) === 20:53:17Z === g3-vision === 20:53:17Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g g3 minicpm-v4.5:latest -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=5596931685 size_vram=5596931685 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 6693.0} contention gate: busy: busiest card mean 9.7% > 5.0% -- waiting (attempt 1/6) run 1: decode=122.173 tok/s (wall 121.003) ttft=159.3 ms W=142.23 J/1k=1164.17 temp=42.0C fan=None% clk=180.0-1860.0MHz mem=405.0MHz NEITHER cap: the SM clock varied with draw at 85.2% of the cap and the card at 42 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it run 2: decode=120.945 tok/s (wall 119.718) ttft=140.25 ms W=59.67 J/1k=493.37 temp=42.0C fan=None% clk=1590.0-1807.0MHz mem=14001.0MHz NEITHER cap: the SM clock varied with draw at 39.8% of the cap and the card at 42 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it run 3: decode=120.957 tok/s (wall 119.73) ttft=137.15 ms W=94.72 J/1k=783.09 temp=42.0C fan=None% clk=1800.0-1800.0MHz mem=14001.0MHz no clock drop during decode spread gate: spread 1.0% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g3-one-dynboost-g3-vision.json wrote 1 file(s) clocks at [g3-vision] close: 1762,14001,3090 (sm,mem,max.sm MHz) posture [g3-vision close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=14 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=14) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g3-vision.json posture [g4-embed open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g4-embed] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [g4-embed]: every 5s -> results/cap-witness-dynamic-boost-g4-embed.csv (pid 1310319) === G4: embeddings === 20:54:23Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. keep-alive: bench instance: held for the arm's duration batch 1: 235.17 ms, 272.147 texts/s, W=22.54 batch 2: 228.16 ms, 280.504 texts/s, W=22.65 batch 3: 233.01 ms, 274.663 texts/s, W=22.63 batch 4: 237.08 ms, 269.952 texts/s, W=22.62 batch 5: 237.55 ms, 269.421 texts/s, W=22.47 one-dynboost: 272.1 texts/s on a batch of 64 (235.17 ms per batch) -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-embed-one-dynboost.json clocks at [g4-embed] close: 2310,14001,3090 (sm,mem,max.sm MHz) posture [g4-embed close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=5 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g4-embed.json posture [idle open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [idle] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [idle]: every 5s -> results/cap-witness-dynamic-boost-idle.csv (pid 1311172) === idle receipt at dynamic-boost === 20:54:46Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. empty: no model on the cards cards empty, nothing loaded: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 7.02 W | pair 7.02 W | UPS whole box None W (None %) loading gemma4:26b (num_ctx=131072, num_gpu=None) fit: fits: fully resident on the card model resident, not generating: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 10.98 W | pair 10.98 W | UPS whole box None W (None %) with gemma4:26b resident and not generating, the card draws 10.98 W (board power), against 7.02 W with the cards empty -- a resident cost of 3.96 W -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-idle-gemma4-dynboost.json clocks at [idle] close: 187,14001,3090 (sm,mem,max.sm MHz) posture [idle close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 140 W, max 150 W, mean 144.0 W, n=18 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 140, n=18) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-idle.json === dynamic-boost block finished === 20:56:12Z stopping bench-5090laptop-cpu-box --- receipt: units --- --- receipt: ports --- no bench port listening (1147[01]) -- clean index, uuid, power.limit [W], enforced.power.limit [W], persistence_mode, temperature.gpu, fan.speed [%], memory.used [MiB], clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz] 0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, [N/A], 150.00 W, Enabled, 33, [N/A], 633 MiB, 187 MHz, 14001 MHz, 3090 MHz posture [block dynboost close]: dynamic-boost — enforced/default/max = 150 95 175 W (the envelope at the block's open was 140 95 175 W. The enforced limit moving between the two IS the measurement; the DEFAULT or the MAX moving would not be, and cap_stage_close_posture refuses on that at every stage boundary.) the UPS is absent on this box -- see results/gpu-5090-laptop-24g-ups-gate-dynboost.txt for the withheld-figure receipt written at the block's open. Board watts are the measured quantity in every table this bench fills. LEG 1 at dynboost exited 0 envelope between the legs: 150 95 175 W ---- LEG 2: /workshop/bench-laptop-5090-2026-09-21/harness-render/run_rung.sh dynboost ---- == render rung dynboost · started 2026-09-21T20:56:13Z == card expected NVIDIA GeForce RTX 5090 Laptop GPU comfy root /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1 comfy python /workshop/ComfyUI/.venv/bin/python comfy version 0.21.1 (26515acd) receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. card reads 0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 150.00 W, 24463 MiB, 8, 16, 817 MHz, 810 MHz posture dynamic-boost — enforced/default/max = 150 95 175 W. The enforced limit FLOATS while these arms render; every result file's trace reduces it to min/median/max and THAT is the cap cell. clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 817, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock lock unprobed by the CARD (when the lock is in force this board cannot clock below 1200 MHz and a power posture means something different from what it means on the unlocked desktop cards this leg compares to). The unit state is not evidence; see above. port band 18190-18199 free; this box's own comfyui.service is inactive -- arm inventory · suffix inventory-dynboost · 2026-09-21T20:56:19Z -> results/inventory-inventory-dynboost.json -- arm inventory exited 0 · 2026-09-21T20:56:19Z posture [after inventory at dynboost]: enforced/default/max = 150 95 175 W -- arm r1 · suffix r1-dynboost · 2026-09-21T20:56:19Z Traceback (most recent call last): File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1571, in main() ~~~~^^ File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1563, in main payload = arm_r1(args.graphs.split(","), batches, args.images, name) File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1200, in arm_r1 "preflight": box_preflight(), "graphs": {}, "refusals": []} ~~~~~~~~~~~~~^^ File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1055, in box_preflight free = assert_card_available() File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1035, in assert_card_available raise CardBusy("a process this bench does not know about holds the card, refusing to " f"start: {apps['unexpected']}") CardBusy: a process this bench does not know about holds the card, refusing to start: [{'pid': 1172409, 'process_name': '/usr/local/lib/ollama/llama-server', 'used_mib': 610.0, 'gpu_uuid': 'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff'}] -- arm r1 exited 1 · 2026-09-21T20:56:19Z posture [between R1 and R4 at dynboost]: enforced/default/max = 150 95 175 W -- arm r4 · suffix r4-dynboost · 2026-09-21T20:56:19Z Traceback (most recent call last): File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1571, in main() ~~~~^^ File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1565, in main payload = arm_r4(args.graph, batches[0], args.minutes, name) File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1345, in arm_r4 "preflight": box_preflight(), "cards": {}} ~~~~~~~~~~~~~^^ File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1055, in box_preflight free = assert_card_available() File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1035, in assert_card_available raise CardBusy("a process this bench does not know about holds the card, refusing to " f"start: {apps['unexpected']}") CardBusy: a process this bench does not know about holds the card, refusing to start: [{'pid': 1172409, 'process_name': '/usr/local/lib/ollama/llama-server', 'used_mib': 610.0, 'gpu_uuid': 'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff'}] -- arm r4 exited 1 · 2026-09-21T20:56:19Z posture [render rung dynboost close]: enforced/default/max = 150 95 175 W == render rung dynboost · done 2026-09-21T20:56:19Z (R1 exit 1, R4 exit 1) == results/r1-dynboost.json and results/r4-dynboost.json the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally. the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it. LEG 2 at dynboost exited 1 envelope at the pass's close: 150 95 175 W artifact receipt -> /workshop/bench-laptop-5090-2026-09-21/RUNG-DYNBOOST-FAILED.txt PASS dynboost — FAILED (rc=1) posture dynamic-boost cap argument dynboost written 2026-09-21T20:56:19Z box_class gpu-5090-laptop-24g card GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff (NVIDIA GeForce RTX 5090 Laptop GPU) cap cell A RANGE, never a number — this board refuses -pl, so the boost pass has no cap to name. Per-stage ranges are in /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/*-cap-posture-dynamic-boost-*.json and in the 2 Hz trace of every scored result file. Do not type a wattage here. envelope now 150 95 175 (enforced,default,max W) clocks 180,810,3090 (sm,mem,max.sm MHz) clock lock OFF-BY-CARD, read FROM THE CARD clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. (the unit ai-perf.service reads active, which is NOT evidence: it is a oneshot and stays 'active' for the uptime after -rgc.) nvidia-powerd active LEG 1 exit 0 LEG 2 exit 1 PASS exit 1 LEG 1 result files carrying dynboost in their name: f367c81079f80faa gpu-5090-laptop-24g-card-mapping-dynboost.json df3c5e9db262767d gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-dynboost.txt a88d501608f73c7d gpu-5090-laptop-24g-concurrency-one-dynboost.json ffe2685821e8d37b gpu-5090-laptop-24g-embed-one-dynboost.json aa45d3fa436c34c0 gpu-5090-laptop-24g-g2-doorman-dynboost.json 9844175d89bf1e4e gpu-5090-laptop-24g-g3-one-dynboost-g3-vision.json 6d799ccc4655deaa gpu-5090-laptop-24g-idle-gemma4-dynboost.json 8de3494229148885 gpu-5090-laptop-24g-m4-one-dynboost-auto-ladder.json e639a799c13f0cdf gpu-5090-laptop-24g-m4-one-dynboost-forced-at131072.json bbc191ec30061012 gpu-5090-laptop-24g-m4-one-dynboost-forced-at98304.json 1a0712337e7ce3d0 gpu-5090-laptop-24g-m5-one-dynboost-auto-ladder.json 05f60af105c83df1 gpu-5090-laptop-24g-m5-one-dynboost-filled.json 571a60ba467861b2 gpu-5090-laptop-24g-m5-one-dynboost-kvf16-at131072.json dd8e3b901821c6a9 gpu-5090-laptop-24g-m5-one-dynboost-kvq4_0-at131072.json 335fcf2fd0d5de1c gpu-5090-laptop-24g-ups-gate-dynboost.txt LEG 2 result files carrying dynboost in their name: f68c3a43144f343e inventory-inventory-dynboost.json stated absences at this pass: UPS no `upsc` on this box: NUT is not installed, nut-server and nut-monitor are inactive, the battery exposes no power_now/energy_now, and /sys/class/powercap/*/energy_uj is root-only (all read 2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason. Board watts — which is what every table in this ladder prints — are read normally at 2 Hz. fan fan.speed reads [N/A] on this board (2026-09-21T14:52Z): the driver reports no fan for it. The fan column is a declared non-figure with this reason, not a gap to be filled later. memory temp temperature.memory reads [N/A], as on every board of this ladder. The core-temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument. arm T the training arm's preprocessed tensor set lives on benchbox, which was unreachable from 2026-09-21T15:03Z. Arm T is a SECOND TRAIN and legs 1 and 2 never depended on it. pass dynboost finished rc=1 at 2026-09-21T20:56:19Z memory guard restored: memshed.timer is active (it was active); MEMSHED_DRY_RUN unset free receipt: journalctl --user -u memshed --since '2026-09-21T20:23:53Z' | grep SHEDDING — every line there is a shed this bench WOULD have taken, with its signature. PASS — the laptop's 5090, posture DYNAMIC-BOOST (the shipped posture) started 2026-09-21T21:01:46Z on this laptop harness /workshop/bench-laptop-5090-2026-09-21/harness-llm render /workshop/bench-laptop-5090-2026-09-21/harness-render log /workshop/bench-laptop-5090-2026-09-21/logs/rung-dynboost.log ⚠ this pass has NO CAP. The board refuses -pl, so there is nothing to set and nothing to wait for: nvidia-powerd floats the enforced limit between this board's default (95 W) and its maximum (175 W) against the CPU's draw. Every figure's cap cell is a RANGE out of the 2 Hz trace, and a single wattage typed into one would be a fabrication. The per-stage ranges land in results/*-cap-posture-dynamic-boost-*.json. ⚠ the clocks run UNLOCKED here, exactly as in the 95 W rung: the operator released the boot-time -lgc in Step 1 and re-applies it only in the runsheet's LAST step. The verdict is read back FROM THE CARD at this pass's open and written into its receipt. receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. envelope at this pass's open: 150 95 175 W (enforced, default, max) memory guard: memshed.timer was active at the window's open (remembered in /workshop/bench-laptop-5090-2026-09-21/.memshed-was) memory guard: memshed.timer is now inactive; MEMSHED_DRY_RUN=1 (STOPPED, never masked: the unit file is a symlink into ~/estate/systemd and 'mask --force' would destroy it. glimmer-check.timer is left alone — it next fires Tue 14:43Z, outside any daytime window, and masking a timer this bench does not need to touch is one more thing to forget to restore.) ================================================================================ PASS dynboost — posture dynamic-boost — LEG 1 (the eleven-arm language bank), then LEG 2 card GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff (NVIDIA GeForce RTX 5090 Laptop GPU) class gpu-5090-laptop-24g cap arg dynboost start 2026-09-21T21:01:46Z ================================================================================ receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. envelope at the pass's open: enforced/default/max = 150 95 175 W the enforced limit FLOATS in this posture. Its cell is a range from the 2 Hz trace. clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 1590, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock-lock receipt -> /workshop/bench-laptop-5090-2026-09-21/logs/clock-lock-dynboost.txt ---- LEG 1: /workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh dynboost ---- card name read back: 'NVIDIA GeForce RTX 5090 Laptop GPU' (arm one; exact equality, not a substring) === the board, before anything is started (this is the dynamic-boost block) === 21:01:51Z index, uuid, name, power.limit [W], enforced.power.limit [W], power.default_limit [W], power.max_limit [W], persistence_mode, pcie.link.width.current, clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz], temperature.gpu, fan.speed [%] 0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 140.00 W, 95.00 W, 175.00 W, Enabled, 8, 180 MHz, 405 MHz, 3090 MHz, 31, [N/A] receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. posture [block dynboost open]: dynamic-boost — enforced/default/max = 140 95 175 W the enforced limit FLOATS here. Its cell is a RANGE from the 2 Hz trace, never a number. memory guard: memshed.timer is inactive; MEMSHED_DRY_RUN=1 === the UPS at the door (expected: absent on this box) === 21:01:52Z UPS UNREADABLE at block start -- whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason, and board watts are read normally: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (read 2026-09-21T14:52Z). Whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason; board watts, which is what every table in this ladder prints, are read normally from the card at 2 Hz. receipt -> results/gpu-5090-laptop-24g-ups-gate-dynboost.txt clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock-lock receipt -> results/gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-dynboost.txt === instance: one card, kv q8_0, parallel 1 === 21:02:00Z "persistence_mode_at_start": "0, Enabled;", "pcie_link_width_at_start": "0, 8;" } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [card-mapping open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [card-mapping] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [card-mapping]: every 5s -> results/cap-witness-dynamic-boost-card-mapping.csv (pid 1322828) === card mapping: watch which physical card's memory rises === 21:02:13Z one: MATCHES: declared soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c, and that is the card whose memory grew every shape checked runs on the card it declares -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-card-mapping-dynboost.json clocks at [card-mapping] close: 1597,14001,3090 (sm,mem,max.sm MHz) posture [card-mapping close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=2 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-card-mapping.json posture [armA-ladder-gemma4 open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA-ladder-gemma4] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armA-ladder-gemma4]: every 5s -> results/cap-witness-dynamic-boost-armA-ladder-gemma4.csv (pid 1323050) === armA-ladder-gemma4 === 21:02:21Z run 1: decode=155.236 tok/s ttft=430.94 ms W=115.01 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0} contention gate: busy: busiest card mean 9.2% > 5.0% -- waiting (attempt 1/6) run 2: decode=155.045 tok/s ttft=434.2 ms W=148.79 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0} contention gate: busy: busiest card mean 17.8% > 5.0% -- waiting (attempt 1/6) run 3: decode=155.536 tok/s ttft=393.96 ms W=146.41 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0} --- num_ctx 98,304 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0}; runner RSS 1.48 GiB run 1: decode=155.643 tok/s ttft=380.47 ms W=104.35 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0} run 2: decode=155.779 tok/s ttft=397.57 ms W=128.14 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0} run 3: decode=155.838 tok/s ttft=350.69 ms W=147.68 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0} --- num_ctx 131,072 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}; runner RSS 1.52 GiB run 1: decode=155.128 tok/s ttft=372.92 ms W=98.58 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0} run 2: decode=155.251 tok/s ttft=420.66 ms W=146.1 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0} run 3: decode=155.593 tok/s ttft=318.23 ms W=122.57 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0} the biggest context this arm holds for gemma4:26b is 131,072 tokens -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-auto-ladder.json clocks at [armA-ladder-gemma4] close: 1635,14001,3090 (sm,mem,max.sm MHz) posture [armA-ladder-gemma4 close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=94 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=94) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA-ladder-gemma4.json arm A's headline rung at dynamic-boost, read from its own result file: num_ctx 131072 posture [armA2-kv-axis open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA2-kv-axis] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armA2-kv-axis]: every 5s -> results/cap-witness-dynamic-boost-armA2-kv-axis.csv (pid 1337089) === A2 instance kv=f16 === 21:10:13Z } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvf16-p1.json to every bencher so the result files carry it === armA2-kvf16-at131072 === 21:10:49Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=17477150965 size_vram=17477150965 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20696.0} run 1: decode=166.175 tok/s (wall 165.208) ttft=451.55 ms W=114.49 J/1k=688.97 temp=44.0C fan=None% clk=2122.0-2265.0MHz mem=14001.0MHz NEITHER cap: the SM clock varied with draw at 76.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it run 2: decode=165.656 tok/s (wall 164.53) ttft=436.29 ms W=116.01 J/1k=700.31 temp=44.0C fan=None% clk=2115.0-2287.0MHz mem=14001.0MHz NEITHER cap: the SM clock varied with draw at 77.3% of the cap and the card at 44 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it run 3: decode=165.545 tok/s (wall 164.615) ttft=439.78 ms W=144.06 J/1k=870.21 temp=44.0C fan=None% clk=2107.0-2265.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 96.0% of the 150 W cap, at 44 C -- well below any throttling temperature spread gate: spread 0.4% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-kvf16-at131072.json wrote 1 file(s) kv=f16 holds num_ctx 131072 at dynamic-boost -- that is this KV type's largest window === A2 instance kv=q4_0 === 21:11:59Z } next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq4_0-p1.json to every bencher so the result files carry it === armA2-kvq4_0-at131072 === 21:12:07Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=17747851344 size_vram=17747851344 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19116.0} run 1: decode=155.641 tok/s (wall 154.613) ttft=367.61 ms W=109.21 J/1k=701.68 temp=45.0C fan=None% clk=1597.0-2377.0MHz mem=14001.0MHz NEITHER cap: the SM clock varied with draw at 72.8% of the cap and the card at 45 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it contention gate: busy: busiest card mean 46.5% > 5.0% -- waiting (attempt 1/6) run 2: decode=155.712 tok/s (wall 154.825) ttft=319.39 ms W=148.71 J/1k=955.03 temp=44.0C fan=None% clk=2212.0-2377.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 106.2% of the 140 W cap, at 44 C -- well below any throttling temperature contention gate: busy: busiest card mean 44.5% > 5.0% -- waiting (attempt 1/6) run 3: decode=155.908 tok/s (wall 155.047) ttft=336.23 ms W=149.81 J/1k=960.89 temp=44.0C fan=None% clk=2182.0-2385.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 107.0% of the 140 W cap, at 44 C -- well below any throttling temperature spread gate: spread 0.2% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-kvq4_0-at131072.json wrote 1 file(s) kv=q4_0 holds num_ctx 131072 at dynamic-boost -- that is this KV type's largest window clocks at [armA2-kv-axis] close: 2055,14001,3090 (sm,mem,max.sm MHz) posture [armA2-kv-axis close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=40 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=40) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA2-kv-axis.json === restore the q8_0 instance (the seat posture) === 21:13:30Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [armA3-filled open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armA3-filled] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armA3-filled]: every 5s -> results/cap-witness-dynamic-boost-armA3-filled.csv (pid 1344787) === armA3-filled === 21:14:06Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. ### gpu-5090-laptop-24g m5 gemma4:26b -- FILLED WINDOW at num_ctx 131,072 (75% target = 98,304 tokens) ### fit at the window: fits: fully resident on the card per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0} filler round 1: 183 copies -> 97992 tokens (target 98304) window filled: 97,992 tokens of 131,072 (74.8% occupied) COLD prefill: 2611.3 tok/s over 98025 tokens, ttft 38644.1 ms contention gate: busy: busiest card mean 19.0% > 5.0% -- waiting (attempt 1/6) run 1: prefill=1306089.808 tok/s (97992 tokens) ttft=721.95 ms decode=72.159 tok/s W=125.09 run 2: prefill=1512105.547 tok/s (97992 tokens) ttft=710.84 ms decode=72.024 tok/s W=103.46 contention gate: busy: busiest card mean 38.0% > 5.0% -- waiting (attempt 1/6) run 3: prefill=1265196.509 tok/s (97992 tokens) ttft=737.75 ms decode=72.171 tok/s W=123.99 prompt tokens: 97992 of 131072 (0.7476) prefill: 1306089.808 tok/s ttft: 721.95 ms decode at depth: 72.159 tok/s -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-dynboost-filled.json clocks at [armA3-filled] close: 1590,14001,3090 (sm,mem,max.sm MHz) posture [armA3-filled close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=32 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=32) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armA3-filled.json posture [armD-ladder-mistral open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armD-ladder-mistral] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armD-ladder-mistral]: every 5s -> results/cap-witness-dynamic-boost-armD-ladder-mistral.csv (pid 1349835) === armD-ladder-mistral === 21:16:46Z per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0}; runner RSS 0.87 GiB run 1: decode=44.93 tok/s ttft=174.14 ms W=131.22 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0} contention gate: busy: busiest card mean 14.9% > 5.0% -- waiting (attempt 1/6) run 2: decode=44.846 tok/s ttft=215.65 ms W=141.06 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0} contention gate: busy: busiest card mean 39.6% > 5.0% -- waiting (attempt 1/6) run 3: decode=44.799 tok/s ttft=205.62 ms W=140.72 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0} --- num_ctx 65,536 --- fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 1440.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0}; runner RSS 0.89 GiB run 1: decode=44.963 tok/s ttft=199.94 ms W=135.76 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0} contention gate: busy: busiest card mean 19.8% > 5.0% -- waiting (attempt 1/6) run 2: decode=44.888 tok/s ttft=203.72 ms W=141.78 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0} run 3: decode=44.897 tok/s ttft=216.32 ms W=131.27 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0} --- num_ctx 98,304 --- does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM); per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 2550.0} per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 22855.0}; runner RSS 14.18 GiB ladder stops at num_ctx 98,304: does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM) the biggest context this arm holds for mistral-small3.2:24b is 65,536 tokens -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-auto-ladder.json clocks at [armD-ladder-mistral] close: 180,9001,3090 (sm,mem,max.sm MHz) posture [armD-ladder-mistral close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=102 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=102) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armD-ladder-mistral.json posture [armD-forced-probe open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armD-forced-probe] open: 180,9001,3090 (sm,mem,max.sm MHz) cap witness [armD-forced-probe]: every 5s -> results/cap-witness-dynamic-boost-armD-forced-probe.csv (pid 1365426) === D-probe: the planner SPILLED at num_ctx 98304 -- re-running it forced === 21:25:16Z === armD-forced-at98304 === 21:25:16Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=23109042175 size_vram=23109042175 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 23170.0} run 1: decode=44.963 tok/s (wall 44.769) ttft=204.64 ms W=140.48 J/1k=3124.35 temp=48.0C fan=None% clk=1650.0-1657.0MHz mem=14001.0MHz no clock drop during decode (draw held at 93.7% of the 150 W cap) contention gate: busy: busiest card mean 13.7% > 5.0% -- waiting (attempt 1/6) run 2: decode=44.917 tok/s (wall 44.729) ttft=208.97 ms W=141.68 J/1k=3154.26 temp=47.0C fan=None% clk=1642.0-1657.0MHz mem=14001.0MHz no clock drop during decode (draw held at 101.2% of the 140 W cap) run 3: decode=44.946 tok/s (wall 44.756) ttft=209.17 ms W=136.42 J/1k=3035.21 temp=48.0C fan=None% clk=1650.0-1657.0MHz mem=14001.0MHz no clock drop during decode (draw held at 90.9% of the 150 W cap) spread gate: spread 0.1% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-forced-at98304.json wrote 1 file(s) === D-probe: the forced rung FIT -- climbing one more to find the first that does not === 21:26:38Z === armD-forced-at131072 === 21:26:38Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=None size_vram=None -> refused: /api/ps reported no size for this model per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 0.0} -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-dynboost-forced-at131072.json wrote 1 file(s) clocks at [armD-forced-probe] close: 1590,14001,3090 (sm,mem,max.sm MHz) posture [armD-forced-probe close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=19 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=19) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armD-forced-probe.json posture [armE-concurrency open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [armE-concurrency] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [armE-concurrency]: every 5s -> results/cap-witness-dynamic-boost-armE-concurrency.csv (pid 1370521) === E instance: parallel 4 === 21:26:51Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p4.json to every bencher so the result files carry it === E: concurrency 1 / 2 / 4 at dynamic-boost === 21:27:26Z ### concurrency level 2 ### level 2: sum 254.139 tok/s, wall 202.815 tok/s over 2.5245s, W=113.41 temp=46.0C fan=None% per stream: [123.085, 131.054] level 2: sum 257.258 tok/s, wall 202.821 tok/s over 2.5244s, W=127.02 temp=47.0C fan=None% per stream: [128.66, 128.598] level 2: sum 256.955 tok/s, wall 203.071 tok/s over 2.5213s, W=135.75 temp=47.0C fan=None% per stream: [128.452, 128.503] ### concurrency level 4 ### contention gate: busy: busiest card mean 22.0% > 5.0% -- waiting (attempt 1/6) level 4: sum 348.883 tok/s, wall 290.894 tok/s over 3.5202s, W=136.04 temp=46.0C fan=None% per stream: [85.421, 87.856, 87.817, 87.789] level 4: sum 352.368 tok/s, wall 277.108 tok/s over 3.6953s, W=112.29 temp=47.0C fan=None% per stream: [88.106, 88.134, 88.078, 88.05] level 4: sum 351.997 tok/s, wall 283.934 tok/s over 3.6065s, W=115.27 temp=48.0C fan=None% per stream: [87.994, 87.993, 88.016, 87.994] one-dynboost takes 4 parallel streams at 352.0 tok/s aggregate (sum of streams) / 283.9 tok/s over wall -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-concurrency-one-dynboost.json clocks at [armE-concurrency] close: 180,9001,3090 (sm,mem,max.sm MHz) posture [armE-concurrency close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=33 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=33) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-armE-concurrency.json posture [g2-doorman open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g2-doorman] open: 180,9001,3090 (sm,mem,max.sm MHz) cap witness [g2-doorman]: every 5s -> results/cap-witness-dynamic-boost-g2-doorman.csv (pid 1375395) === G2: the doorman on mistral at num_ctx 65536 (1 caller, then 4) === 21:29:36Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. fit: SPLIT and scored as one: 56.2% VRAM / 43.8% RAM contention gate: busy: busiest card mean 6.9% > 5.0% -- waiting (attempt 1/6) concurrency 1: 10 calls, median 438.98 ms, p95 468.26 ms, max 468.26 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 32.11} concurrency 4: 8 calls, median 948.56 ms, p95 4567.46 ms, max 4567.46 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 73.56} a doorman call on mistral-small3.2:24b costs 438.98 ms at the median and 468.26 ms at p95, one caller at a time -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g2-doorman-dynboost.json clocks at [g2-doorman] close: 1590,14001,3090 (sm,mem,max.sm MHz) posture [g2-doorman close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=11 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g2-doorman.json === G3/G4 instance: back to parallel 1 === 21:30:31Z next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it posture [g3-vision open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g3-vision] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [g3-vision]: every 5s -> results/cap-witness-dynamic-boost-g3-vision.csv (pid 1378485) === G3: a vision seat (TEXT path only -- no image is read or sent by this bench) === 21:31:06Z === g3-vision === 21:31:06Z prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101 base=http://127.0.0.1:11470 arm=one live=False runs=3 ### gpu-5090-laptop-24g g3 minicpm-v4.5:latest -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ### memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. UPS before arm: load=None% status=None keep-alive: bench instance: held for the arm's duration /api/ps: size=5596931685 size_vram=5596931685 -> fits: fully resident on the card per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 6696.0} run 1: decode=121.722 tok/s (wall 120.536) ttft=132.94 ms W=82.0 J/1k=673.66 temp=43.0C fan=None% clk=1800.0-1830.0MHz mem=14001.0MHz no clock drop during decode contention gate: busy: busiest card mean 9.7% > 5.0% -- waiting (attempt 1/6) run 2: decode=121.733 tok/s (wall 120.551) ttft=132.89 ms W=145.85 J/1k=1198.11 temp=43.0C fan=None% clk=180.0-1875.0MHz mem=14001.0MHz the POWER CAP doing its job: the SM clock fell while draw sat at 104.2% of the 140 W cap, at 43 C -- well below any throttling temperature contention gate: busy: busiest card mean 19.4% > 5.0% -- waiting (attempt 1/6) run 3: decode=121.762 tok/s (wall 120.59) ttft=147.37 ms W=144.74 J/1k=1188.72 temp=42.0C fan=None% clk=1800.0-1860.0MHz mem=14001.0MHz no clock drop during decode (draw held at 103.4% of the 140 W cap) spread gate: spread 0.0% of the median within the 15% gate -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g3-one-dynboost-g3-vision.json wrote 1 file(s) clocks at [g3-vision] close: 1635,14001,3090 (sm,mem,max.sm MHz) posture [g3-vision close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 150 W, max 150 W, mean 149.0 W, n=17 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 150, n=17) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g3-vision.json posture [g4-embed open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [g4-embed] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [g4-embed]: every 5s -> results/cap-witness-dynamic-boost-g4-embed.csv (pid 1380748) === G4: embeddings === 21:32:30Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. keep-alive: bench instance: held for the arm's duration batch 1: 243.23 ms, 263.126 texts/s, W=22.54 batch 2: 236.42 ms, 270.7 texts/s, W=22.52 batch 3: 236.56 ms, 270.55 texts/s, W=22.54 batch 4: 246.38 ms, 259.765 texts/s, W=22.55 batch 5: 237.93 ms, 268.988 texts/s, W=22.54 one-dynboost: 269.0 texts/s on a batch of 64 (237.93 ms per batch) -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-embed-one-dynboost.json clocks at [g4-embed] close: 2325,14001,3090 (sm,mem,max.sm MHz) posture [g4-embed close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 150 W, median 150 W, max 150 W, mean 150.0 W, n=5 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 150 W, flat ⚠ THE LIMIT DID NOT MOVE over this stage. That is a finding, not a range: check that nvidia-powerd is actually floating the TGP. stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-g4-embed.json posture [idle open]: dynamic-boost — enforced/default/max = 150 95 175 W the limit is FLOATING by design; this gate records it and never stops on a change. clocks at [idle] open: 1590,14001,3090 (sm,mem,max.sm MHz) cap witness [idle]: every 5s -> results/cap-witness-dynamic-boost-idle.csv (pid 1382200) === idle receipt at dynamic-boost === 21:32:54Z memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead. empty: no model on the cards cards empty, nothing loaded: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 7.14 W | pair 7.14 W | UPS whole box None W (None %) loading gemma4:26b (num_ctx=131072, num_gpu=None) fit: fits: fully resident on the card model resident, not generating: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 11.34 W | pair 11.34 W | UPS whole box None W (None %) with gemma4:26b resident and not generating, the card draws 11.34 W (board power), against 7.14 W with the cards empty -- a resident cost of 4.2 W -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-idle-gemma4-dynboost.json clocks at [idle] close: 180,14001,3090 (sm,mem,max.sm MHz) posture [idle close]: dynamic-boost — enforced/default/max at open 150 95 175 W, at close 150 95 175 W enforced.power.limit over this stage: min 140 W, median 140 W, max 150 W, mean 144.0 W, n=19 the board envelope it floated inside: default 95.0 W, max 175.0 W cap cell for this stage: 140-150 W (median 140, n=19) stage envelope -> results/gpu-5090-laptop-24g-cap-posture-dynamic-boost-idle.json === dynamic-boost block finished === 21:34:26Z stopping bench-5090laptop-cpu-box --- receipt: units --- --- receipt: ports --- no bench port listening (1147[01]) -- clean index, uuid, power.limit [W], enforced.power.limit [W], persistence_mode, temperature.gpu, fan.speed [%], memory.used [MiB], clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz] 0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, [N/A], 150.00 W, Enabled, 33, [N/A], 15 MiB, 937 MHz, 810 MHz, 3090 MHz posture [block dynboost close]: dynamic-boost — enforced/default/max = 150 95 175 W (the envelope at the block's open was 140 95 175 W. The enforced limit moving between the two IS the measurement; the DEFAULT or the MAX moving would not be, and cap_stage_close_posture refuses on that at every stage boundary.) the UPS is absent on this box -- see results/gpu-5090-laptop-24g-ups-gate-dynboost.txt for the withheld-figure receipt written at the block's open. Board watts are the measured quantity in every table this bench fills. LEG 1 at dynboost exited 0 envelope between the legs: 150 95 175 W ---- LEG 2: /workshop/bench-laptop-5090-2026-09-21/harness-render/run_rung.sh dynboost ---- == render rung dynboost · started 2026-09-21T21:34:26Z == card expected NVIDIA GeForce RTX 5090 Laptop GPU comfy root /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1 comfy python /workshop/ComfyUI/.venv/bin/python comfy version 0.21.1 (26515acd) receipt: nvidia-powerd is active — Dynamic Boost IS floating this board's limit, which is the whole point of this posture. Its cap cell is a RANGE, never a number. card reads 0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 150.00 W, 24463 MiB, 8, 16, 937 MHz, 810 MHz posture dynamic-boost — enforced/default/max = 150 95 175 W. The enforced limit FLOATS while these arms render; every result file's trace reduces it to min/median/max and THAT is the cap cell. clock lock [by the CARD]: OFF-BY-CARD evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 1155, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. the unit ai-perf.service reads active — NOT EVIDENCE: it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise. clock lock unprobed by the CARD (when the lock is in force this board cannot clock below 1200 MHz and a power posture means something different from what it means on the unlocked desktop cards this leg compares to). The unit state is not evidence; see above. port band 18190-18199 free; this box's own comfyui.service is inactive -- arm inventory · suffix inventory-dynboost · 2026-09-21T21:34:32Z -> results/inventory-inventory-dynboost.json -- arm inventory exited 0 · 2026-09-21T21:34:32Z posture [after inventory at dynboost]: enforced/default/max = 140 95 175 W -- arm r1 · suffix r1-dynboost · 2026-09-21T21:34:32Z card available at 2026-09-21T21:34:32Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0% cap 140.0 W AT THIS INSTANT (posture dynamic-boost: nvidia-powerd is floating it; the cell is a RANGE from each arm's trace) · UPS load unreadable R1 klein-4b-8st on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load)) r1-r1-dynboost-klein-4b-8st-c0-b3-warm[0] 9.959s exec / 3 img = 3.320s per image r1-r1-dynboost-klein-4b-8st-c0-b3-timed[0] 3.605s exec / 3 img = 1.202s per image r1-r1-dynboost-klein-4b-8st-c0-b3-timed[1] 3.871s exec / 3 img = 1.290s per image R1 klein-4b-32st on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load)) r1-r1-dynboost-klein-4b-32st-c0-b3-warm[0] 57.573s exec / 3 img = 19.191s per image r1-r1-dynboost-klein-4b-32st-c0-b3-timed[0] 13.752s exec / 3 img = 4.584s per image r1-r1-dynboost-klein-4b-32st-c0-b3-timed[1] 14.145s exec / 3 img = 4.715s per image R1 klein-4b-8st-pixel4 on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load)) r1-r1-dynboost-klein-4b-8st-pixel4-c0-b3-warm[0] 5.921s exec / 3 img = 1.974s per image r1-r1-dynboost-klein-4b-8st-pixel4-c0-b3-timed[0] 3.583s exec / 3 img = 1.194s per image r1-r1-dynboost-klein-4b-8st-pixel4-c0-b3-timed[1] 3.887s exec / 3 img = 1.296s per image -> results/r1-r1-dynboost.json -- arm r1 exited 0 · 2026-09-21T21:40:06Z posture [between R1 and R4 at dynboost]: enforced/default/max = 150 95 175 W -- arm r4 · suffix r4-dynboost · 2026-09-21T21:40:06Z card available at 2026-09-21T21:40:06Z · the image service holds 0 MiB · quiet: mean 5.0% <= 5.0% cap 140.0 W AT THIS INSTANT (posture dynamic-boost: nvidia-powerd is floating it; the cell is a RANGE from each arm's trace) · UPS load unreadable r4-c0-warm[0] 3.374s exec / 1 img = 3.374s per image r4-c0-warm[1] 1.461s exec / 1 img = 1.461s per image r4-c0-warm[2] 1.530s exec / 1 img = 1.530s per image -> results/r4-r4-dynboost.json -- arm r4 exited 0 · 2026-09-21T21:50:30Z posture [render rung dynboost close]: enforced/default/max = 150 95 175 W == render rung dynboost · done 2026-09-21T21:50:30Z (R1 exit 0, R4 exit 0) == results/r1-dynboost.json and results/r4-dynboost.json the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally. the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it. LEG 2 at dynboost exited 0 envelope at the pass's close: 150 95 175 W artifact receipt -> /workshop/bench-laptop-5090-2026-09-21/RUNG-DYNBOOST-RECEIPT.txt PASS dynboost — COMPLETE posture dynamic-boost cap argument dynboost written 2026-09-21T21:50:30Z box_class gpu-5090-laptop-24g card GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff (NVIDIA GeForce RTX 5090 Laptop GPU) cap cell A RANGE, never a number — this board refuses -pl, so the boost pass has no cap to name. Per-stage ranges are in /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/*-cap-posture-dynamic-boost-*.json and in the 2 Hz trace of every scored result file. Do not type a wattage here. envelope now 150 95 175 (enforced,default,max W) clocks 1590,14001,3090 (sm,mem,max.sm MHz) clock lock OFF-BY-CARD, read FROM THE CARD clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 1590, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force. (the unit ai-perf.service reads active, which is NOT evidence: it is a oneshot and stays 'active' for the uptime after -rgc.) nvidia-powerd active LEG 1 exit 0 LEG 2 exit 0 PASS exit 0 LEG 1 result files carrying dynboost in their name: 06df284c27c686e7 gpu-5090-laptop-24g-card-mapping-dynboost.json 24d417db90f76467 gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-dynboost.txt 71e300b9836387cd gpu-5090-laptop-24g-concurrency-one-dynboost.json 64d224a82c4c9be5 gpu-5090-laptop-24g-embed-one-dynboost.json 30a621a2e58b0062 gpu-5090-laptop-24g-g2-doorman-dynboost.json 8eb12e48694523df gpu-5090-laptop-24g-g3-one-dynboost-g3-vision.json 521fa966f2b50918 gpu-5090-laptop-24g-idle-gemma4-dynboost.json fc6abb5882799903 gpu-5090-laptop-24g-m4-one-dynboost-auto-ladder.json 1a68b29a37429b4c gpu-5090-laptop-24g-m4-one-dynboost-forced-at131072.json eac870e4bd188cc1 gpu-5090-laptop-24g-m4-one-dynboost-forced-at98304.json 94d14f844cf40c7e gpu-5090-laptop-24g-m5-one-dynboost-auto-ladder.json 44f30729477a5b4e gpu-5090-laptop-24g-m5-one-dynboost-filled.json 02f2a15ff0f6317f gpu-5090-laptop-24g-m5-one-dynboost-kvf16-at131072.json a8b7335348bb840b gpu-5090-laptop-24g-m5-one-dynboost-kvq4_0-at131072.json bac0c2297443dfdd gpu-5090-laptop-24g-ups-gate-dynboost.txt LEG 2 result files carrying dynboost in their name: c23585260c80ce96 inventory-inventory-dynboost.json 7404ad4f039e2011 r1-r1-dynboost.json 9d83565278654d17 r4-r4-dynboost.json stated absences at this pass: UPS no `upsc` on this box: NUT is not installed, nut-server and nut-monitor are inactive, the battery exposes no power_now/energy_now, and /sys/class/powercap/*/energy_uj is root-only (all read 2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason. Board watts — which is what every table in this ladder prints — are read normally at 2 Hz. fan fan.speed reads [N/A] on this board (2026-09-21T14:52Z): the driver reports no fan for it. The fan column is a declared non-figure with this reason, not a gap to be filled later. memory temp temperature.memory reads [N/A], as on every board of this ladder. The core-temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument. arm T the training arm's preprocessed tensor set lives on benchbox, which was unreachable from 2026-09-21T15:03Z. Arm T is a SECOND TRAIN and legs 1 and 2 never depended on it. pass dynboost finished rc=0 at 2026-09-21T21:50:30Z memory guard restored: memshed.timer is active (it was active); MEMSHED_DRY_RUN unset free receipt: journalctl --user -u memshed --since '2026-09-21T21:01:46Z' | grep SHEDDING — every line there is a shed this bench WOULD have taken, with its signature.