{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "speeds-from-ram-alone",
 "title": "Speeds from RAM alone",
 "dek": "How fast language models write when a computer's own memory, and no graphics card, does the work, measured on three machines under one registration frozen before the first run: a mixture of experts of about 180 billion parameters at 6.5 to 8.3 tokens a second from 128 GB of eight-channel memory; a 26-billion mixture at 14.6 on a laptop and 13.6 on a desktop with 32 GB, each at its engine's default threads (4 and 8); a 24-billion dense model at 2.6 to 5.5 on the same two across the thread settings tried. The house's bandwidth rule held on the laptop; the small desktop wrote 2.6 times faster than the rule predicts from its measured bandwidth, a verdict this page prints rather than explains.",
 "published": "2026-11-07",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/speeds-from-ram-alone/",
  "md_url": "https://research.strata2signal.com/speeds-from-ram-alone/index.md",
  "md_sha": "9aa3d3d4b5f3137ff800e4d5238e4234b2435928a1a9685c09a936938ebbe53d",
  "html_sha": "715b5cf0e426c12d93f60ad19c9b22e9773c087102a529f1d3665f0f886bbf79",
  "anchors_sha": "24f390714e0f63ddb35413cbce2a74cd4269aac56b64a1d9764677fb3f48bf8e"
 },
 "short": {
  "paragraph": "By the rule this house published in August, a language model with no graphics card writes about as fast as its computer can stream the model's working weights out of memory, once for every token. This page measured that on three more machines, every card in them hidden from the run and proven unused. A laptop with an Intel Core Ultra 9 290HX Plus and 64 GB of DDR5-5600 wrote 14.6 tokens a second with a 26-billion-parameter mixture of experts and 2.65 with a 24-billion dense model at the engine's default of 4 threads; a desktop with an AMD Ryzen 7 3700X and 32 GB of DDR4 wrote 13.6 and 2.57 at its default of 8; a desktop with an AMD EPYC 7532 and 128 GB of DDR4-3200 in eight channels wrote 6.5 to 8.3 with a mixture of about 180 billion parameters whose 54.8 GB of weights are more than any gaming card holds. The laptop's memory moved 59.8 GB/s and its writing speed followed the rule; the small desktop's read 21.4 GB/s and the model wrote 2.6 times faster than the rule predicts (2.3 times the top of the band it registered), a verdict printed and not explained. More threads made the dense model twice as fast on the laptop and the mixture a third slower, and the slower mixture runs were the cooler ones: none throttled while writing. The big mixture's first word on a 4,038-token question took 98 seconds at best.",
  "from_dek": false,
  "counts": {
   "words": 3363,
   "minutes": 15,
   "tables": 4,
   "kit": true
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "why-this-page",
   "heading": "Why this page",
   "level": 2,
   "span": [
    2299,
    3192
   ],
   "chunks": [
    [
     2299,
     3192
    ]
   ],
   "chars": 893,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "a-few-words-this-page-leans-on",
   "heading": "A few words this page leans on",
   "level": 2,
   "span": [
    3192,
    5184
   ],
   "chunks": [
    [
     3192,
     5184
    ]
   ],
   "chars": 1992,
   "digest": "this section defines technical terms used in the benchmarks, including token, tokens a second, prefill and decode, memory bandwidth, threads, dense and mixture of experts, Q4_K_M and IQ3_S, NOT WALL, and THROTTLED.",
   "digest_skipped": null
  },
  {
   "id": "memory-speed-and-the-rule-this-bench-registered",
   "heading": "Memory speed, and the rule this bench registered",
   "level": 2,
   "span": [
    5184,
    8125
   ],
   "chunks": [
    [
     5184,
     8125
    ]
   ],
   "chars": 2941,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-it-was-measured",
   "heading": "How it was measured",
   "level": 2,
   "span": [
    8125,
    11158
   ],
   "chunks": [
    [
     8125,
     11158
    ]
   ],
   "chars": 3033,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-laptop-intel-core-ultra-9-290hx-plus-64-gb-of-ddr5-5600",
   "heading": "The laptop: Intel Core Ultra 9 290HX Plus, 64 GB of DDR5-5600",
   "level": 2,
   "span": [
    11158,
    14147
   ],
   "chunks": [
    [
     11158,
     14147
    ]
   ],
   "chars": 2989,
   "digest": "the laptop with an Intel Core Ultra 9 290HX Plus and 64 GB of DDR5-5600 was measured on mains. it shows performance for gemma4:26b and mistral-small3.2:24b at different thread counts, noting THROTTLED labels.",
   "digest_skipped": null
  },
  {
   "id": "the-two-channel-desktop-amd-ryzen-7-3700x-32-gb-of-ddr4",
   "heading": "The two-channel desktop: AMD Ryzen 7 3700X, 32 GB of DDR4",
   "level": 2,
   "span": [
    14147,
    15813
   ],
   "chunks": [
    [
     14147,
     15813
    ]
   ],
   "chars": 1666,
   "digest": "the AMD Ryzen 7 3700X desktop with 32 GB of DDR4 was measured with no other work running. it shows that halving the threads for the mixture cost 13 per cent and for the dense model bought 4.",
   "digest_skipped": null
  },
  {
   "id": "the-eight-channel-desktop-amd-epyc-7532-128-gb-of-ddr4-3200",
   "heading": "The eight-channel desktop: AMD EPYC 7532, 128 GB of DDR4-3200, the big mixture",
   "level": 2,
   "span": [
    15813,
    20156
   ],
   "chunks": [
    [
     15813,
     20156
    ]
   ],
   "chars": 4343,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-the-numbers-do-not-say",
   "heading": "What the numbers do not say",
   "level": 2,
   "span": [
    20156,
    21077
   ],
   "chunks": [
    [
     20156,
     21077
    ]
   ],
   "chars": 921,
   "digest": "this section clarifies that the machines are not joined to the battery page, no wall watts are measured, and the two-channel desktop's above has no explanation.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    21077,
    21815
   ],
   "chunks": [
    [
     21077,
     21815
    ]
   ],
   "chars": 738,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    21815,
    22663
   ],
   "chunks": [
    [
     21815,
     22663
    ]
   ],
   "chars": 848,
   "digest": "this section lists the bench's own result files, registration, amendments, and several other published pages used as reference.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-4c2cfbdc",
   "q": "how does the article define tokens a second?",
   "a": "Tokens a second describes the speed at which an answer appears. The house rule of thumb suggests that 10 or more tokens per second feels fluid, while 4 to 6 feels comfortable.",
   "cites": [
    "a-few-words-this-page-leans-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-864194be",
   "q": "how is the memory speed rule calculated?",
   "a": "The rule states that tokens a second is approximately bandwidth divided by bytes streamed per token. A result is considered to hold if the ratio k lands between 0.20 and 0.28.",
   "cites": [
    "memory-speed-and-the-rule-this-bench-registered"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-7c055799",
   "q": "how were the model runs isolated from graphics hardware?",
   "a": "Runs were performed inside a sandbox with no graphics device files, emptied card library device variables, and no network except for a loopback. The absence of a card was verified through multiple checks.",
   "cites": [
    "how-it-was-measured"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-fc0759d7",
   "q": "how does increasing threads affect the laptop's dense model?",
   "a": "Increasing from four to twenty threads made the dense model 2.08 times faster and reduced the first word time from 83 s to 27.9 s, despite the mixture model slowing down.",
   "cites": [
    "the-laptop-intel-core-ultra-9-290hx-plus-64-gb-of-ddr5-5600"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-ef954072",
   "q": "how does thread count impact the big mixture's performance?",
   "a": "Increasing threads from 8 to 32 improved writing speed from 6.51 to 8.32 tokens a second and increased the share of the ceiling from 34 % to 44 %.",
   "cites": [
    "the-eight-channel-desktop-amd-epyc-7532-128-gb-of-ddr4-3200"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-ebfeb47b",
   "q": "what limitations exist regarding the reported energy usage?",
   "a": "The energy readings are not wall watts; they come from the processor's own energy counter. Additionally, the eight-channel desktop's UPS carries other devices, so its reading is not the computer's total draw.",
   "cites": [
    "what-the-numbers-do-not-say"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § The laptop: Intel Core Ultra 9 290HX Plus, 64 GB of DDR5-5600 · § Renders"
  },
  {
   "slug": "electricity-per-image",
   "why": "shares ground with § What the numbers do not say · § What this page does not say"
  },
  {
   "slug": "flash-next-on-three-cards",
   "why": "shares ground with § The laptop: Intel Core Ultra 9 290HX Plus, 64 GB of DDR5-5600 · § The three cards, and how they were run"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe registration, the kit that ran it and the three result folders are the record; every figure above is read from them, and [the data kit](/speeds-from-ram-alone/data/) beside this page carries each leg's summary, triad and receipt line. The writing stood on Ollama 0.32.15 and llama.cpp beneath it, the big mixture on llama.cpp build 11514. The models are Google's Gemma 4 at 26 billion, Mistral Small 3.2 from Mistral AI, and Qwen3.8-Flash-Next; each earlier page names the licence it found. A small human team pasted the stages, armed the night window and signed the amendments; a fleet of AI agents built the kit, read the files and wrote this page under that team's rulings.",
 "kit": {
  "url": "https://research.strata2signal.com/speeds-from-ram-alone/data/",
  "licence": "CC BY 4.0",
  "files": [
   "README.md",
   "eight-channel-desktop/RECEIPT.txt",
   "eight-channel-desktop/summary.json",
   "eight-channel-desktop/triad.json",
   "laptop/RECEIPT.txt",
   "laptop/summary.json",
   "laptop/triad.json",
   "two-channel-desktop/RECEIPT.txt",
   "two-channel-desktop/summary.json",
   "two-channel-desktop/triad.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 9,
  "digests_kept": 5,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 1,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 8
  },
  "new_noun_tokens_checked": [
   "0.20",
   "0.28",
   "0.32.15",
   "1.279",
   "10",
   "11514",
   "128",
   "13",
   "13.5",
   "2.08",
   "24b",
   "26b",
   "27.9",
   "290HX",
   "3080",
   "32",
   "34",
   "3700X",
   "4",
   "44",
   "47",
   "54",
   "6",
   "6.51",
   "64",
   "65",
   "7",
   "7532",
   "76",
   "8",
   "8.32",
   "83",
   "9",
   "98.3",
   "AMD",
   "C",
   "Core",
   "DDR4",
   "DDR4-3200",
   "DDR5-5600",
   "GB",
   "Gemma",
   "IQ3S",
   "Intel",
   "Mistral",
   "Ollama",
   "Plus",
   "Q4KM",
   "RealKeep",
   "RuleSage",
   "Ryzen",
   "THROTTLED",
   "UPS",
   "Ultra",
   "W",
   "WALL",
   "gemma4",
   "mistral-small3.2"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 180,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 59,
    "text_sha": "bedfdd82b478",
    "figure": "54",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "a context reading, not a finding: the section prints the hottest sensor's 47 to 54 °C to say why the throttle label cannot read on that processor, and as a summary line alone it carries neither that point nor any figure the page concludes on (the rule exhibit 103's E6 drew for summary lines, as exhibit 104's R2 applied it)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: a context reading, not a finding: the section prints the hottest sensor's 47 to 54 °C to say why the throttle label cannot read on that processor, and as a summary line alone it carries neither that point nor any figure the page concludes on (the rule exhibit 103's E6 drew for summary lines, …"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 51,
    "text_sha": "4612f8ee6028",
    "figure": "76",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it drops the sentence's condition, 'in each run, under its 95 °C limit', which is why the figure is printed (no run reached the throttle limit, so 0 of 3 THROTTLED); a context reading, not a finding (exhibit 103's E6, exhibit 104's R2)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it drops the sentence's condition, 'in each run, under its 95 °C limit', which is why the figure is printed (no run reached the throttle limit, so 0 of 3 THROTTLED); a context reading, not a finding (exhibit 103's E6, exhibit 104's R2). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 37,
    "text_sha": "808720eb9fcd",
    "figure": "13.5",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "a package counter figure without its label: the page prints 'Idle, the package drew 13.5 W (NOT WALL)', and every package watt this page prints carries NOT WALL; the line drops it, and an idle reading is a context figure, not a finding",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: a package counter figure without its label: the page prints 'Idle, the package drew 13.5 W (NOT WALL)', and every package watt this page prints carries NOT WALL; the line drops it, and an idle reading is a context figure, not a finding. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 224,
    "text_sha": "b5abad88dad3",
    "section": "why-this-page",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it says the tier of machines runs the biggest mixture of experts from memory alone, where the section says this page is that tier on three more machines and the biggest mixture ran on one of them; and it keeps the price's date but drops the price it cites ($399.99)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it says the tier of machines runs the biggest mixture of experts from memory alone, where the section says this page is that tier on three more machines and the biggest mixture ran on one of them; and it keeps the price's date but drops the price it cites ($399.99). Not a gate drop, and recor…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 211,
    "text_sha": "3bd3e164e829",
    "section": "memory-speed-and-the-rule-this-bench-registered",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it says the article establishes the rule, which is the battery page's (published in August; this bench registered a check of it), and that the rule was tested with the 26-billion mixture on the eight-channel desktop, where that model was not run and the rule is read as a ceiling for the big mixture instead",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it says the article establishes the rule, which is the battery page's (published in August; this bench registered a check of it), and that the rule was tested with the 26-billion mixture on the eight-channel desktop, where that model was not run and the rule is read as a ceiling for the big m…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 239,
    "text_sha": "47efe2aa7baf",
    "section": "how-it-was-measured",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it says the warm-ups and the three scored runs were there to ensure no card was used; the section proves no card by its own checks (the sandbox's probe, the engine's own evidence, the host's card monitor, no voided run, and the ceiling on one machine), and the warm-up and three runs are how each arm was counted",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-06: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-06 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it says the warm-ups and the three scored runs were there to ensure no card was used; the section proves no card by its own checks (the sandbox's probe, the engine's own evidence, the host's card monitor, no voided run, and the ceiling on one machine), and the warm-up and three runs are how e…"
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 232,
    "text_sha": "ea3f49623454",
    "chip_id": "c-3cd8c9cb",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it says halving the threads 'improved the dense model by 4' with no unit, where the section says it 'bought the dense model 4' per cent beside the mixture's 13 per cent; a bare figure that lost its unit, the claim of chip c-ca45fea2 removed by hand on 2026-11-06, drafted again under a re-worded question, so its frozen id moved and the carry could not match it",
    "detail": "removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it says halving the threads 'improved the dense model by 4' with no unit, where the section says it 'bought the dense model 4' per cent beside the mixture's 13 per cent; a bare figure that lost its unit, the claim of chip c-ca45fea2 removed by hand on 2026-11-06, drafted again under a re-worded question, so its frozen id moved and the carry could not match it. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 223,
    "text_sha": "8563eb7fec80",
    "chip_id": "c-7aca6874",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-06",
    "why": "it says the page documents 'performance tiers' and focuses specifically on the largest mixture of experts, where the section says the page is the battery page's one emergency tier measured on three more machines, with the biggest mixture run on one of them (the misreading the why-this-page digest was removed by hand for on 2026-11-06)",
    "detail": "removed from this pack by hand on 2026-11-06 (UTC), after the writer ran and after the gates had counted: it says the page documents 'performance tiers' and focuses specifically on the largest mixture of experts, where the section says the page is the battery page's one emergency tier measured on three more machines, with the biggest mixture run on one of them (the misreading the why-this-page digest was removed by hand for on 2026-11-06). Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-11-06T06:45:27Z"
}