The hardware roster · a derived index, not an exhibit

The hardware roster.

Rebuilt from the published exhibits · 7 graphics cards, 6 processors, 10 whole machines, 2 memory classes and 13 parts · 1054 figures, from 24 exhibits · named in 77 · last built 2026-11-01

This roster is in beta and still being tested: every figure is proved against the page it came from, and the way they are read and shown is still changing.

Every card, CPU and machine this shelf has named or measured, the exhibits each appears in, and the figures each exhibit published for it — quoted with their conditions, proved against the row they came from, ranked by nobody.

Its sibling is the model roster, which does the same job keyed by model.

How to read this page

A figure here is one published number: a value an exhibit printed, carried over with the conditions it printed beside it.

  • Every figure is quoted with the conditions its page states — the model, how it was quantised, the runtime and its version, the power cap, the context — and never rewritten to match another page. A figure without its conditions is not a smaller fact, it is a different one.
  • Nothing here is ranked. Where this shelf holds one card measured under two conditions, both readings are printed, one under the other, and you decide which one answers your question.
  • A figure an exhibit later corrected carries the correction, and says so. It is struck through, and its page’s own words about what was wrong are printed underneath.

How this page is built →

Narrow these 38 cards, CPUs and machinesevery kindkind: Graphics cards · 7 of 38 cards, CPUs and machineskind: Processors · 6 of 38 cards, CPUs and machineskind: Whole machines · 10 of 38 cards, CPUs and machineskind: Memory classes · 2 of 38 cards, CPUs and machineskind: Parts priced, never benched · 13 of 38 cards, CPUs and machinesevery makermaker: AMD · 3 of 38 cards, CPUs and machinesmaker: ASUS · 1 of 38 cards, CPUs and machinesmaker: Bestoss · 1 of 38 cards, CPUs and machinesmaker: Corsair · 2 of 38 cards, CPUs and machinesmaker: Crucial · 1 of 38 cards, CPUs and machinesmaker: CyberPower · 1 of 38 cards, CPUs and machinesmaker: Fractal Design · 1 of 38 cards, CPUs and machinesmaker: Intel · 3 of 38 cards, CPUs and machinesmaker: KingBank · 1 of 38 cards, CPUs and machinesmaker: KingSpec · 1 of 38 cards, CPUs and machinesmaker: MSI · 3 of 38 cards, CPUs and machinesmaker: Minisforum · 1 of 38 cards, CPUs and machinesmaker: NVIDIA · 7 of 38 cards, CPUs and machinesmaker: Noctua · 1 of 38 cards, CPUs and machinesmaker: Rimlance · 1 of 38 cards, CPUs and machinesmaker: TEAMGROUP · 1 of 38 cards, CPUs and machinesmaker: not published · 9 of 38 cards, CPUs and machinesmeasured and named-only alikemeasured here · 17 of 38 cards, CPUs and machinesnamed, not measured · 21 of 38 cards, CPUs and machinesevery exhibitexhibit: A 3090, in its own words · 2 of 38 cards, CPUs and machinesexhibit: A harness that rewrites itself · 1 of 38 cards, CPUs and machinesexhibit: A laptop, asked the desktop's questions · 6 of 38 cards, CPUs and machinesexhibit: A Rig Your Friend Already Owns · 5 of 38 cards, CPUs and machinesexhibit: A sealed 3090 Ti FTW3 Ultra, four years after EVGA left the business · 3 of 38 cards, CPUs and machinesexhibit: A Short History of Mistral · 4 of 38 cards, CPUs and machinesexhibit: A short history of Ollama · 3 of 38 cards, CPUs and machinesexhibit: A short history of Qwen · 3 of 38 cards, CPUs and machinesexhibit: A short history of vLLM · 2 of 38 cards, CPUs and machinesexhibit: Ask About This Page · 1 of 38 cards, CPUs and machinesexhibit: The August arrivals · 2 of 38 cards, CPUs and machinesexhibit: The chair trials · 1 of 38 cards, CPUs and machinesexhibit: Teaching a Music Model Chopin in Five Minutes · 2 of 38 cards, CPUs and machinesexhibit: It’s Not a Metaphor: A Language Model Is a Compression of Its Corpus · 1 of 38 cards, CPUs and machinesexhibit: The narrator’s chair, refused. · 1 of 38 cards, CPUs and machinesexhibit: Electricity per million tokens, card by card · 9 of 38 cards, CPUs and machinesexhibit: Elo ratings, from chess to image models · 1 of 38 cards, CPUs and machinesexhibit: EmbeddingGemma 2, measured on release day against the embedder we already run · 1 of 38 cards, CPUs and machinesexhibit: EmbeddingGemma 2's shared space for text, images and audio, measured on what we publish · 2 of 38 cards, CPUs and machinesexhibit: Fifty-Seven Milliseconds: What a Diffusion Step Actually Buys · 2 of 38 cards, CPUs and machinesexhibit: Qwen3.8-Flash-Next on an RTX 5090 laptop, an RTX PRO 6000 Blackwell and an RTX 3090 Ti: the Strata engine and llama.cpp, the numbers · 6 of 38 cards, CPUs and machinesexhibit: Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench · 6 of 38 cards, CPUs and machinesexhibit: FLUX 3 Image Klein, ranked blind in a game's own painted style · 1 of 38 cards, CPUs and machinesexhibit: FLUX 3 Image Klein, run again with its maker's prompt upsampler · 1 of 38 cards, CPUs and machinesexhibit: Four homes for one reranker: what RuleSage's slowest stage cost at each stop · 4 of 38 cards, CPUs and machinesexhibit: GPUs and orchids · 6 of 38 cards, CPUs and machinesexhibit: Half an Hour with Dead Composers · 1 of 38 cards, CPUs and machinesexhibit: The cost to purchase and run a very capable home AI rig · 12 of 38 cards, CPUs and machinesexhibit: How a Vision Model Sees: Eyes for a Machine That Reads · 1 of 38 cards, CPUs and machinesexhibit: How an embedding model turns anything into a list of numbers · 1 of 38 cards, CPUs and machinesexhibit: How the long table works · 2 of 38 cards, CPUs and machinesexhibit: How the print lab works · 3 of 38 cards, CPUs and machinesexhibit: How the speed tool works · 4 of 38 cards, CPUs and machinesexhibit: Indra OCR, measured on RuleSage's rulebook pages · 1 of 38 cards, CPUs and machinesexhibit: The licence ledger · 1 of 38 cards, CPUs and machinesexhibit: llm-d 1.0, and what it means if you run one or two GPUs · 1 of 38 cards, CPUs and machinesexhibit: Nemotron 3.5 Rerank 8B, measured on our own workloads · 1 of 38 cards, CPUs and machinesexhibit: Nine Worlds, One Dog · 1 of 38 cards, CPUs and machinesexhibit: Ollama and vLLM on one RTX 3090 · 4 of 38 cards, CPUs and machinesexhibit: Ollama or vLLM: which to run, and how · 2 of 38 cards, CPUs and machinesexhibit: RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench · 2 of 38 cards, CPUs and machinesexhibit: One 3090 out, one 3090 Ti in, and the cap that decided it · 4 of 38 cards, CPUs and machinesexhibit: OpenJev as a reranker for RuleSage: measured · 1 of 38 cards, CPUs and machinesexhibit: The outside judges · 1 of 38 cards, CPUs and machinesexhibit: Qwen-Image-2.1-Turbo, judged blind in a game's own painted style · 1 of 38 cards, CPUs and machinesexhibit: Reading the answer instead of writing it · 2 of 38 cards, CPUs and machinesexhibit: RTX 5080 vs RTX 3080 Ti · 6 of 38 cards, CPUs and machinesexhibit: The seat trials · 2 of 38 cards, CPUs and machinesexhibit: Six graphics cards, the same questions · 10 of 38 cards, CPUs and machinesexhibit: Six Worlds for the Print Lab · 1 of 38 cards, CPUs and machinesexhibit: Small open deciders, measured: a word counter kept up on our hardest question · 2 of 38 cards, CPUs and machinesexhibit: Ten Minutes with Living Artists · 4 of 38 cards, CPUs and machinesexhibit: The 3080 Rig, Priced · 3 of 38 cards, CPUs and machinesexhibit: The one-card RTX 3090 rig, repriced · 9 of 38 cards, CPUs and machinesexhibit: The Ceiling Is Not the Corpus · 2 of 38 cards, CPUs and machinesexhibit: The cost of intelligence, as a utility · 3 of 38 cards, CPUs and machinesexhibit: The Dog, the Dice, and the Painter: How a Small World Draws Itself · 1 of 38 cards, CPUs and machinesexhibit: The free speed wasn’t free · 1 of 38 cards, CPUs and machinesexhibit: The Instrument Travels · 3 of 38 cards, CPUs and machinesexhibit: The move · 2 of 38 cards, CPUs and machinesexhibit: qwen3.8:27b across four house benches: one seat filled, one floor missed · 1 of 38 cards, CPUs and machinesexhibit: The open call · 2 of 38 cards, CPUs and machinesexhibit: The quantization ladder · 1 of 38 cards, CPUs and machinesexhibit: The same eighteen pictures, card by card · 11 of 38 cards, CPUs and machinesexhibit: The Same Sixteen · 1 of 38 cards, CPUs and machinesexhibit: The Two-Hour Machine, Priced · 5 of 38 cards, CPUs and machinesexhibit: The Typist and the Developer · 2 of 38 cards, CPUs and machinesexhibit: Everyone on the payroll, three at the table · 1 of 38 cards, CPUs and machinesexhibit: Three librarians and a careful reader · 1 of 38 cards, CPUs and machinesexhibit: Three New Voices at the Narrator's Chair · 1 of 38 cards, CPUs and machinesexhibit: Two 3090s out, one 3090 Ti in · 2 of 38 cards, CPUs and machinesexhibit: Two Hours at 12 tok/s · 7 of 38 cards, CPUs and machinesexhibit: Two 3080s against one 3090, and the cap decides · 9 of 38 cards, CPUs and machinesexhibit: The voice trials · 3 of 38 cards, CPUs and machinesexhibit: What 150 Watts Buys · 2 of 38 cards, CPUs and machinesexhibit: What it costs to run a high-end AI rig at home · 6 of 38 cards, CPUs and machinesexhibit: Where your question goes · 2 of 38 cards, CPUs and machines
What kind of thing
Who makes it
Whether this shelf has figures for it
Which exhibit names it (77)
The rail is CSS and nothing else — no script runs on this page. Its four questions narrow independently, and the line above says what each one is set to; the size of two of them together is the one number it cannot count. A narrowed view is not part of the address, so it cannot be linked — each card’s own page is a URL, and that is the thing to share.

Graphics cards

Named by model, every one of them.

GPU91 figures

RTX PRO 6000 Blackwell (96 GB)

96 GB · NVIDIA

91 figures · named in 48 exhibits

models measured on itACE-Step 1.5 · FLUX.1 dev · FLUX.2 [klein] 4B · gemma4:26b · Qwen3.8-Flash-Next · FLUX.1 schnell · gemma4:12b · HiDream-O1-Image-Dev · and 12 more on its own page

what the pages say about the power cap300 W, 420 W, 450 W, 450 W, against the same-day 600 W rung, 500 W, against the same-day 600 W rung, 600 / 500 / 450 W, 600 W, 600 → 450 W, 600 → 500 → 450 W, uncapped in that arm

GPU344 figures

RTX 3090 (24 GB)

24 GB · NVIDIA

344 figures · named in 43 exhibits

models measured on itACE-Step 1.5 · FLUX.2 [klein] 4B · gemma4:12b · klein-4b-32st · klein-4b-8st-pixel4 · klein-4b-8st · gemma4:26b · HiDream-O1-Image-Dev · and 11 more on its own page

what the pages say about the power cap250 W, 250 W · 300 W, 250 watts, matched at 250 W, 280 W, 280 W, 80 per cent of the card's 350 W rating, 300 W, 350 W, 350 W (the board's default), a 3090 at 250 W, the workstation card at 420 W, measured before this workshop capped its cards, the lower cap

GPU221 figures

RTX 5090 Laptop GPU (24 GB)

24 GB · NVIDIA

221 figures · named in 12 exhibits

models measured on itFLUX.2 [klein] 4B · Qwen3.8-Flash-Next · gemma4:26b · HiDream-O1-Image-Dev · Kandinsky 5.0 Image Lite · Krea 2 Turbo · MiniCPM-V 4.5 · Mistral Small 3.2 · and 3 more on its own page

what the pages say about the power cap95 W, a limit of 135-150 W, 140 to 150 W, 140 to 150 W (n=19), 140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them), 140–150 W (n=374; 150 W in 335), 150 W in every one of the 24 samples of its own 2 Hz trace, 150 W over their short stages (n=11 and n=5 at five-second intervals), 150 W, flat, 150 W, flat (n=385), 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches, 150 W, flat — every one of 1,140 samples at 2 Hz, 150 W, flat — every one of 1,141 samples at 2 Hz, at every posture

GPU22 figures

RTX 3080 (10 GB)

10 GB · NVIDIA

22 figures · named in 9 exhibits

models measured on itklein-4b-32st · klein-4b-8st-pixel4 · klein-4b-8st · gemma4:26b · Mistral Small 3.2 · a 14-billion model

what the pages say about the power cap250 W, cap not set by that harness

GPU211 figures

RTX 3090 Ti (24 GB)

24 GB · NVIDIA

211 figures · named in 17 exhibits

models measured on itFLUX.2 [klein] 4B · Qwen3.8-Flash-Next · gemma4:26b · klein-4b-32st · klein-4b-8st · klein-4b-8st-pixel4 · Mistral Small 3.2

what the pages say about the power cap300 W, 350 W, 400 W, 450 W

GPUnamed, not measured

RTX 3080 Ti (12 GB)

12 GB · NVIDIA

the model of two boards this workshop owns, a Founders Edition and an EVGA XC3 Ultra, both 12 GB. The laptop bench’s data kit named it first, listing it among the boards measured; “RTX 5080 vs RTX 3080 Ti” measured the Founders Edition in print beside an RTX 5080 from the same seat, with each board’s price before tax and the day it was ordered; “The same eighteen pictures, card by card” measured the Founders Edition again on its render grid, with the EVGA board held idle below it. This roster has not carried those rows yet, and the word each listing used for its condition is still written down off this shelf

no figure carried yet · named in 9 exhibits

GPUnamed, not measured

RTX 5080 (16 GB)

16 GB · NVIDIA

a card this workshop bought new and measured on 2026-09-24 (UTC), the first 16 GB card through this shelf, in the same seat as a 3080 Ti Founders Edition and an EVGA 3080 FTW3 Ultra; “RTX 5080 vs RTX 3080 Ti” and “The same eighteen pictures, card by card” print its figures in their own tables, and this roster has not carried those rows yet

no figure carried yet · named in 8 exhibits

Processors

Four processor models exist on this shelf in total, and two of them are in one markdown file inside one kit. The thinness is the finding.

Whole machines

A box the pages name as a box. Some hold a card this roster has registered; some hold a card the shelf has never named, and those say so.

machine18 figures

the mini PC

Minisforum

a fanless-class mini desktop, CPU-only inference, 32 GB of memory and no discrete GPU — the machine of the two-hour battery test

18 figures · named in 8 exhibits

models measured on itgemma4:12b · gemma4:26b · Mistral Small 3.2 · nemotron-3.5-lightning:30b-a3b · phi4:14b · qwen3.5:9b · a 26-billion-parameter model

machine13 figures

the server

a CPU-only server: 128 GB of ECC memory, a server-class processor, and a GPU present in the box and deliberately parked

13 figures · named in 1 exhibit

models measured on itgemma4:12b · gemma4:26b · Mistral Small 3.2 · nemotron-3.5-lightning:30b-a3b · phi4:14b · qwen3.5:9b

machine14 figures

Rig B — the gaming laptop

MSI

the workshop's gaming laptop: an RTX 5090 Laptop GPU for the diffusion rematch, deliberately unused for the CPU-only bench

14 figures · named in 5 exhibits

models measured on itgemma4:12b · gemma4:26b · Mistral Small 3.2 · nemotron-3.5-lightning:30b-a3b · phi4:14b · qwen3.5:9b

machine10 figures

the darkroom's box — two RTX 3090s

a consumer tower built around two renewed RTX 3090s on an ordinary two-slot board, with desktop memory and a desktop power supply; the box the price series prices, which its own page described as never off until a dated correction of 2026-09-18 made it a bench machine, switched on when there is a bench to run

10 figures · named in 2 exhibits

what the pages say about the power cap280 W caps for six hours, idle the other eighteen, each card at 280 W, both cards at their caps all day

machine81 figures

the pair — two RTX 3080s in one 2021 desktop

two GeForce RTX 3080 10 GB cards of 2021 in a desktop of the same vintage, one at PCIe x16 and one at x4 of an x16-capable slot, on one supply — 20 GB of graphics memory between them, and the rig exhibit fifty measures against a single card in the same box

81 figures · named in 3 exhibits

models measured on itgemma4:26b · Mistral Small 3.2 · one 26-billion model split across both cards · two copies of a 12-billion model, one per card

what the pages say about the power cap250 W each — 500 W between them, 250 W each — 500 W between them · 300 W each — 600 W between them, 300 W each — 600 W between them

machinenamed, not measured

the 3090's own tower

the ordinary consumer tower its exhibits benched the single 3090 in — board, memory, disk and case all roughly five years old, and the point of the exhibit that benches it

no figure carried yet · named in 2 exhibits

machine14 figures

CyberPower PR1500

CyberPower

a consumer rack/tower uninterruptible power supply — 1500 VA on the label, 900 W by its own firmware's nominal — and the subject of the two-hour battery test, which reads its own meter through four phases either side of the plug coming out

14 figures · named in 2 exhibits

models measured on ita 26-billion-parameter model

machine5 figures

the machine that draws today

the print lab's drawing machine as the exhibit names it — the one machine the lab was pinned to from 2026-09-08, after two machines had shared one line

5 figures · named in 1 exhibit

models measured on itFLUX.2 [klein] 4B

Memory classes

A class registered as an identity in its own right, because on this shelf it is more than one thing and inference is not a road this page takes.

Parts — priced, never benched

Real named parts with prices, in the price pages, and no measurement of their own.

partnamed, not measured

Noctua NH-D15

Noctua

the CPU cooler in the two-card rig's bill

no figure carried yet · named in 2 exhibits

partnamed, not measured

ASUS ROG Strix B550-E

ASUS

the board named in the bill's two notes, beside the one that was bought

no figure carried yet · named in 1 exhibit

partnamed, not measured

MSI PRO Z490-A PRO

MSI

the motherboard in the two-3080 rig's bill, priced refurbished

no figure carried yet · named in 1 exhibit

How this page is built

  • A mention is a row in a table we published, or a line in the data file behind it. That is what the count of exhibits on a card is counting.
  • One thing can hold several published spellings; older pages name a card by its memory class and newer ones name the board, and this page joins them — for the models measured on it too, where the model roster knows the tag behind a page’s own words.
  • Every row is checked against its page’s own row, word for word, by a gate that runs before anything is published; the paragraph under this list says how many rows a program read off a page and how many a person carried across, counted off the rows themselves, and each row in hardware/rows.json names the road it took.

Nothing on this page is typed by a generator: the whole of it is rendered by tools/hardware_page.py out of one committed file of figures. 839 of them were read off a published page’s own table by the roster’s reader (tools/hardware_extract.py) and 215 were typed in from the page by a person; every one is then proved against that page’s own row by tools/hardware_gate.py before any deploy could serve it, and each row says which road it came by.

And one fence this page does not have. The model roster checks what kind of thing each entry is against the licence ledger’s own heading for it, and refuses to build when the two disagree. The ledger carries rows for models, weights and software and not one for hardware, so the identity table here is typed with no second surface to check it against. What stands in for it: an unregistered spelling is a broken build, a spelling two identities claim is a broken build, and every figure on every card page had to survive a gate that re-read it against the page it came from.

What is not on this page, and why

Two kinds of thing are missing on purpose, and both are listed rather than left out, because a name absent from an index reads the same as a name nobody looked for.

Strings that name a ROLE rather than a machine. A kit or a page substitutes a job for a host — and half of these are the published OUTPUT of a redaction pass, so resolving one to a box would not be filling a gap, it would be undoing a publication decision from the other side of the fence.

  • <harness host> — the same placeholder for the other end of the same run — the machine that drove the bench, not the machine under test
  • <serving host> — a PLACEHOLDER the power bench's rung summaries write where a hostname was (what-150-watts-buys/data/rungs/*/summary.json) — the angle brackets are the kit saying the value was withheld, not a name
  • card A — how-the-print-lab-works names it as a convention NOT YET IN USE — "if a second card ever joins it"
  • cloud (agent harness) — the same — somebody else's machine
  • cloud (hosted endpoint) — not this workshop's hardware (outside-judges · candidates[].hardware_class)
  • http://127.0.0.1:11434 — a transport — every machine's own address
  • http://<local daemon — address withheld> — a transport with its host withheld
  • http://[gpu-box-daemon] — the same published alias, as a bracketed daemon address
  • http://the-gpu-box[:port redacted] — the same published alias, as an endpoint with its port withheld
  • local — a transport, not a machine
  • the 96 GB box's production seat A — a SEAT on a card, not a card
  • the 96 GB box's production seat B — a SEAT on a card, not a card
  • the bench endpoint — a role the kit substitutes for a host (the-instrument-travels, the-same-sixteen)
  • the inference workstation (box-neutral by house law) — a role — the kit says so inside the string itself
  • the public host — a ROLE: the web server a plate or a page is served from (how-the-print-lab-works, the-ceiling-is-not-the-corpus/data/provenance.json). It is a job in a pipeline, not a benched machine
  • the reading seat — a seat on the print lab's models, not a machine; its seconds are a model figure and belong to the model roster
  • the training box — a role (the-ceiling-is-not-the-corpus · corpus.host / dataset.host)
  • the-gpu-box — a PUBLISHED ALIAS substituted for a host name (reading-is-fast, the-free-speed-wasnt-free, the-move)

Published exhibits that name no hardware at all. Real exhibits; they simply measure something other than a machine.

  • The night watch: how one agent a night will choose what a small telescope sees — a FIELD GUIDE to a plan with no bench: it names the planned telescope parts (a Sky-Watcher AZ-GTi mount, an Askar FMA180 Pro refractor, a ZWO ASI585MC camera, a ZWO EAF focuser) and "a small always-on computer" by role only; it names no card, processor or class, and records no timing, so there is nothing on it for a hardware row to carry
  • How OCR differs from vision: a field guide, with eight images read three ways — a FIELD GUIDE with a demonstration and no bench: its two OCR engines ran on a laptop CPU and its vision model on the workshop's inference server, both named by role only; it names no card, processor or class, and records no timing, so there is nothing on it for a hardware row to carry
  • How the search works — a FIELD GUIDE to this hub's own search page: it runs nothing on any card or machine, and the card numbers it prints (3090, 3080, 5090) are words typed into the search box and an article title the search ranks, never hardware this page measures; the older iPhone or iPad it names is the reader's own browser
  • Inside a GGUF file — a FIELD GUIDE to a file format: it reads model files' headers and runs no model on any card or machine; it names no card or class, and its one machine, a 64 GB laptop, is quoted with its figure from the three-card bench page, whose hardware line names it
  • CallHome: what local AI tools send out by default, in one table — a WORKSHOP NOTE introducing a sibling site: it runs nothing on any card or machine and names no card, class or machine; the machines it mentions are the reader's own and, by role only, the workshop's own (“these machines”, “the machines”), which is the register's own law; every count on it is read from the register's data file, and nothing on the page measures a machine
  • Uncensored models: what they are and how they work — an EXPLAINER with no bench: it runs no model on any card or machine, and its figures are the file sizes and sha256 values Hugging Face publishes about one upload. The one machine it mentions, a mini PC with no graphics card, is exhibit thirty-two's, named in a link to that page and in a Sources line, and that page carries it on its own hardware line; “one graphics card and the rest of an ordinary PC” describes what an engine needs, “using only one GPU” is a paper's own words about its run, and the offline section's laptop is a reader's imagined machine that nothing on the page measures
  • ONNX Runtime's telemetry on Linux, measured: on by default since 1.29, and the switch that stops it — a BENCH page of one LIBRARY, run on one laptop's processor inside a container, and it names no card, class or machine: its machines are described by role only (“one laptop plugged in to mains power”, “the inference server”), and its figures are the library's, not a board's
  • Who makes FLUX: a short history of Black Forest Labs — a HISTORY of a company; it measures nothing and names no card, class or machine. Its one sentence about hardware says the diffusion bench timed two of the company's models on “several of the workshop's graphics cards”, and that page carries the hardware line; this one has no figure a card could own
  • Introducing the playground — the playground's introduction and a note about a LAYOUT. Every figure on it is a measurement of a rendered page, a file on disk or the arithmetic of a grid; it names open-weight models and libraries but no card, class or machine, and its only words for where the work happens are “machines this workshop runs”
  • The indifferent prediction — a prose piece about somebody else's result; it measures nothing and names no machine this shelf has ever put under test — the one card word on it is a 1997 board in a story, and the fleet beside it is named by generation and never by model
  • Listen for Yourself — an audio shelf — the renders' machine is named on the exhibit that made them, never here
  • Hear It for Yourself — the same: 252 published digests and a player, and no machine string anywhere on the page or in the kit
  • It Was Already There: Where New Knowledge Actually Comes From — a prose piece; names no machine
  • A Dinner Party for the Dead — a prose piece; names no machine
  • The Beat Lab, asked and answered — a Q&A page; names no machine
  • How we work — the method page. It describes how this shelf works and names no card, class or machine anywhere — measured over its twin, its pack and its kit, which carry no card token at all
  • Two New Frontier Models at the Rules Desk — GENERIC ONLY — the graphics card it needs to answer users, which names no card. (Its 140902Z timestamp is the second false-positive class: 4090 inside a stamp.)
  • The compressed photograph — GENERIC ONLY — VRAM figures throughout and no card class or model string anywhere
  • Reading is fast, writing is slow — names hardware only as the-gpu-box, a PUBLISHED ALIAS substituted for a host — a NOT_A_HARDWARE label, not an appearance
  • The map nobody picks up — the same published alias, in five kit files; no card, class or machine string on the page
  • How the Beat Lab works — GENERIC ONLY, and it is worth saying which words: our own machines, graphics cards we own, the ordinary processor of our web server and a rented virtual machine we administer. No card, no class, no machine identity — a service page about where the work happens, which is not the same as a page that says on what

Some hardware here was never benched. The power supply, the board, the memory, the uninterruptible power supply: real named parts with prices, in the price pages, and no measurement of their own. They get a card that says what the thing is and stops there — not a card with an empty table, which reads as rows gone missing rather than as a thing nobody ran a stopwatch on.

Provenance

  • Read or carried, then proved by machine — every figure on every card page was taken off a published page — most by the roster’s own table reader, the rest typed in by a person — and written down with the conditions that page states. Before any deploy can serve it, tools/hardware_gate.py re-reads each one against that page’s markdown twin: the section id must still exist, the byte range must still lie inside it, the quote and the value must still be there word for word, and the card’s own name must be where the row says it is. A row that fails any of those is a broken build.
  • It fails closed — if an exhibit names a card this roster has never seen, the build refuses rather than dropping the card. A missing card is a broken build, not a quiet omission.
  • No new claims — no ranking, no computed ratio, no normalised per-step or per-watt figure the page did not publish. Where a comparison would need judgement the hub has not published, this page links to the page that did the work.
  • No script — the filters are CSS radios and sibling selectors. Open devtools: there is nothing to find.
  • The data — this page is CC BY 4.0. Attribution: strata→signal research, research.strata2signal.com/hardware/.
  • Authorship — benched, drafted and audited by the workshop’s own agents under a human operator’s rulings, the same arrangement told in full here. If something on this page is wrong, tell us.

elsewhere in the workshop

a strata→signal property · hello@strata2signal.com · say hello