DOCS
How it works.
Open the page, your GPU gets benched. Here is exactly what is measured, where every price comes from, and what we cannot know.
The bench
- Detect. WebGPU adapter info, plus the WebGL renderer string as a fallback for the model name. We normalise it to a key like
rtx-4090. If we can't match it, you pick your card. - Compute. A WGSL compute shader runs long chains of
fmaonvec4<f32>across ~1M threads. We calibrate the work so each dispatch is short (no frozen tab), warm up, then take the median of 5 runs. Timed with GPU timestamp queries when the browser allows, otherwise withonSubmittedWorkDone()wall clock. - Memory. A storage-buffer stream kernel copies up to 256 MB; bytes read + written over time = GB/s, median of 5 runs.
- Budget. Under 30 seconds total. Without WebGPU (some Firefox / Linux setups) we say so and you can still pick a card and get live prices.
- Browser ≠ native. WebGPU numbers are lower than native CUDA/ROCm numbers. We show the spec-sheet number next to yours and never call the browser number native.
Prices
Prices are what renters pay on Vast.ai right now: dph_total of rentable, on-demand, single-GPU offers, collected every 10 minutes. We keep count, min, p25, median and p75 per card. History exists only from when our collector started — nothing is backfilled.
A host takes home less than the renter price. We don't hard-code a fee: the marketplace cut field is your input (default 0). Vast has said live prices are typically about 25% above what hosts earn — if you want to model that, try a cut around 20.
Electricity: board power (typical, from public spec sheets — editable) × hours × your $/kWh. The default 0.15 $/kWh is an example; change it.
HTTP API
JSON, CORS open, times in ISO-8601 UTC. Base: https://benched.world
GET /api/prices # every card: live stats + spec
GET /api/prices/:model?range=7d # one card + hourly history
GET /api/quote?model=rtx-4090&hours=12&watts=450&kwh=0.15&cut=0
GET /api/value # ranked by $/TFLOP-hour and $/GB-VRAM-hour
GET /api/board # on-chain board
GET /api/token # {ca: null} until launch
POST /api/attest # plausibility check + EIP-712 signature
MCP server
Tools: gpu_price, rent_quote, cheapest_compute, compare_gpus, board_top. Every result carries source and fetchedAt.
Claude Code (HTTP)
claude mcp add --transport http benched https://benched.world/mcp
Claude Desktop (stdio, from a checkout of the repo)
{
"mcpServers": {
"benched": {
"command": "node",
"args": ["/path/to/256/server/bin/mcp-stdio.js"],
"env": { "BENCHED_API_URL": "https://benched.world" }
}
}
}
The board
A wallet can record its result on BenchedBoard on Robinhood Chain. The server checks the numbers are physically plausible for the card (measured ≤ 1.15 × spec, above zero, known model) and signs them; your wallet submits. The contract keeps the best score per wallet and card. It holds no funds.
Contract (Robinhood Chain, chain id 4663): — view on Blockscout (verified source).
Limits / not audited
- Not audited: neither the server nor the contract.
- Prices are from one marketplace (Vast.ai), on-demand, single-GPU offers, at fetch time. Not what a host earns, and not a promise anyone will rent your card.
- Spec numbers are vendor peak; browser WebGPU measurements are lower than native.
- Browser benchmarks can be spoofed. The board is for fun and comparison, not proof of hardware.
- Other tabs, power settings, thermals and the 3D scene itself can lower your result. Run again with nothing else open for your best number.
- We don't run a compute network, don't pay anyone for jobs, and nothing here is income advice.