BENCHED

WEBGPU BENCH · LIVE

Your GPU is
being benched.

No install, no wallet. We measure this card in your browser and price it against what renters pay right now.

Detecting GPU… starting
FP32 measured—TFLOPS
Memory measured—GB/s
Browser numbers run below native CUDA numbers.

01 · SHROUD + FANS

Heat is
the bill.

Every watt the card draws is paid for at your wall. We use the card's typical board power and your electricity price.

Board power
—
Power cost / day
—

02 · FIN STACK

Copper glows
with the watts.

The heat-pipe colour in the scene follows board power. Fans spin with the benchmark's load, not on a loop.

03 · BACKPLATE

Look
underneath.

The sticker under the backplate carries your bench time in UTC. It is blank until your card has been measured.

04 · PCB

Memory moves
the money.

Pulses on the copper traces run at a speed set by the bandwidth we measured: a storage-buffer stream kernel, median of five runs.

Measured
—
Trace speed
idle

05 · VRAM

Renters pay
for memory.

Big models need big VRAM, so the market prices memory as much as compute. Cheapest live offers per GB of VRAM:

  1. Loading live prices…

06 · SILICON

The die,
up close.

Spec sheets quote native FP32 peak. A browser shader gets a share of that. We show both, side by side, and never call the browser number native.

Measured in browser
—
Spec sheet
—

07 · INSIDE THE DIE

A hall of
cores.

Lit cores = your measured TFLOPS against the fastest card in our table. Before a result, the hall idles with a neutral scan.

08 · OUT AGAIN

What it would rent for.

Card—
Offers live—
Prices fetched—
Renters pay / day—
Your power / day—
Margin / day—

Price = what renters pay on Vast.ai (dph_total, on-demand, 1-GPU offers). Host take-home is lower: set the marketplace cut yourself. Electricity default is an example — change it. Source

LIVE VALUE RANKING

Cheapest compute right now.

#Card$/TFLOP-hourMedian $/hour
Loading live prices…

FOR AGENTS · API + MCP

Ask what a card rents for.

Same live numbers over plain HTTP and an MCP server: gpu_price, rent_quote, cheapest_compute, compare_gpus, board_top.

Docs + install
# live median rent for a 4090, 12 h/day
curl "https://benched.world/api/quote?model=rtx-4090&hours=12&watts=450&kwh=0.15&cut=0"

ON-CHAIN BOARD

Best scores, per wallet.

The attester only checks numbers are physically plausible for the card. Browser benchmarks can be spoofed; the board is for fun and comparison, not proof of hardware.

Open the board
  1. Loading…