X Compute
The network is open

Prove what your GPU can really do.

A 45-second benchmark, right in your browser. Get a verified score, your rank against identical hardware, and the AI models your machine can run.

How it works

No install. No account. Nothing personal leaves your browser.

$$XCOMP is coming to Robinhood ChainRuns made before launch earn Genesis points. See what the token unlocks.
—
Runs measured
—
Verified
—
Operators
—
Device classes

Live from the network. Every figure on this site is a real run; nothing is simulated.

How it works

Measured, not claimed.

Spec sheets say what hardware should do. This says what yours does, from the browser you already have open.

01

Measure

Real WebGPU compute kernels move 512 MiB through your GPU's memory and multiply dense matrices, timed by the GPU's own clock.

02

Verify

A signed, single-use ticket opens every run. The server recomputes the score and checks it against what the hardware can physically do.

03

Rank

Your result lands in its device class, never against different hardware. It becomes a die on the public wafer and earns points.

The run

Six phases. One number.

Each phase isolates one thing an AI workload asks of your hardware. Together they make the X-Score.

Read

GB/s

Generating one token reads every active weight once. Read throughput is the hard ceiling on tokens per second.

Can I run it?

From a score to a model.

Pick a machine and see which open models fit in its memory, and the fastest they can possibly decode.

Open the full planner
Weights

Most popular, about 4.85 bits per weight.

Memory
12 GB
Published peak
504 GB/s
3 of 7 models run on this machine
  • Llama 3.2 3B
    Fits3.2 of 12 GB
    259tok/s
  • Llama 3.1 8B
    Fits6.4 of 12 GB
    104tok/s
  • Llama 3.3 70B
    Too large45.9 of 12 GB
    —
  • Phi-4
    Fits10.7 of 12 GB
    56.6tok/s
  • Gemma 2 27B
    Too large18.7 of 12 GB
    —
  • Qwen 2.5 32B
    Too large22.2 of 12 GB
    —
  • gpt-oss-20b3.6B active
    Too large14.7 of 12 GB
    —

The ceiling is memory bandwidth divided by the bytes of weights read per token. It is an upper bound from the memory side; real runtimes land below it. Memory need is the weights plus an estimate for the KV cache and runtime.

Leaderboard

Ranked against its own kind.

A laptop never competes with a workstation card. Every board is one device class, so a rank says something about your machine, not your budget.

See every class
Latest runs

Runs appear here the moment they are verified.

$XCOMP

A token with a job.

$XCOMP is wired into the product: it multiplies what a run earns, unlocks the pro tools, and pays for certificates and compute as the network grows.

Explore $XCOMP
HoldLive at token launch

Multiply every run

Link a wallet that holds $XCOMP and every verified run earns up to 2× the points. The balance is read on-chain when the run is submitted.

UnlockLive at token launch

Pro tools for holders

Deep reports with per-sample traces and export, API keys with raised limits, and fleet mode to track every machine you own.

EarnIn development

Season reward pools

Each weekly season has a fixed pool, split by the points earned that week. Boosted points mean a larger share of the same pool.

SpendIn development

Hardware certificates

Spend $XCOMP to anchor a verified result on-chain as a certificate: proof of what a machine does, for resale listings and rental marketplaces.

SupplyOn the roadmap

Verified compute

Machines that prove their performance can serve inference jobs. Buyers pay in $XCOMP, and a verified score is the entry ticket.

VoteOn the roadmap

Set the rules

Holders vote on season themes, point rules and the reference machine behind the score.

Questions

Asked often.

Is the benchmark real, or an animation?+

Real. It compiles compute shaders, allocates 512 MiB on your GPU and times hundreds of passes with the GPU's own clock. The 3D wafer only draws between timed passes, so it never touches a measurement.

What do I need to run it?+

A recent Chrome, Edge or other browser with WebGPU. Browsers without it fall back to a CPU measurement: you still get a score, but it is not ranked.

Can results be faked?+

Browser numbers cannot be proven honest. They can be made hard to fake by accident and awkward to fake on purpose: signed single-use tickets, a minimum run time, a score recomputed on the server, and a ceiling set by the hardware's own memory bus.

What do you collect?+

The scores, the GPU model your browser reports, and your browser and OS family. No account, no cookies, no IP address stored, no fingerprint. Your identity is a random key that lives in your browser.

What does $XCOMP actually do?+

It multiplies the points a verified run earns, unlocks the pro tools, and is the unit for season pools, certificates and compute as they ship. Each function is labelled with its real status on the $XCOMP page.

Are points worth money?+

No. Points are a ledger of verified work. They decide how a funded reward pool is split, and a pool pays only once it has been funded on-chain. Nothing here is a promise of return.

Your turn

Forty-five seconds. One number.

Find out what your machine is worth to an AI model.