$XCOMP is live.
Proof of Compute: a 45-second benchmark that runs in your browser and proves what your GPU can really do.
Verified score. Rank against identical hardware. The AI models you can run.
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
https://t.co/J55bfJCjSf
One signature. Zero transactions.
Linking a wallet to your runs is a signed message. No gas. No approval. Nothing leaves your wallet.
Unlink it whenever you want. Your points stay.
https://t.co/J55bfJCjSf
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
Every run gets its own page.
A permanent public link with the score, the six phases and the verdict. Paste it in a listing, a forum, a group chat.
Selling a GPU? Show what it does, not what the box says.
https://t.co/J55bfJCjSf
A score is a summary. The deep report is the evidence.
Every phase, sample by sample. The spread between the 5th and 95th percentile. Your run against the median of its class. Export to JSON or CSV.
Open to $XCOMP holders.
https://t.co/J55bfJCjSf
No WebGPU in your browser? You still get a number.
With WebGPU, the run is measured on the GPU and ranked.
Without it, the same six phases run on the CPU: scored, never ranked.
Either way, the result says which path it took.
https://t.co/J55bfJBM2H
Real workloads don't read memory in a straight line.
Embedding lookups and sparse layers jump around. So one phase makes 16 million scattered reads per pass, at addresses no prefetcher can guess.
Fast streaming and fast jumping are two different machines.
https://t.co/xTpEWDJ3Ph
Memory feeds the model. Math runs it.
Attention and MLP layers are matrix products, so one phase multiplies 512 × 512 matrices: 268 million operations per pass, counted in GFLOP/s.
It weighs 22% of your X-Score.
https://t.co/J55bfJBM2H
The data is public.
Scores, device classes, medians, daily proofs: all of it is one request away. No API key. No account.
GET /api/leaderboard
GET /api/stats
GET /api/proof
Build on it.
https://t.co/J55bfJCjSf
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
Every verified run earns points.
+100 for the run
+100 for your first run on a GPU model
+250 if that model is still rare on the network
+500 for your first run on the Genesis wafer
Capped per day, so they can't be farmed.
https://t.co/J55bfJCjSf
$XCOMP is live.
Proof of Compute: a 45-second benchmark that runs in your browser and proves what your GPU can really do.
Verified score. Rank against identical hardware. The AI models you can run.
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
https://t.co/J55bfJCjSf
CA added on the website.
The official $XCOMP contract is now live on https://t.co/J55bfJCjSf, with price, market cap and liquidity read in real time.
CA: 0x8ef46Aa084Cbe825D8060232322e9464fEab5721
Always check the address before you trade.
https://t.co/J55bfJCjSf
One lucky sample doesn't make a fast GPU.
Every phase takes dozens of samples, throws out the outliers and reports the median.
Run it twice and you land on the same score. That is the point of a benchmark.
https://t.co/J55bfJCjSf
Small benchmarks lie. They fit in the cache and measure the cache.
Ours moves 512 MiB: two 256 MiB buffers, touched in turn. A byte is only seen again after half a gigabyte of other traffic.
What you get is your memory's real speed.
https://t.co/J55bfJBM2H
New on X Compute: duels.
Think your GPU beats your friend's? Prove it.
1. Run the benchmark
2. Hit "Challenge a friend"
3. Send the link
They run it, and you both get a duel page: two scores, six phases, one winner.
https://t.co/xTpEWDJ3Ph
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
A new season every Monday.
Bandwidth week. Compute week. Laptop week. Endurance week. Open week.
Each one ranks a single thing, always against the median of your own device class. A laptop can top the board.
https://t.co/J55bfJCjSf
Hold. Don't stake.
$XCOMP holders earn more points per run: ×1.25, ×1.5, up to ×2.
Your tier is read from your wallet when a run is submitted. Nothing is locked. Nothing leaves your wallet.
https://t.co/J55bfJCjSf
CA: 0x8ef46aa084cbe825d8060232322e9464feab5721
A chat lasts longer than a benchmark.
So the last phase holds the load for 14 seconds and compares the end with the start.
Some machines hold 99%. Some drop once they heat up. Now you know which one you own.
https://t.co/J55bfJBM2H
Every GPU has a speed limit: memory bus width × data rate ÷ 8.
RTX 5090: 1,792 GB/s RTX 4090: 1,008 GB/s RTX 4070: 504 GB/s
No honest run can beat its own bus. One that claims to is thrown out.
https://t.co/J55bfJBM2H
The X-Score has no secret sauce.
It is a weighted geometric mean of six phases. The reference machine scores 1,000. A machine twice as fast scores 2,000.
The server recomputes it for every run. Your browser's number is never trusted.
https://t.co/J55bfJBM2H
A benchmark that animates while it measures is measuring its own animation.
Ours doesn't. While a pass is timed, the page holds still. It draws one frame, then the next pass starts.
Every sample is timed by the GPU's own clock.
https://t.co/J55bfJCjSf