Vote: pick the GPU + model I optimize next
Pick the GPU, the system RAM and the model I work on next. The interesting ones are RAM-starved: a model bigger than the GPU's memory, often an MoE, where the engine streams weights from system RAM and the usual numbers fall apart. That's why system RAM is its own question.
1 of 4
GPU
Every card is a GitHub issue. Vote with a ๐ on the issue, fund it on Polar, or both. At the cutoff, the card with the highest score is the one I work on, and the results land on this site.
Round 1 ยท cutoff 2026-12-31
No cards yet. Pick a combo above and start the first one.
How a card wins
Score = ๐ votes + โ(dollars funded). One vote per GitHub account. Money helps, but the square root means no single sponsor can buy a round: $100 counts as 10 votes, $10,000 as 100.
At the cutoff the highest score wins. Funding for every other card stays on its issue and carries into the next round. Nothing is spent until a card wins.
What the winning card gets
- A BoltBeam audit of the card: every role and kernel of a token against its roofline, published here.
- The work: engine patches like the llama.cpp mods, tuned settings, or a new kernel, whatever closes the most of the gap.
- Before-and-after numbers on this site, measured the same way as everything else, with the exact engine build and date.
- Everything released under MIT.
How the money works
Funding goes through Polar. Polar holds it on the issue and pays out only when the issue is closed with the work done. If a card never wins, its funding never pays out. Sponsors are credited on the results if they choose to be.
This is paying for work, not a donation: there is no charity and no tax deduction.