Open source
TT-Bio runs folding, binder design and protein embeddings on Tenstorrent. One Blackhole Galaxy matches an NVIDIA DGX B200 on throughput, for a fifth of the price.
Benchmarks
Predictions per hour, measured on the same 512-residue protein, with each stack on its own shipped settings. Then divided by what the hardware costs.
Every number here, and how it was measured
Per-model tables, the exact seconds behind every bar, run conditions, the cost model and every price and power source.
Coverage
Every model runs on the same stack, from the kernels on the chip up to the serving layer. Speed up one part and every model gains.
Structure prediction
Binder design
Embeddings
Accuracy
Every model reproduces its official reference implementation, within that reference's own run-to-run spread.
A single reference run is not a target. The reference itself varies between seeds, so we measure that spread and land inside it.
Submodule activations by correlation, end-to-end outputs on identical noise, and published benchmarks reproduced.
Per-target results, thresholds and reproduction commands are in the repository, along with the legs that do not yet pass.
Why it works
Serving is steady, predictable and large. Move it to Tenstorrent and your GPU fleet is left for training and model development, where it earns the most.
No sharding and no interconnect ceiling. A prediction runs on a single AI Processor, so throughput scales linearly with the number of them.
Agents port a model to the card in a day or two, on your machine. Your weights never leave you.
Blackhole pairs GDDR6 with large on-chip SRAM, so the HBM shortage does not gate delivery.
JapanFold runs TT-Bio on Tenstorrent Galaxy hardware as a public service. Fold something today, with no account and no card.