Benchmarks
Every model, measured on both sides
What TT-Bio runs, how fast each model is on Tenstorrent and on NVIDIA, what that costs per prediction, and how every number was taken.
What TT-Bio runs
Structure and binding affinity prediction
| --model | Folds | Affinity | MSA |
|---|---|---|---|
| openfold3 | protein, RNA, DNA | no | on by default |
| boltz2 | protein, DNA, RNA, ligand | yes | on by default |
| opendde | multi-chain protein complexes | no | on by default |
| protenix-v2 | protein, DNA, RNA, ligand | no | on by default |
| rf3 | protein, nucleic acid, ligand | no | on by default |
| esmfold2 | single protein chain | no | not needed |
| openbind | protein, RNA, DNA, ligand by SMILES or CCD | no | on by default |
| protenix-v1 | protein, DNA, RNA, ligand | no | on by default |
| mol1 | protein, DNA, RNA, ligand | no | on by default |
| af2ig | a designed complex, from its own coordinates | no | not needed |
| nesso1 | nothing, it returns scalars | yes | not needed |
OpenFold3, Protenix-v2, Protenix-v1 and Mol-1 are AlphaFold3-family. AF2-IG scores a
designed complex rather than folding a sequence. Nesso-1 runs under tt-bio affinity and
returns scalars, not coordinates. esmfold2-fast is lighter, opendde-abag is
the antibody-antigen checkpoint.
Protein design
| --model | Model | Designs | Input |
|---|---|---|---|
| rfd3 | RFdiffusion3 | all-atom binders, motif scaffolds, nucleic-acid binders | JSON spec with contig strings |
| boltzgen | BoltzGen | protein, peptide, nanobody, antibody binders | target structure in a design YAML |
| rfd4 | RFD4-Proteina | binders, motif scaffolds, enzymes, antibody and nanobody CDRs, unconditional monomers | RFD4 design spec (YAML or JSON) |
| pxdesign | PXDesign | binder backbones, no sequence | target structure in a design YAML |
Protein embeddings
| --model | Embeddings | Output | Sizes, as --model |
|---|---|---|---|
| esmc | sequence, ESMC language model | per-residue, pooled | esmc-300m, esmc-600m, esmc-6b |
| saprot | structure-aware, ESM-2 over a fused amino-acid and Foldseek-3Di vocabulary | per-residue, pooled | saprot-35m, saprot-650m, saprot-1.3b |
SaProt's structure input is optional; sequence-only costs accuracy below 1.3B.
Performance and cost
Throughput per dollar of purchase price
DerivedThroughput per dollar of total cost of ownership
DerivedPredictions and designs per hour, per server
DerivedWhat the server costs
Price and powerOne AI Processor the measured seconds behind every server figure
Prediction time, one AI Processor
MeasuredSeconds per design or affinity prediction, one AI Processor
MeasuredThe exact figures every bar as a table
Measured, seconds per prediction on one AI Processor
Measured, seconds per design or affinity prediction on one AI Processor
Derived, predictions and designs per hour per server
Derived, throughput per dollar of purchase price
Derived, throughput per dollar of total cost of ownership
Methods and sources what ran, how cost is computed, prices and power
Sources
Caveats
Every seconds figure is measured on the hardware named. Everything else is derived from those measurements and the published prices and power ratings. Updated .
Tenstorrent, Blackhole and Galaxy are trademarks of Tenstorrent Inc. NVIDIA, DGX, H200, B200 and A100 are trademarks of NVIDIA Corporation. Hardware names are used to identify the systems measured.