TT-Bio benchmarks: biomolecular models on Tenstorrent

What TT-Bio runs

Structure and binding affinity prediction

--model	Folds	Affinity	MSA
boltz2	protein, DNA, RNA, ligand	yes	on by default
protenix-v1	protein, DNA, RNA, ligand	no	on by default
protenix-v2	protein, DNA, RNA, ligand	no	on by default
openfold3	protein, RNA, DNA	no	on by default
openbind	protein, RNA, DNA, ligand by SMILES or CCD	no	on by default
esmfold2	single protein chain	no	not needed
opendde	multi-chain protein complexes	no	on by default
rf3	protein, nucleic acid, ligand	no	on by default
af2ig	a designed complex, from its own coordinates	no	not needed
nesso1	nothing, it returns scalars	yes	not needed

Protenix-v1, Protenix-v2 and OpenFold3 are AlphaFold3-family. Protenix-v1 is upstream's own v0.5.0 base checkpoint: half the pair width and 4 trunk recycles against Protenix-v2's 10. OpenDDE runs the Protenix-v2 stack plus a structural-token expander. OpenBind-0 is the same OpenFold3 stack on upstream's v0.5.0 checkpoint, tuned for protein-ligand co-folding. AF2-IG takes a target plus a binder backbone and the binder's sequence, and reports ipTM and interface pAE for the design; it is the selection filter in a binder pipeline, not a folder you hand a sequence to. Nesso-1 runs under tt-bio affinity rather than predict: no coordinates, no PAE, just an affinity value and ten other scalars. esmfold2-fast is lighter, opendde-abag is the antibody-antigen checkpoint.

Protein design

--model	Model	Designs	Input
boltzgen	BoltzGen	protein, peptide, nanobody, antibody binders	target structure in a design YAML
rfd3	RFdiffusion3	all-atom binders, motif scaffolds, nucleic-acid binders	JSON spec with contig strings
pxdesign	PXDesign	binder backbones, no sequence	target structure in a design YAML

PXDesign conditions on a distogram of the target only and writes the binder as GLY, because it generates a backbone with no sequence.

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

Derived

Throughput per dollar of total cost of ownership

Derived

Predictions and designs per hour, per server

Derived

What the server costs

Price and power

One AI Processor the measured seconds per prediction and per design that every server figure above is built from

Prediction time, one AI Processor

Measured

Seconds per design or affinity prediction, one AI Processor

Measured

The exact figures the measured seconds behind every bar, and the derived tables

Measured, seconds per prediction on one AI Processor

Measured, host and device inside one prediction

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 was run, how cost is computed, where every price and power figure comes from

Sources

Caveats

Every seconds-per-prediction and seconds-per-design figure is measured on the hardware named. Everything else is derived from those measurements and from the published prices and power ratings listed above. Nothing is estimated. Updated .

Every model here reproduces its reference implementation within that reference's own run-to-run noise.

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.