Folding@home for Apple Silicon Mac GPUs

People have asked for Folding@home on Apple GPUs since 2019 (openmm#2489) and 2024 (fah-client-bastet#303). Here it is. minicore runs Folding@home's own OpenMM work units end to end on the Mac's GPU, through OpenMM's Metal platform, and folds them to the standard of Folding@home's own core.

A FAH work unit goes into minicore, which runs it through OpenMM Metal with df64 mixed precision on the Apple GPU and writes FAH's result files, with the state tests — FAH's own work unit in, FAH's own result files out, all on the Mac.

Checked against FAH's core 0x27. 21 of 21 state tests passed on a whole real FAH unit (P18264, 919,221 atoms, 5 ns), as FAH's core 0x27 passed 21. Six files in the result, in core 0x27's order; viewerTop.json, xtcAtoms.csv.bz2 and checkpointIntegrator.xml byte-identical to core 0x27's. 101 of 101 core-parity checks passed, and 65 of 65 failure and stop-resume cases. 0.054 kJ/mol/nm worst state-test force RMSE, beside CUDA's 0.052, against FAHBench's limit of 5. 16 s to fold the bundled test unit on an M3 Ultra after ./build.sh

The five checks as text

	Checked against FAH's core 0x27
21 of 21	state tests passed on a whole real FAH unit (P18264, 919,221 atoms, 5 ns), as FAH's core 0x27 passed 21
Six files	in the result, in core 0x27's order; viewerTop.json, xtcAtoms.csv.bz2 and checkpointIntegrator.xml byte-identical to core 0x27's
101 of 101	core-parity checks passed, and 65 of 65 failure and stop-resume cases
0.054	worst state-test force RMSE in kJ/mol/nm, beside CUDA's 0.052, against FAHBench's limit of 5
16 s	to fold the bundled test unit on an M3 Ultra after `./build.sh`, so you can check it yourself (Try it)

Accuracy

Worst state-test force and energy error per test system against the Reference platform, Metal and CUDA — Worst state-test errors, Metal beside CUDA, far inside FAHBench's limits.

Energy drift of DHFR in NVE, Metal on the old fork before and after #5492, Metal on upstream OpenMM 9e320c6c3 without and with patch 0005 and on this build, c403e95b3 with it, CUDA mixed and CPU — Energy drift in NVE. This build drifts -0.0005 ± 0.0007 kJ/mol/ps over 10 runs, beside CUDA's -0.0001 ± 0.0011 (Welch t -0.31), or -4.1e-6 kT per ns per degree of freedom. The old pin, 9e320c6c3, drifted -0.0010 ± 0.0007 with the same patch (Welch t +0.50 against this build). Upstream OpenMM 9e320c6c3 without patch 0005 drifts -0.0050 ± 0.0009 over 10 runs (Welch t -3.50 against CUDA). The fork this repo first built sits beside CUDA too (Welch t -0.73).

Distance from FAH's trajectory against the distance between two CUDA seeds — Metal parts from FAH's trajectory as fast as two CUDA runs part from each other.

Parity with Folding@home's core

Parity with core 0x27 item by item, minicore on Apple Silicon

Every behaviour FAH's client acts on matches core 0x27. The rest needs FAH's own signing keys or is a format the client never reads. PARITY.md lists every row.

Speed

Same card, like for like

On one NVIDIA RTX 5060 Ti, minicore built for CUDA against FAH's core 0x27, over the same first 11% of P18264:

Core	OpenMM	Per 1%
FAH's core 0x27	8.2.0	187 s
minicore, CUDA	8.2.0	184 s
minicore, CUDA	8.6.1	180 s

Seconds per 1% over the first 11% of P18264 on one RTX 5060 Ti, FAH's core 0x27 against minicore built for CUDA — minicore keeps pace with FAH's own core on FAH's own hardware.

On a 5,000-step cut on that card, minicore's viewerTop.json and xtcAtoms.csv.bz2 were byte-identical to FAH's core's. Run conditions are in EVIDENCE.md.

On Apple Silicon

The M3 Ultra folded the whole unit in 5.06 h, 182 s per 1%. Run conditions are in EVIDENCE.md.

FAQ

Is it as accurate as Folding@home's core?

Yes. It passed all 21 of the unit's state tests, as FAH's core did, and its errors sit with CUDA's. Its energy drift in NVE is -0.0005 ± 0.0007 kJ/mol/ps over 10 runs, -4.1e-6 kT per ns per degree of freedom. CUDA's is -0.0001 ± 0.0011, and 10 runs each don't tell the two apart (Welch t -0.31, p 0.76).

That takes patch 0005. Upstream OpenMM wraps molecules back into the box by the double precision box, which the Metal kernels never use in mixed precision (they see it rounded to float), so each reorder changes every pair length across that face, measured at +0.002 kJ/mol per reorder; patch 0005 wraps by the float box instead. Without it, upstream 9e320c6c3, the pin before c403e95b3, drifts -0.0050 ± 0.0009 over 10 runs (Welch t -3.50 against CUDA); that arm was measured on 9e320c6c3 only. I found the cause in lab lane pin-drift-bisect (lab commit cdf10b66) and measured 10 runs of 9e320c6c3 with the patch in lane pin-drift-confirm (lab commit 26a0344c). The 10 runs of this build, c403e95b3 with the patch, are lanes pr5524-check and pin-c403-drift (lab commit 3de1b173, raw energies in pin-c403-drift/results/drift/s5..s10 and pr5524-check/results/drift/s1..s4). EVIDENCE.md has the tables.

Can Folding@home use the Apple Silicon GPU?

Yes. minicore runs Folding@home's real work units on M1 and later GPUs.

Can I run Folding at home on my Mac's GPU?

Not through the official client yet: it folds on the Mac's CPU only. fah-core-metal is the GPU core that would change that, and you can test it on your Mac today with ./report.sh.

How fast is it?

On one RTX 5060 Ti, minicore built for CUDA took 184 s per 1% of P18264's first 11% on OpenMM 8.2.0 and 180 s on 8.6.1, where FAH's core 0x27 took 187 s on the same card. On Apple Silicon, the M3 Ultra folded the whole unit in 5.06 h.

Can I fold for Folding@home points with it today?

No. FAH signs its cores and their results with its own private key, so that step is FAH's.

The source, how to build and run it, and the evidence for each number are in the repository: the README and docs/EVIDENCE.md. Packaging it for FAH's client on macOS is in docs/PACKAGING.md.

Ran it on your Mac? Post your result, whether it passed or not.