ChimeraX on Apple M1 CPUs

Tom Goddard
November 2, 2021

We are making a version of our ChimeraX molecular graphics program that runs natively on Apple's new M1 CPUs for faster interactive calculations.
We'll report some speed-up timings and describe difficulties porting from Intel to the Apple M1 CPU.
A native Apple M1 version of ChimeraX is not yet available, but we expect to release it within 6 months.

About ChimeraX

Display protein, nucleic acid, and ligand atomic models, sequences and microscopy data.
Compute surfaces, hydrogen bonds, align structures, molecular dynamics
Building models in 3D electron microscopy maps.
A hundred other tools for analysis of molecular structures.
About 1/3 of ChimeraX downloads are for Mac (Windows 23000 (51%), Mac 15500 (34%), Linux 6400 (14%) past year)

New Mac Computers

Most new Mac computers use Apple's M1 CPU
- All new Mac laptops, Mac mini, iMac 24" use Apple M1 CPUs
- iMac 27" and Mac Pro still use Intel CPUs
- In about 1 more year Apple will not sell any Intel CPU machines.
Mac computers with M1 CPUs can run all older Intel Mac applications using Rosetta 2 emulation.
Support for Intel Mac hardware and software will continue for many (5?) years.

Speed of Apple M1 ChimeraX vs Intel ChimeraX

ChimeraX is 70% Python and 30% C++, the C++ is used for speed.
C++ code runs about 1.1x - 1.5x longer with Intel emulation on Apple M1.
Python code runs about 1.5x - 1.9x longer with Intel emulation on Apple M1.
Apple M1 in Mac Mini is about 1.3x - 2x faster than 2019 MacBook Pro Intel i9 CPU.
Apple M1 Pro GPU is about the same speed as 2019 MacBook Pro discrete AMD Vega 20 GPU.
Render speed comparison, Windows Nvidia GTX 1080 Ti (released 2017) is about 3x faster than M1 Pro.

Calculation¹	M1 CPU²		M1 Pro CPU³		Intel CPU⁴	Notes
Calculation¹	M1 ChimeraX	Intel ChimeraX	M1 ChimeraX	Intel ChimeraX	Intel ChimeraX	Notes
open PDB 3j3q	3.8 sec	5.6	4.0	6.0	6.9	Parsing 230 MByte text file in C++. 2.4 million atoms.
7n30 nucleotides slabs	6.7 sec	12.6	6.6	13.0	14.1	Python calculation. 150,000 atom ribosome structure.
addh on 7n30	18.0 sec	28.7	16.9	28.8	38.7	Mixed Python and C++ calculation.
compute surface for 7n30	8.9 sec	9.6	8.0	9.5	11.4	C++ calculation. Multithreaded.
Map contour calculation EMDB 22657	0.67 sec	1.02	0.67	1.02	1.17	C++ calculation on 300 MByte data set. Apoferritin 420x420x420.
Rendering surface 7n30 full lighting	12.8 frames/sec	11.8	25.0	25.7	27	Graphics rendering speed test. Intel CPU column has AMD Vega 20 discrete GPU.

Timings done using the ChimeraX time command. Benchmark details.
2021 Mac Mini (MacMini9,1), macOS 11.6, Apple M1 CPU, 16 Gbytes memory.
2021 MacBook Pro 16" (MacBookPro18,1), macOS 12.0.1, Apple M1 Pro CPU, 16 Gbytes memory.
2019 MacBook Pro (MacBookPro15,3), macOS 11.6, 32 Gbytes memory, Intel i9 8-core 2.4 GHz CPU, AMD Vega 20 GPU.

Difficulties porting ChimeraX to Apple M1 CPUs

ChimeraX Python and C++ code needs no changes.
ChimeraX uses 90 packages developed by others.
- 60 are pure Python from the PyPi repository.
- 30 are binary packages that need Apple M1 versions.
- 6 binary packages do not have Apple M1 distributions: ambertools, h5py, imagecodecs, netcdf4, pytables, scipy.
Qt 6 window toolkit is distributed for Apple M1 but not Qt 5.
- ChimeraX uses Qt 5, the stable Qt version from 2012 - 2021.
- Qt 6 with html support was released September 2021.
Apple M1 applications must be either all native M1 binaries or all Intel binaries, no mixing.
Need to distribute either a large univeral package that includes both Intel and M1 binaries, or two separate ChimeraX versions.

Potential advantages of native Apple M1 ChimeraX

Better OpenGL driver stability with Apple M1 GPU.
- No graphics driver crashes among 43 ChimeraX bug reports in 2021 with Apple M1.
- About 100 ChimeraX graphics driver crashes reported on Intel Macs in past 2 years.
Better C++ crash stack traces with native M1 app than with Intel emulation.
- Intel ChimeraX crashes on M1 often give no C++ stack trace.