We recently bought two Mac Mini with new apple M1 chip. Although Apple provide seamless simulator Rosetta 2 for Intel-based software, I wonder how the performance of new Apple M1 chip dose. So here is a simple test using Phenix and one crystallographic dataset collected in my lab. We also compared the computation speed between Mac and Linux, Desktop PC and Xeon workstation. This is not a professional benchmarks but a reference of some structural biologists to evaluate whether switching to M1 chip is a good decision. I hope open source programmers for structural biology will provide M1-based versions too. If Apple’s GPU can be used for computation, then we really don’t count on no-where-to-buy Nvidia GPU cards at all.

We use a HKL2000 indexed sca file (1.68 Å) and a PDB template for “molecular replacement” first. Phaser (version 2.8.3) with simple mode was used. The phased data was then sent to “Autobuild” using 6 parallel threads (actually only need 5). The best fitted auto-built model was submitted to “refinement” to refinement structures.

Here are the 5 different computers including desktop, laptop and workstations used for comparison. To be noted first, due to varied purchased time, I don’t use same Phenix version (1.17 to 1.19) to do all performance benchmarks.

	Linux-1	Linux-2	Macbook Pro 2018	Mac mini M1-2021	Mac mini -Intel 2019
OS	Ubuntu 18	Ubuntu 16	Mojave, 10.14	Big Sur, 11	Catalina, 10.15
CPUs	Intel i7, 3.2GHz	Intel Xeon silver, 2.2GHz	Intel i5, 2.3GHz	Apple M1, 3.2GHz	Intel i3, 3.2GHz
cores	6	12	4	4	4
threads	12	24	8	8	8

Below are snapshots of parameters used for Phaser (simple mode), autobuild and Refinement

The benchmarks are shown here.

	Linux-1	Linux-2	Macbook Pro 2018	Mac mini M1-2021	Mac mini, Intel-2019
Phenix	1.17.1-3660	1.17.1-3600	1.18.2-3874	1.19.2	1.18.2-3874
Phaser (seconds)	622	1046	604.5	3677-3900	533
Autobuild (minutes)	241	373	N/A	320	440
Refinement (seconds)	1334	2362	N/A	2470	3150

Few facts regarding the benchmarks and my observations.

• All Phaser tests returned consistent results that a single solution with LLG values greater than 65,000.
  
• The Phaser in Apple M1 is unknowingly slow. I did 3 times and the programs took 3600-3900 seconds to perform molecular replacement. I don’t know whether it was version differences of Phenix or M1 Rosetta2 simulation.
  
• In order to check whether Phenix 1.19.2 slows down Phaser, I installed Phenix 1.19.2 on the Linux-1 (Ubuntu 18), it took 600.2 seconds to finish molecular replacement. So it is unclear how Phaser on M1-based mac mini processes slowly.
  
• Autobuild always returns well built chains A and B. No significant differences between 5 test platforms. The Rfree and Rwork are always around 0.19 and 0.17, respectively.
  
• Autobuild performance speeds were similar at all 5 platforms. Xeon CPU has slower wall clock rates and it is indeed slow to do autobuild compared to Intel i7 desktop-grade CPU. Mac M1 seems to be a good replacement of Intel-based Mac products. Perhaps when more and more ARM-based native scientific programs written, the power of M1 chip will be more enhanced.
  
• The Refinement fine tunes structural quality to 96-98% favored dihedral angles in Ramachandran plots, no Cbeta outliers, no cis/twist residues, no rotatmer outliers. Molprobity clash scores are 3-5.