3 ms·

I'm having a hard time finding mips/flops for the M1 CPU, can you measure one of them?

by bullen 5y ago

brrrrrm 5y ago
~100G fp32 FMA/s per core on my M1 air last I checked.
- bullen 5y ago
  Is that CPU or GPU? Sounds a bit high when 2Gflops/watt (but I guess at 40W for all cores =2,5Gflops/W it could make sense?!) is the industry standard! Can you turn off the acceleration to get CPU only figures? What is FMA/s? Edit: Ok that is some sort of SIMD/GPU... I need figures for the raw CPU that can also do general purpose calculations and branches! Maybe MIPS is a better "apples to apples" measurement?
  - brrrrrm 5y ago
    Ah, it's actually 100gflops (not 100Gfmas) because the fmla instruction is fused. I just ran it on CPU, single core (no pinning), just using the aarch64 fmla instruction. Using pure BLAS (multithreaded, with the MM ISA extension), I get like 1.2TFlops in fp32 on the whole device. Code for the blas test: https://jott.live/code/blas_test.cc https://jott.live/code/blas_test.cc
    - bullen 5y ago
      So a friend that has a M1 did this test: https://github.com/brianolson/flops/blob/master/flops.c https://github.com/brianolson/flops/blob/master/flops.c And the 5nm M1 has ~2.5Gflops/W which is not a huge increase compared to the 28nm Pi 4 at 2Gflops/W. No-moores law in effect. Game Over!
    - bullen 5y ago
      Can you run it without SIMD (fmla)? Same as these tests: http://web.eece.maine.edu/~vweaver/group/green_machines.html http://web.eece.maine.edu/~vweaver/group/green_machines.html
  - brrrrrm 5y ago
    Yea I tested it SIMD (8 ops per instruction) and it's hitting 100gflops, so that'd be 12.5k MIPS, right?
ReaLNero 5y ago
Could you give me a benchmark in particular? Or maybe this one works: https://github.com/mmperf/mmperf https://github.com/mmperf/mmperf. I'll run it in an hour.
- bullen 5y ago
  So a friend that has a M1 did this test: https://github.com/brianolson/flops/blob/master/flops.c https://github.com/brianolson/flops/blob/master/flops.c And the 5nm M1 has ~2.5Gflops/W which is not a huge increase compared to the 28nm Pi 4 at 2Gflops/W. No-moores law in effect. Game Over!
- bullen 5y ago
  Can you try this one? Dunno how hard it would be to make it work... https://github.com/deater/performance_results https://github.com/deater/performance_results Just want something that is more "apples to apples" (no pun intended but yes... lol)

I'm having a hard time finding mips/flops for the M1 CPU, can you measure one of them?

by bullen 5y ago

I'm having a hard time finding mips/flops for the M1 CPU, can you measure one of them?

brrrrrm 5y ago

~100G fp32 FMA/s per core on my M1 air last I checked.

bullen 5y ago
Is that CPU or GPU? Sounds a bit high when 2Gflops/watt (but I guess at 40W for all cores =2,5Gflops/W it could make sense?!) is the industry standard! Can you turn off the acceleration to get CPU only figures? What is FMA/s? Edit: Ok that is some sort of SIMD/GPU... I need figures for the raw CPU that can also do general purpose calculations and branches! Maybe MIPS is a better "apples to apples" measurement?
- brrrrrm 5y ago
  Ah, it's actually 100gflops (not 100Gfmas) because the fmla instruction is fused. I just ran it on CPU, single core (no pinning), just using the aarch64 fmla instruction. Using pure BLAS (multithreaded, with the MM ISA extension), I get like 1.2TFlops in fp32 on the whole device. Code for the blas test: https://jott.live/code/blas_test.cc https://jott.live/code/blas_test.cc
  - bullen 5y ago
    So a friend that has a M1 did this test: https://github.com/brianolson/flops/blob/master/flops.c https://github.com/brianolson/flops/blob/master/flops.c And the 5nm M1 has ~2.5Gflops/W which is not a huge increase compared to the 28nm Pi 4 at 2Gflops/W. No-moores law in effect. Game Over!
  - bullen 5y ago
    Can you run it without SIMD (fmla)? Same as these tests: http://web.eece.maine.edu/~vweaver/group/green_machines.html http://web.eece.maine.edu/~vweaver/group/green_machines.html
- brrrrrm 5y ago
  Yea I tested it SIMD (8 ops per instruction) and it's hitting 100gflops, so that'd be 12.5k MIPS, right?

ReaLNero 5y ago

Could you give me a benchmark in particular? Or maybe this one works: https://github.com/mmperf/mmperf https://github.com/mmperf/mmperf. I'll run it in an hour.

bullen 5y ago
So a friend that has a M1 did this test: https://github.com/brianolson/flops/blob/master/flops.c https://github.com/brianolson/flops/blob/master/flops.c And the 5nm M1 has ~2.5Gflops/W which is not a huge increase compared to the 28nm Pi 4 at 2Gflops/W. No-moores law in effect. Game Over!
bullen 5y ago
Can you try this one? Dunno how hard it would be to make it work... https://github.com/deater/performance_results https://github.com/deater/performance_results Just want something that is more "apples to apples" (no pun intended but yes... lol)