Changelog

• 24/08/20

  • Replaced some links with permanent ones (Internet Archive)

  • Added additional links.

  • Added clarifications to the SME introduction.

  • Updated all E-Core performance numbers (QOS_CLASS_UTILITY instead of QOS_CLASS_BACKGROUND).

  • Added performance for SME2 multi-vector group FMLA.

  • Updated multithreading numbers (up to 10 threads).

  • Replaced copy bandwidth benchmark with separate read/write benchmarks.

  • Data: bandwidth_p_1_2d9566d.log, micro_p_1_2d9566d.log, micro_p_2_2d9566d.log, micro_p_3_2d9566d.log, micro_p_4_2d9566d.log, micro_p_5_2d9566d.log, micro_p_6_2d9566d.log, micro_p_7_2d9566d.log, micro_p_8_2d9566d.log, micro_p_9_2d9566d.log, micro_p_10_2d9566d.log, micro_e_1_2d9566d.log, micro_e_2_2d9566d.log, micro_e_3_2d9566d.log, micro_e_4_2d9566d.log, micro_e_5_2d9566d.log, micro_e_6_2d9566d.log, micro_e_7_2d9566d.log, micro_e_8_2d9566d.log, micro_e_9_2d9566d.log, micro_e_10_2d9566d.log

• 24/06/15

  • Added results of copy benchmark.

  • Data: copy_p_1_61d3860.log.

• 24/05/26

  • Added C+=AB^T kernel with M=N=K=32.

  • Data: gemm_p_1_23efa23.log, gemm_p_2_23efa23.log, gemm_p_3_23efa23.log, gemm_p_4_23efa23.log.

• 24/05/21

  • Added outer-product microbenchmarks for FP64, BF16-BF16-FP32, FP16-FP16-FP32, INT8-INT8-INT32 and INT16-INT16-INT32.

  • Added missing multi-core results on 4, 5 and 6 e-cores.

  • Data: micro_p_1_7e5e61b.log, micro_e_1_7e5e61b.log, micro_e_4_499197c.log, micro_e_5_499197c.log, micro_e_6_499197c.log.

• 24/05/18

  • Added AMX peak.

  • Added performance results for changing SME Pstate.

  • Added multi-core results.

  • Pushed code to GitHub.

  • Data: micro_p_1_499197c.log, micro_p_2_499197c.log, micro_p_3_499197c.log, micro_p_4_499197c.log, micro_e_1 499197c.log, micro_e_2_499197c.log, micro_e_3_499197c.log.

• 24/05/17:

  • Added performance results for efficiency cores.

  • Added microbenchmark for tile reuse.

  • Added benchmark results for Acccelerate’s GEMM performance.

  • Data: driver, kernels, logs p-core, logs e-core.

• 24/05/15

  • Added introduction.

  • Added first microbenchmarks measuring the peak of M4’s performance cores when using Neon, Streaming SVE and SME.

  • Data: driver, kernels, logs.