CoralGemm - Matrix multiply stress test for AMD ROCm

From Define Wiki
Jump to navigation Jump to search
# rocm 6.0.2, mi210, ubuntu 2204

  git clone https://github.com/AMD-HPC/CoralGemm.git
  cd CoralGemm/
  cd src/
  make

# execute
# dgemm mi210 64GB device
./gemm R_64F R_64F R_64F R_64F OP_N OP_T 8640 8640 8640 8640 8640 8640 36 300

# sgemm 64GB device 
./gemm R_32F R_32F R_32F R_32F OP_N OP_T 8640 8640 8640 8640 8640 8640 72 300

# Simultaneous execution , load on the 8 GPUs :
./gemm R_64F R_64F R_64F R_64F OP_N OP_T 4224 3840 9216 4224 3840 4224 18 24 strided

# Expected results : 23.5-24.5 TF per GPU (i.e 47-49 TF per MI250)