search

nihui / vkpeak

A tool which profiles Vulkan devices to find their peak capacities
MIT License
99 stars 3 forks source link

measure fp16-matrix on intel graphics, reduce VRAM usage #13

Closed nihui closed 4 months ago

nihui commented 4 months ago

i5-12400

nihui@nihui-alpc:~/vkpeak/build$ ./vkpeak 0
device       = Intel(R) UHD Graphics 730 (ADL-S GT1)

fp32-scalar  = 260.38 GFLOPS
fp32-vec4    = 281.95 GFLOPS

fp16-scalar  = 661.49 GFLOPS
fp16-vec4    = 881.09 GFLOPS
fp16-matrix  = 101.24 GFLOPS

fp64-scalar  = 0.00 GFLOPS
fp64-vec4    = 0.00 GFLOPS

int32-scalar = 121.34 GIOPS
int32-vec4   = 124.15 GIOPS

int16-scalar = 259.06 GIOPS
int16-vec4   = 266.58 GIOPS

nihui@nihui-alpc:~/vkpeak/build$ ./vkpeak 1
device       = llvmpipe (LLVM 18.1.1, 256 bits)

fp32-scalar  = 96.01 GFLOPS
fp32-vec4    = 344.58 GFLOPS

fp16-scalar  = 18.62 GFLOPS
fp16-vec4    = 51.70 GFLOPS
fp16-matrix  = 0.00 GFLOPS

fp64-scalar  = 95.09 GFLOPS
fp64-vec4    = 206.31 GFLOPS

int32-scalar = 63.29 GIOPS
int32-vec4   = 191.45 GIOPS

int16-scalar = 83.67 GIOPS
int16-vec4   = 150.67 GIOPS