clpeak icon indicating copy to clipboard operation
clpeak copied to clipboard

result for AMD RX 6600 in macOS 14.4

Open cfwen opened this issue 3 months ago • 3 comments

Platform: Apple Device: AMD Radeon RX 6600 Compute Engine Driver version : 1.2 (Feb 21 2024 21:44:18) (Macintosh) Compute units : 28 Clock frequency : 2750 MHz

Global memory bandwidth (GBPS)
  float   : 199.54
  float2  : 207.46
  float4  : 211.33
  float8  : 209.29
  float16 : 177.12

Single-precision compute (GFLOPS)
  float   : 4390.59
  float2  : 4382.19
  float4  : 4342.84
  float8  : 4324.32
  float16 : 4204.27

No half precision support! Skipped

Double-precision compute (GFLOPS)
  double   : 554.49
  double2  : 553.16
  double4  : 551.83
  double8  : 544.15
  double16 : 545.67

Integer compute (GIOPS)
  int   : 1735.16
  int2  : 1721.56
  int4  : 1717.33
  int8  : 1692.05
  int16 : 1706.22

Integer compute Fast 24bit (GIOPS)
  int   : 7663.00
  int2  : 7561.06
  int4  : 7510.31
  int8  : 7338.48
  int16 : 7299.52

Transfer bandwidth (GBPS)
  enqueueWriteBuffer              : 11.93
  enqueueReadBuffer               : 13.49
  enqueueWriteBuffer non-blocking : 12.64
  enqueueReadBuffer non-blocking  : 13.55
  enqueueMapBuffer(for read)      : 91.02
    memcpy from mapped ptr        : 10.43
  enqueueUnmap(after write)       : 34458.03
    memcpy to mapped ptr          : 10.52

Kernel launch latency : 5.57 us

cfwen avatar Mar 22 '24 00:03 cfwen

single precision performance is only half of the theoretic number

cfwen avatar Mar 22 '24 00:03 cfwen

for same hardware, in Ubuntu 20.04, single precision performance is close to theoretic performance

cfwen avatar Mar 22 '24 00:03 cfwen

Apple is not actively maintaining support for OpenCL

krrishnarraj avatar Apr 07 '24 05:04 krrishnarraj