results for NVIDIA GeForce RTX 4090 (overclocked) in Windows 11

[moyang]

Core: +200 MHz, VRAM +1000 MHz, Power limit: 600W

Platform: NVIDIA CUDA
Device: NVIDIA GeForce RTX 4090
Driver version : 531.61 (Win64)
Compute units : 128
Clock frequency : 2520 MHz

Global memory bandwidth (GBPS)
  float   : 954.06
  float2  : 983.28
  float4  : 1001.36
  float8  : 1013.59
  float16 : 1017.66

Single-precision compute (GFLOPS)
  float   : 90262.02
  float2  : 85753.74
  float4  : 90346.06
  float8  : 89091.80
  float16 : 89121.65

No half precision support! Skipped

Double-precision compute (GFLOPS)
  double   : 1496.53
  double2  : 1494.58
  double4  : 1488.91
  double8  : 1482.93
  double16 : 1470.34

Integer compute (GIOPS)
  int   : 46283.53
  int2  : 46459.81
  int4  : 45872.11
  int8  : 46332.95
  int16 : 46330.86

Integer compute Fast 24bit (GIOPS)
  int   : 46572.11
  int2  : 46336.07
  int4  : 46324.10
  int8  : 46139.03
  int16 : 45105.82

Transfer bandwidth (GBPS)
  enqueueWriteBuffer              : 20.85
  enqueueReadBuffer               : 20.48
  enqueueWriteBuffer non-blocking : 20.84
  enqueueReadBuffer non-blocking  : 20.46
  enqueueMapBuffer(for read)      : 9.07
    memcpy from mapped ptr        : 28.45
  enqueueUnmap(after write)       : 26.86
    memcpy to mapped ptr          : 28.05

Kernel launch latency : 9.42 us

[RhynarAI]

Thanx for sharing. I wonder why FP16 is shown as "no support"... Ada Lovelace has FP16 Support.

[moyang]

I guess it means no "native support" for half-precision. FWIW Ada emulates FP16 using FP32, hence FP16 and FP32 have the same tflops. In contrast, recent AMD architectures (RDNA, CDNA) has FP16 performance 2x as much as FP32.

[RhynarAI]

Yeah, thats probably the reason. But it still is missleading - because of course one can do FP16 on Ada and the Tensor Cores do have native FP16 support, so its not like the card can´t process or store FP16/BF16..
Is Tensor core support possible in OPENCL ?