Looking for clpeak results on Jetson Orin Nano Super

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
1

The title says it all. I would appreciate the results if someone has them.

4

Hi,
We have benchmark on the blog:
NVIDIA Jetson Orin Nano Developer Kit Gets a “Super” Boost | NVIDIA Technical Blog

But we don’t run clpeak. Please share the steps, and we can set up developer kit and try.

5

The easiest would be to install a binary package with sudo apt install clpeak. If not available, this should work:

git clone https://github.com/krrishnarraj/clpeak
cd clpeak
mkdir build
cd build
cmake ..
cmake --build .
./clpeak
8

Hi,
For reference, we set up Orin Nano developer kit in Jetpack 6.2.2 Super config and run the application:

$ git submodule update --init --recursive --remote
$ cmake -S . -B build -DCLPEAK_ENABLE_VULKAN=OFF -DCLPEAK_ENABLE_METAL=OFF
$ cmake --build build -j
$ ./build/clpeak
clGetPlatformIDs (-1001)
no platforms found

=== CUDA backend ===

CUDA Device: Orin
  Arch          : sm_87
  Driver        : 12.6
  NVRTC         : 12.6
  SMs           : 8
  VRAM          : 7607 MB

  Single-precision compute (GFLOPS)
    float : 1556.99

  Half-precision compute (GFLOPS)
    half : 1792.71
    half2 : 1794.80

  Double-precision compute (GFLOPS)
    double : 29.22

  Mixed-precision compute fp16xfp16+fp32 (GFLOPS)
    mp : 1342.48

  BF16 compute bf16xbf16+fp32 (GFLOPS)
    bf16 : 1758.27

  WMMA tensor-core compute (TFLOPS)

  WMMA fp16xfp16+fp32 16x16x16 (TFLOPS)
    wmma_fp16 : 116.70

  WMMA bf16xbf16+fp32 16x16x16 (TFLOPS)
    wmma_bf16 : 39.28

  WMMA tf32xtf32+fp32 16x16x8 (TFLOPS)
    wmma_tf32 : 22.02

  WMMA fp64xfp64+fp64 8x8x4 (TFLOPS)
    wmma_fp64 : 0.10

  FP8(E4M3) mma.sync m16n8k32+fp32 (TFLOPS)
    FP8 mma.sync requires sm_89 or newer (Ada/Hopper+)! Skipped

  FP8(E5M2) mma.sync m16n8k32+fp32 (TFLOPS)
    FP8 mma.sync requires sm_89 or newer (Ada/Hopper+)! Skipped

  cuBLASLt GEMM peak (TFLOPS)
    fp32 : 1.43
    fp64 : 0.03
    tf32 : 4.47
    fp16 : 13.43
    bf16 : 13.05

  Integer compute (32-bit IMAD) (GOPS)
    int : 925.59

  INT8 dot-product compute (__dp4a) (GOPS)
    int8_dp : 2795.85
    int8_dp2 : 2850.52
    int8_dp4 : 3071.38
    int8_dp8 : 3046.26

  Packed INT4 compute (emulated) (GOPS)
    int4_packed : 591.12

  WMMA int8xint8+int32 16x16x16 (TOPS)
    wmma_int8 : 197.53

  INT8 mma.sync m16n8k32+int32 (TOPS)
    int8_k32 : 25.62

  INT4 mma.sync m8n8k32+int32 (TOPS)
    int4 : 27.12

  BMMA b1 mma.sync m8n8k128+int32 xor.popc (TOPS)
    bmma_b1 : 108.69

  cuBLASLt GEMM peak (TOPS)
    int8 : 13.09

  Atomic throughput (GOPS)
    global : 22.21
    local  : 222.88

  Global memory bandwidth (GBPS)
    float   : 62.94
    float2  : 85.33
    float4  : 92.51

  Local memory bandwidth (GBPS)
    float  : 447.01
    float2 : 689.57
    float4 : 770.06

  Image memory bandwidth (GBPS)
    float4 : 231.31

  Transfer bandwidth (GBPS)
    Host->Dev (pinned) : 25.73
    Dev->Host (pinned) : 25.79

  Kernel launch latency (us)
    dispatch  : not measurable via CUDA driver API
    roundtrip : 16.17
9

Thank you for running the benchmark. Was it on Orin Nano Super or the older Orin Nano?

10

Hi,
It is Orin Nano developer kit with Super config. The flash command is
Quick Start — NVIDIA Jetson Linux Developer Guide

$ sudo ./tools/kernel_flash/l4t_initrd_flash.sh --external-device nvme0n1p1 \
  -c tools/kernel_flash/flash_l4t_t234_nvme.xml -p "-c bootloader/generic/cfg/flash_t234_qspi.xml" \
  --showlogs --network usb0 jetson-orin-nano-devkit-super internal