Xavier benchmarks

Are there any performance benchmarks for the Xavier, even just comparing it to the TX2? Currently I’m seeing little difference between the Xavier and TX2 with nvidia’s FFT libraries. Just trying to find some controlled test that shows the performance.

We’ll be covering benchmarks in next week’s webinar on Jetson AGX Xavier.

Regarding cuFFT, have you tried FP16 mode of the library and using different nvpmodel power settings (see this thread).

Here’s the settings for the Xavier:

  sudo nvpmodel -q
    NV Power Mode: MAXN

and here’s the settings for the TX2:

  sudo nvpmodel -q
    NV Power Mode: MAXN

so the power models are the same.

We haven’t tried the FP16 mode yet. Thanks for the webinar info, hopefully I can gain additional insight into the performance issues.

It’s possible that cuFFT isn’t fully optimized yet for Xavier’s integrated GPU, historically cuFFT has lagged a bit behind cuDNN and TensorRT on Jetson in that respect. We’re continuing to make performance improvements to the components in JetPack, as this is initial software is an early-access release.

You can have a look at Fastvideo SDK for image processing on Tegra:

Fastvideo SDK benchmarks for Tegra X2:

Fastvideo SDK benchmarks for Tegra Xavier: