The NVIDIA Jetson TX2 Performance Has Evolved Nicely Since Launch

Word this week of the NVIDIA Jetson Xavier Development Kit being up for pre-order reminded me of some benchmarks I had been meaning to do of seeing how the NVIDIA Jetson TX2 developer kit's performance has evolved since its launch a year and a half ago. There's actually a quite measurable improvement in performance with the latest software/drivers compared to it was at launch.

The Jetson Xavier Development Kit is open for pre-order this week. This new kit costs a whopping $2,499 USD or $1,299 for a limited "first kit" special price for existing NVIDIA Developer Program members. As a reminder, the Jetson Xavier Development Kit features eight ARMv8.2 CPU cores, a 512-core Volta GPU with Tensor cores, 16GB of LPDDR4 memory, 32GB eMMC, dual NVDLA deep learning accelerators, a 7-way VLIW vision processor, and the Jetson Xavier itself comes in the same "module" form factor that can be attached to the developer kit or other use-cases.

The Jetson TX2 that launched back in March of 2017 features two Denver CPU cores plus four ARM Cortex-A57 cores, a NVIDIA Pascal-based GPU with 256 CUDA cores, 8GB of LPDDR4 memory, 32GB eMMC, and lacks the tensor cores and NVDLA accelerators of the soon-to-ship Xavier Development Kit.

Anyhow, the benchmarks today are seeing how the performance of the Jetson TX2 has changed since launch. In particular, still using the Ubuntu 16.04 "Linux 4 Tegra" root file-system but with having upgraded to the L4T R28.2.1 release that was released in June.

The updated Linux 4 Tegra stack uses Ubuntu 16.04 with the same GCC 5.4.0 compiler, X.Org Server 1.18., and uses the Linux 4.4 LTS kernel series still. Though the kernel revision is newer in going from Linux 4.4.15 to 4.4.38 plus having all of the various Tegra kernel bits back-ported. The L4T stack compared to launch also means moving from CUDA 8.0 to CUDA 9.0. NVIDIA continues using the Tegra CPUFreq Schedutil CPU frequency scaling driver/governor.

Most of the tests in this article are looking at how the general CPU performance has improved over the past eighteen months worth of software updates but the CUDA performance has also improved too as to be shown.