While last week we were able to write about the NVIDIA Jetson TX1 development board, at that time we weren't able to share any benchmarks or hands-on experience with this ARM board powered by NVIDIA's Tegra X1 SoC. The embargo on that has now expired and as such this morning there are a lot of benchmarks to share with you. There are many benchmarks looking at different areas of the Jetson TX1 including power consumption and thermal. For kicks I've also done some comparisons against the Tegra 2 and Tegra 3 as well as other ARM hardware like the now defunct Calxeda ARM server and Raspberry Pi 2.
As a recap, the NVIDIA Jetson TX1 is comprised of four ARM 64-bit Cortex-A57 CPU cores (max frequency of 1.91GHz) and four Cortex-A53 low-power cores, 4GB of LPDDR4 memory with 25.6GB/s of memory bandwidth, and has a 256-core Maxwell graphics processor capable of 1 TFLOP/s. The Tegra X1 is a very nice upgrade over the Jetson TK1 with the Tegra 1 predecessor that was 32-bit ARM and using a Kepler graphics processor. There is 16GB of eMMC storage onboard as well as a SD slot for expanded storage. The Jetson TX1 has onboard 802.11ac WiFi as well as Bluetooth while still having a Gigabit Ethernet port, one USB port, and a PCI Express x4 slot. NVIDIA refers to the Jetson TX1 as a super-computer on a module, thanks to the 400-pin board-to-board connector that will allow the module to be adapted for other purposes rather than needing to fabricate your own board around a Tegra X1 SoC.
With NVIDIA having talked up the Jetson TX1 a lot for deep learning purposes and being able to fit the JTX1 module on drones and other portable, low-power devices, of course there's some deep learning benchmarks. It's why last week I added Caffe AlexNet to the Phoronix Test Suite along with some other interesting CUDA benchmarks.
Besides looking at the performance around NVIDIA's advertised use-cases for the Jetson TX1, I ran plenty of other benchmarks too for both the four A57 cores and Maxwell GPU... The Jetson TX1 is the first board to market I find interesting since they're using the more powerful A57 cores rather than A53 cores found on the cheaper AArch64 development boards that are out there now. Plus with the powerful Maxwell graphics could make this a nice 64-bit ARM desktop for Linux developers. For those that missed the earlier article, the Jetson TX1 development board kit retails for $599 USD although an educational-discounted version will sell for $299.
Before getting to the performance results on the Jetson TX1, there's a few items to talk about. First, the form factor of the Jetson TX1 board is great: it's mini-ITX compatible! ARM development boards these days come in all shapes and sizes while it's fantastic that this one can be mounted in a mini-ITX chassis. The board layout is also better than the Jetson TK1 where the power connection and SD card slot were on the opposite side as the rest of the connectors: with the Jetson TX1, all of the necessary connectors are on the back side.
At first I was very nervous whether the mini-ITX board could fit in a 1U chassis due to its rather large cooler and looked like it would be a tad too big. Fortunately, the Jetson TX1 just perfectly fit inside a thin mini-ITX 1U rackmount case. This was the case I bought and fit fine with the Jetson TX1 for less than $50 while being rackmount friendly. Great for making it stackable and racked up alongside the dozens of other benchmarking systems at Phoronix. The board layout is certainly a nice improvement over the Jetson TK1.
The Tegra X1 SoC supports accelerated VP9 and H.265 video decoding and is capable of decoding at 4K 60 FPS. There is also accelerated encode for H.264, H.265, and VP8 (no VP9 mentioned for encode). The HDMI connection supports HDMI 2.0 and HDCP 2.2 and is capable of outputting at 4K 60 FPS. I've been testing the Jetson TX1 a lot from a Dell 4K display and this sub-10W board has no problems driving at 3840 x 2160!