NVIDIA Jetson TX2 Linux Benchmarks

[phoronix]

Phoronix: NVIDIA Jetson TX2 Linux Benchmarks

Last week we got to tell you all about the new NVIDIA Jetson TX2 with its custom-designed 64-bit Denver 2 CPUs, four Cortex-A57 cores, and Pascal graphics with 256 CUDA cores. Today the Jetson TX2 is shipping and the embargo has expired for sharing performance metrics on the JTX2.

http://www.phoronix.com/vr.php?view=24362

[coder111]

I wonder how would they compare to AMD Embedded R-series CPU based SOCs, say RX-421BD. Although that one seems to consume on 2x more power.

How is OpenCL support on AMD Open-Source drivers? Even with proprietary drivers, how would AMD OpenCL performance compare to NVidia CUDA or NVidia OpenCL?

In terms of maturity and availability of computational libraries, is CUDA or OpenCL ecosystem more active?

[RussianNeuroMancer]

I wonder how would they compare to AMD Embedded R-series CPU based SOCs

There is long discussion about comparison with AMD R-Series in this thread.

[dungeon]

This should be compared to Intel's Goldmont i guess... to something like this $99 x86:

https://www.newegg.com/Product/Produ...82E16813157731

As i see, there are some results on openbenchmarking... at least C-Ray goes better

https://openbenchmarking.org/s/ASRock%20J4205-ITX

[coder111]

Hi,

I just read that thread. Cool discussion. Personally I hate vendor traps and I would go with Open solution given a choice, especially for a longer term project.

Have you direct experience working with GPU computations on AMD RX-series? What is the situation today with OpenCL on AMD on Linux? Are you using open-source driver chain, or proprietary? How is the performance on Linux?

I haven't worked in embedded, and I haven't worked directly with GPU computing. But one of coworkers I spoke with a couple of years ago said that performance he was getting with CUDA on NVidia GPUs was miles ahead what he was getting with OpenCL & AMD. That was on Windows, and AFAIR mostly doing Monte-Carlo simulations & some kind of financial risk calculations. Never did go into much detail. Again, this could have been this specific use-case, or maybe situation on the market at that time.

I wonder how are things today.

--Coder

There is long discussion about comparison with AMD R-Series in this thread.

[Michael]

I wonder how would they compare to AMD Embedded R-series CPU based SOCs, say RX-421BD. Although that one seems to consume on 2x more power.

How is OpenCL support on AMD Open-Source drivers? Even with proprietary drivers, how would AMD OpenCL performance compare to NVidia CUDA or NVidia OpenCL?

In terms of maturity and availability of computational libraries, is CUDA or OpenCL ecosystem more active?

The OpenCL is basically non-existant at the moment for AMD open-source... The Clover Gallium3D code is in rough shape and not officially supported. They are working on their OpenCL stack with ROCm, but that isn't yet fully open or mainlined. But even when that stack is there, it seems like they are only supporting Polaris and Fiji with it, unless they end up porting it back to older hardware on AMDGPU.

OpenCL on AMDGPU-PRO performs quite decent compared to NVIDIA.

[RussianNeuroMancer]

Are you using open-source driver chain, or proprietary?

For now you'll need AMDGPU-PRO for OpenCL. (Not sure if ROCm driver release is required - most likely not.)

How is the performance on Linux?

I doesn't seen articles about this and didn't compare it myself, but people talking it used to be better on Linux, at least in the past. I doesn't know current Win vs. Lin situation.

But one of coworkers I spoke with a couple of years ago said that performance he was getting with CUDA on NVidia GPUs was miles ahead what he was getting with OpenCL & AMD.

Depends on algorithm optimisation I guess. You can find many NV vs. AMD OpenCL benchmarks on Phoronix.

[ruthan]

Ad some comparison with x86 and Core i3 and cheap Rapsberry 3, some Games would be nice too. I know that Michael ignoring Android at all, this board is suited for it.

[tildearrow]

Michael Typos:

MAX-P is designed for maximum performance in delivering up to twice the performance of hte JTX1 at less than 15 Watts.

The SoC temperature while running AlexNet was 33~35&C.

[Michael]

Michael Typos:

Fixed, thanks,