Mathematica CUDALink on Nano?

Hi, I’d like to try running Mathematica on the Jetson Nano, using CUDALink to accelerate Mathematica computations. I believe there is a Mathematica kernel installation available for ARM-based Linux, but is the CUDALink extension compatible with the Jetson Nano GPU?


Wolfram would have to provide this (CUDALink is provided by Wolfram). Either in arm64/aarch64/ARMv8-a format, or such that you could compile it on the Jetson. However, even if they provide one in this architecture there may be another catch…read on…

I use an 11.x version of Mathematica, and long ago used C to link, but not CUDA. My thought is that you’re probably not going to get this to work because most of the GPU discovery libraries I’ve seen (but not necessarily for Wolfram…I am generalizing) require a PCIe video card. My Mathematica 11.x version automatically detects my discrete video card, but I doubt it could detect video which is directly integrated to the memory controller.

You might ask Wolfram if they either support (or would consider supporting) the 64-bit ARM (arm64/aarch64/ARMv8-a) architecture, and if so, whether configuration of GPU could be based on non-PCIe functions (the GPU and CUDA can be directly queried, but nvidia-smi does not exist on Jetsons).

FYI, here is a starting reference, but only binary desktop PC formats are shown for Linux:

Ah, thanks. The CUDA Resources page indeed does not have an ARM build for CUDALink. I’ve already reached out to Wolfram on this and we’ll see what they say, but it looks like they would not support an ARM-based Linux board with integrated GPU. Thanks, though!