Advice for a Nvidia desktop GPU fully compatible with TX2

Hello, dear developers!

I’m starting a project for autonomous UAV navigation with real time obstacle avoidance and I’m planning to use TX2 as on-board PC.

I still haven’t purchased TX2 dev board, but I know that I would have to use GPU havily for Point Cloud estimation and manipulation. I already found several open source libraries that works with CUDA and I would like to be able to compile and test them in advance of buying TX2.

So my question is - could you please advice me Nvidia GPU to buy for a desktop PC that would support full CUDA-code compatibility between this GPU and TX2. I want to have all the features that TX2 has and (if possible) similar performance, so that I could simulate it while working with desktop PC. I don’t need the high-end GPU model, but just the one that would work fine. Does GTX 1050 looks like something that I need?

I also don’t have any experience with CUDA, thus I decided to buy the GPU to learn it as well.

Thank you in advance!

It depends on if you require half-precision FP16 support in the discrete GPU or not - not all the dGPU’s have FP16. You would probably also want a card from the same GPU family (Pascal), so that would point you to GP100-based cards like the Quadro GP100 or P100 server card.

If you can forgo FP16 support, then the GTX 1050 or GTX 1030 are good alternatives.

I actually have spare GT 1030, but I found this article saying that it’s not performing correctly:


It says that it was failing some mining tests (irrelevant to me), but as well wasn’t performing well with TensorFlow, which is an issue, since I’m planning to use it as well.

It also says there ‘CUDA Capability Major/Minor version number: 6.1’, does it mean that I can’t use lates CUDA versions with GT 1030?

In this topic, I found that GTX 1050 supports FP16, but low-rate performance. So the CUDA code must me fully compatible, right?
https://devtalk.nvidia.com/default/topic/1023708/gpu-accelerated-libraries/fp16-support-on-gtx-1060-and-1080/

The latest CUDA features target the Volta family, which adds WMMA / TensorCore support, so the Pascal family cards do not support that.

It should at least run, albeit not with ideal performance. Since you are looking to run TensorFlow and may need more memory, you could look for a GTX 1050 Ti 4GB GDDR5, or GTX 1060 6GB GDDR5. Jetson TX2 has 8GB memory capacity. Although if you are trying for an accurate comparison, it makes most sense just to go for the Jetson TX2.

Thank you!