AI system - Pytorch on Linux

Does anybody have a reliable source for Nvidia boards? One of my sons is a PhD AI student and has to build a machine to perform his research. I think he/ we (I’m helping financially here) are looking for a GTX 1080 ti to be the centerpiece of the system.

Of course we are looking for suggestions on this and the other components that need to be acquired. He’s going to be running Pytorch on the Linux OS.

Comments??

2 Likes

So from what I read PyTorch is primarily for Intel Processors and Cuda cores. So I would stay away from AMD CPU’s. This article doesn’t exactly explain min requirements but does stress the need for High End CPU, GPU, and large amounts of RAM.

I am think you are going to want something like this. It is a scaled down version of what our AI server is here at work. Along with your 1 or 2 1080ti’s. Not sure what the budget is like but from what I read in that article I would scale down the CPU budget wise first. I have no actual experience building an AI machine just trying to look at what we have at work for AI and scale it back. Titan V is 3k but meant for AI.

https://www.ibm.com/developerworks/library/cc-get-started-pytorch/index.html

1 Like

It also suggests using these.

1 Like

Maybe it is not as intensive as IBM made it out to be just found this article that may help in the build process.

If you cant tell I love spec’ing out computer hardware hehe. Totally should be working right now.

1 Like

Last links and build then I need to get back to work. Found this thread detailing each aspect so make of it what you will but basically you want a high speed single core due to the bottle neck of DMA direct memory access between the GPU and CPU. You also want at least as much RAM on the MB as your GPU’s have. So a 1080ti has 12 you want 16 on the mother board. So in this instance you want your MB to scale with GPU’s if you add any the MB selected here can go up to 128GB. HDD’s pose a bottle neck also so NVMe ssd’s are in order and this MB can hold 3 NVMe HDD’s total.

This is beginning to make more sense on how our AI servers are built at work. We have large farms of regular SSD’s but all the drives in the actual computer servers are NVMe drives. So basically the entire work load is transferred from 3.5" disk drives or SSD’s to the NVMe drives when the compute is done it transfers it back and starts the next batch. File used often are on the SSDs and the files not used that often on slow disk drives. May want to make a similar setup with this machine on a scale that fits your budget.

If I was to build one for myself here is what it would look like.
https://secure.newegg.com/Wishlist/SharedWishlistDetail?ID=4fgoQs%2BAHUOY%2Fs9xXebp5w%3D%3D

I changed it from a 1080ti as you can get a Titan X for about the same price currently and get better AI performance as it was intended more for this use case than gaming. Titan V would be better but twice the price.

1 Like

💰 YEN · DCTV ·️ Bitcoin Lambo · 10 Days of Bitcoin ·️ CEO's Brainpan 🧠