Epic CPU Workstation with ample RAM for deep learning tasks. Recommended for AI projects. Get advice on optimization!
Epic CPU Workstation with ample RAM for deep learning tasks. Recommended for AI projects. Get advice on optimization!
I'm working on a robust workstation setup for daily Deep Learning (DL) training using PyTorch. I'm currently using an FX-8350 with a 980Ti Hybrid (overclocked) and 16GB DDR3 RAM. While it's served me well, it has two main issues: the CPU is outdated and underperforming, and I lack sufficient memory. During training sessions, I often read from my SSD thousands of times per iteration—which strains the drive and reduces its lifespan. Having the full dataset in memory speeds up training by 4-6 times, which is a major advantage. I plan to handle larger datasets in the future and would like to retain as much data as possible.
Regarding the GPU, I'm evaluating whether to upgrade now or wait for upcoming models and PCIe versions. I'm also curious about the differences between PCIe 3.0 and 4.0, and whether newer NVIDIA releases justify the investment.
My current usage involves gaming as well as intensive computations for DL, often requiring large datasets in RAM. I'm considering a new graphics card now or later, depending on upcoming announcements and performance needs.
I've worked extensively with RNNs. My setup doesn't use AMD; the software is optimized for nVidia GPUs, though AMD tools are limited in comparison. I've used non-Ryzen AMD CPUs before, but the performance didn't match Intel's in similar builds. Would you think about a dual Xeon workstation? It supports ECC DDR3 RAM, though the cost is low—currently I'm using 64GB in quad channel for under $100. PCIe 3 is better than 4, but it depends on how much you value future-proofing and your investment in DL. Are you doing this for work or just a hobby?
I'm interested in learning more about this CPU component in DL. Since most of my experience has been with image and signal processing, the main processing is on my GPU. I haven't realized there are any CPU-based tools for DL yet. I'm curious about the differences between AMD and Intel processors, especially since non-Ryzen models have historically performed poorly. I'm also surprised that Ryzen CPUs didn't meet expectations before them.
This likely highlights the main issue here. Before diving into GPU RNNs, I tested it on just CPUs. I managed a few iterations per second using an Intel C2D chip, but it slowed down to seconds per iteration on an Athlon x4. Then I moved to a GPU, starting with Radeon and comparing it to an nVidia. That was a time when machine learning was still in its early days for regular people, though it’s now very popular. The results are mostly focused on nVidia and Intel lately. It depends on what you’re investing, but I’d recommend sticking with nVidia/Intel if you’re spending a lot.