Rebuilding an old workstation into an AI testbed

Illustration of a microcomputer tower sweating while running generative AI

I had an old (circa 2012?) Hackintosh, which started life as a Core i5 3570K Processor (3.4 GHz 4 Core), installed on a Gigabyte GA-B75M-D3P main board (LGA 1155 socket, B75 chipset). Basically, a lightly tweaked version of this moarfish build. The Corsair Carbide 300R case (or, rather, the plastic drive rails) were showing their age, the power supply had a tendency to run extremely - worryingly - loudly, and so it was just sort of collecting dust. A lot of dust (see below). I had picked up a new Cooler Master case and EVGA power supply to rebuild it, but never got around to it. Instead, I eventually moved to a maxed out (1TB SSD, 16GB RAM, Core i7-3720QM CPU @ 2.60GHz) 2012 Mac mini server, running the latest macOS through OCLP.

Now that I’ve finally decided to pull the trigger and start getting my hands dirty with “AI,” after flopping around for a minute looking at various options (eGPU chassis hooked up to a 2020 MacBook Air running Linux? $249 NVIDIA Jetson Orin Nano Super Developer Kit? Raspberry Pi 5 16GB with a Hailo-10H and an M.2 hat?), I realized, hey, I’ve got this workstation, I can slap a GPU in it and start playing around. At first it was just going to be the existing configuration (i5 and 16GB RAM), plus whatever cheapest GPU would provide any reasonable performance (I was looking at the NVIDIA Tesla P100, but then saw the V100 was the first with Tensor cores and not a huge leap up in price ($150 approx. for a 12GB P100, $275 for a 16GB V100). GPU comparison tool. End of life / support dates for NVIDIA architectures.

But then I started realizing, remembering the NAS upgrade project, how cheap this stuff has gotten as it’s slipped further and further away from “state of the art.” Why stick with a mid-range i5 when I can max it out with the top of the line (for the time) Ivy Bridge Xeon for $35? Why stick with 16GB RAM when I can max it out at 32GB for $48? Why reuse a 13 year old OEM CPU cooler, or the old case fan, when brand new stuff is < $20? And then I realize how little I’m actually re-using in this build (just the motherboard, and the “new” case and power supply I already had on hand). Sigh. Oh well. It set constraints I had to stay within for this project, and that ain’t nothin’.

New stuff:

What I’ll have when I’m done (parts are ordered but not here yet), a Xeon E3-1275v2 (3.50GHz, 8MB L3 cache, integrated P4000 graphics) CPU (could have gone up to a Xeon E3-1290v2 (3.70GHz, 8MB L3 cache, but with no onboard graphics); 32GB RAM; a PCIe x16 3.0 slot, another x16 slot running at x4 (2.0), an x1 slot (2.0), and an old school PCI slot; in that x16 slot an NVIDIA V100 16GB GPU (CUDA cores and 640 Tensor Cores, >100 teraflops); and for storage, a SATA III Samsung  drive. That should do nicely as a starter system.

Picture of the dusty insides of a computer workstation
How it started...

Comments