How I built a €25K Machine Learning Rig | Hacker News Reader

How I built a €25K Machine Learning Rig

emilwallner.com

15 pointsemilwallner9 comments

Hey, I made a machine learning rig with four NVIDIA RTX A6000 and an AMD EPYC 2 with 32 cores, including 192 GB in GPU memory and 256GB in RAM.

I made a 4000-word guide for people looking to build Nvidia Ampere prosumer workstations and servers, including:

- Different budget tiers

- Where to place them, home, office, data center, etc.

- Constraints with consumer GPUs

- Reasons to buy prosumer and enterprise GPUs

- Building a workstation and a server

- Key components in a rig and what to pick

- Lists of retailers and build lists

Let me know if you have any questions!

I'd be really worried about thermals. The cards are very close to each other for an open air case, I'd expect it would need a case and enough fans to create some air pressure.

There is some weird parts about mining rigs that I don't really agree with.

"pcie risers" - as the author pointed out mining rigs usually go for 1x PCIE risers through USB cause they don't care about bandwidth - but that doesn't mean you need to go for that as well, you could have gone for 16x PCIE risers just fine which is just a long ribbon cable usually shielded.

"mining rigs use crappy PSUs from alibaba" - you can use whatever PSU you want.

I converted it into a server, I've been running it on full load for a week and the temperatures are 80C +- 2 degrees. I don't know if there is a way to use x16 cables without transferring electricity? Thus, mixing circuits. Consumer PSUs are maximum 2000W, so you have to turn to Alibaba for +2100 W PSUs.

Even though I’ll never use it, I still enjoyed the read!!!

Thanks!!

Thermaltake box. The best.

Yeah, they are mesmerising!

put your build on pcpartpicker,

https://fr.pcpartpicker.com/b/sDmqqs this is my 2018 build :)

Tight!! I was planning on it, but half of the components are enterprise parts and are not listed on pcpartpicker.

Reply on news.ycombinator.com