Completed my 64GB VRAM rig - dual MI50 build + custom shroud

Hello everyone! A few months ago I started a project to build my own local AI server. After some testing and buying the second GPU, I was able to finalize the setup. \*\*Specs:\*\* \* \*\*Motherboard:\*\* Gigabyte X399 DESIGNARE \* \*\*CPU:\*\* Threadripper 2990WX (32 Cores / 64 Threads) \* \*\*RAM:\*\* 64GB DDR4 \* \*\*GPUs:\*\* 2x AMD Instinct MI50 32GB \*\*Costs:\*\* Everything was bought used, except for the case. \* Motherboard + CPU + RAM + PSU: \\\~690€ \* GPUs: about 330€ each \* Case: \\\~150€ \* \*\*Total:\*\* \\\~1500€ \*\*Software:\*\* \* Ubuntu 24.04 LTS \* ROCm 6.3 \* llama.cpp It runs GLM 4.7 flash Q8\\\_0 at \\\~50 tokens/s (but it drops down fast) I need to tinker a bit more with the setup to test things out. \*\*Custom GPU shroud\*\* One of the major constraint was that the machine needs to not be super loud, as it sits under my desk. For that I designed and 3D printed a custom shroud to ensure proper cooling while keeping it (somewhat) silent. The shroud is open source and licensed under MIT ! Its a modular build, easily printable on small 3D printer, 3 parts assembled with M2 and M3 screws. For cooling it uses a single 92mm fan (Arctic P9 Max), works pretty nicely ! \*\*Repo:\*\* \\\[\[https://github.com/roackim/mi50-92mm-shroud\\\](https://github.com/roackim/mi50-92mm-shroud)\](https://github.com/roackim/mi50-92mm-shroud\](https://github.com/roackim/mi50-92mm-shroud)) \\\*\\\*STLs:\\\*\\\* \\\[\[https://github.com/roackim/mi50-92mm-shroud/releases/tag/1.0.0\\\](https://github.com/roackim)\](https://github.com/roackim/mi50-92mm-shroud/releases/tag/1.0.0\](https://github.com/roackim)) \*\*Details\*\* The cards stay around 18W idle and use about 155W on load. Note: Since my motherboard doesn't expose FAN header controls, I set the speed to \\\~2700rpm. It’s not that loud, but it’s a fixed speed, bummer. Overall happy with the build. It was super fun designing and building the custom shroud for the GPU ! If you guys have any tips to share regarding llama.cpp, dual gpus or amd MI50s I would be very grateful ! Thanks🐔