FusionQuill AI Server: an five AI GPU server Build with RTX 6000s and H200
Story:
As an intern at FusionQuill AI, I had the opportunity to assemble a multi-GPU server used for running local AI Models. Before this, I had built a PC, with dual Nvidia RTX 4090 GPUs. The server build was more complex, with 4 NVIDIA RTX 6000 Ada Generation GPUs and an AMD EPYC CPU. To fit all the parts, I choose the Lian Li Dynamic O11 XL, a case designed for up to e-ATX motherboards. The build overall went well, there were some issues with CPU cooling but were solved.
Then six months later, I had to add the Nvidia H200, a fifth GPU in a case designed for four. Cooling and airflow issues due to the limited number of fans arose and the number of ports was an issue. To solve these concerns, we used PCIe Riser cables to connect the GPUs and replaced the front-panel with a 3D printing mesh and added fans there as well.
Specs:
CPU: AMD Epyc
GPU: RTX ADA 6000 (4x), H200
Memory: 512 GB DDR5
RAM: 4TB M.2 SSD + 24TB HDD
Case: Lian Li 011 XL
Reflection:
The project taught me problem-solving and how the real-world solutions are different from the ones we imagine. To overcome challenges, we need to adapt and re-design until we find the best solution.