In the STH studio, we’ve built a robust AI cluster featuring high-memory systems to efficiently run larger models. Our setup includes 4-5 NVIDIA GB10 machines, four AMD Ryzen AI Max+ 395 systems, and an Apple Mac Studio M3 Ultra. Utilizing Solidigm SSDs with a QNAP NAS, this configuration optimizes storage and networking capabilities. The project emphasizes network-based hosting and fine-tuning of models, reducing local storage costs significantly as machine count increases.
The Solidigm D5-P5336 SSDs, despite being QLC NAND, enhance read-heavy workloads, significantly decreasing model load times compared to traditional hard drives. Integrating multiple high-speed networks, including 25GbE Thunderbolt adapters, ensures efficient data handling across platforms. Our aim is to create a comprehensive AI storage solution, supporting dynamic workloads while balancing performance and cost. This initiative demonstrates the effective use of modern hardware in AI applications, facilitating faster model processing and efficient data management strategies.
Source link
