GigaIO, a provider of workload-defined infrastructure for AI and technical computing workflows, has successfully configured 32 AMD Instinct MI210 accelerators to a single server using its FabreX ultra-low latency PCIe memory fabric. This solution, known as SuperNODE, presents a simplified system that enables the seamless scaling of multiple accelerator technologies, including GPUs and FPGAs, without the typical latency, cost, and power overhead associated with multi-CPU systems.
As the demand for enhanced GPU performance continues to rise, technologies that reduce node-to-accelerator data communications become increasingly vital in delivering the necessary compute power while optimizing infrastructure total cost of ownership (TCO).
GigaIO’s SuperNODE system underwent rigorous testing, utilizing 32 AMD Instinct MI210 accelerators on a Supermicro 1U server powered by dual 3rd Gen AMD EPYC processors. The system delivered impressive scalability in two key tests. Hashcat exhibited a remarkable 100% scale factor, while Resnet50 achieved an impressive 99% scalability.
