Single Root Complex Purley 4U GPU Server for Deep Learning Applications

Increased Total Expansion Capability from Its Previous Generation

A new design based on PNYSER48 series, PNYSRA4X series can commodate up to 10 double-width GPU boards and has an 11th slot to the rear of the chassis, allowing a single-width PCIe x16 card to be installed. This new design change is ideal for NVMe deployment or high speed networking like 100 Gigabit EDR Infiniband, 100 Gigabit Ethernet, or 100 Gigabit Intel Omni-Path fabric. Furthermore, an internal storage mezzanine slot (for SAS storage deployment without taking up PCIe slots) is also available for integrating device controller solutions.

Single-Root Complex Design – Optimized for Deep Learning Applications

As a single-root complex GPU server, PNYSRA4X series deploys a set of two PLX® PEX8796 PCIe switch chips to provide direct connectivity between one single CPU and 10 GPU boards (PCIe x16 slots) and an high-speed network card (PCIe x16 slot). For Deep Learning applications, this design allows data exchanges and instruction signals to be assembled/dispatched more efficiently (between the CPU, the 10 GPU boards and the high-speed network card), both on the hardware level (single or multiple nodes) and in the Deep Learning framework. This also translates into better conservation of precious CPU compute resources (on the 2nd CPU) for faster workload processing (either on different threads apart from the 1st CPU or in collaboration with the 1st CPU).