FuriosaAI and Supermicro Explore 16-Card AI Server Tray Design
South Korean AI chip startup FuriosaAI is collaborating with Supermicro to design a 16-card AI server tray using a dual-layer motherboard structure. This design aims to expand the XPU interconnect domain and reduce data movement across server trays to improve efficiency. Standard non-rack AI systems typically support a maximum of 8 XPUs per node, so doubling this capacity to 16 cards can significantly improve operational efficiency for AI workloads. It demonstrates a novel hardware architecture approach to scaling AI compute density and interconnect bandwidth. The server features a "sandwich" structure with two motherboards sharing a single cooling system, each equipped with 8 PCIe slots. Additionally, it will utilize Astera Labs' high-lane-count switch chips to increase the Network Interface Card (NIC) to switch ratio from 4:1 to 16:1.
## BACKGROUND
An XPU is a generic term for various processing units like GPUs, NPUs, or TPUs used to accelerate AI workloads. In AI clusters, the interconnect domain refers to the network of high-speed connections linking these processors, where larger domains help reduce latency and communication bottlenecks during distributed training.