Home / Technology / Meta's AI Data Center: 1000 GPUs as One
Meta's AI Data Center: 1000 GPUs as One
14 Sep
Summary
- Meta aims for nearly 1,000 AI GPUs in a single data center domain.
- Panmnesia's CXL design links CPUs, accelerators, and memory across racks.
- The architecture boosts accelerator coordination to 16 per CPU.
Meta is collaborating with Panmnesia on an advanced AI data center architecture designed to unify thousands of AI GPUs into a single, coherent operating environment. This groundbreaking proposal leverages Compute Express Link (CXL) technology to connect processors, accelerators, and memory modules across multiple racks without relying on traditional network interfaces.
The new design aims to integrate up to 960 AI accelerators into one coherence domain, effectively enabling close to 1,000 GPUs to function as a unified system. This architecture addresses the growing challenge of coordinating massive numbers of accelerators in large-scale AI training. By minimizing unpredictable communication delays between racks, which are common with Ethernet or InfiniBand, CXL provides a more consistent processing environment.
Panmnesia's proposed system significantly enhances accelerator connectivity. Under this architecture, a single CPU could manage up to 16 accelerators, an eightfold increase compared to existing configurations like NVIDIA's GB200 NVL72. Approximately 60 such groups could then form a coherence domain, comprising around 960 accelerators. This cross-rack access is expected to improve performance dramatically, with timing reductions from microseconds to hundreds of nanoseconds.
The architecture also introduces flexibility by allowing individual failed devices to be replaced without disrupting the entire server. While electrical CXL signalling has distance limitations, Panmnesia is exploring optical CXL links for extended reach, having validated this approach in a hardware proof-of-concept.
