The completed system will span 64 racks, each with a CS-2 accelerator, and is expected to deliver four exaFLOPS of sparse FP16 performance. The system will be housed at Colovore's Santa Clara facility due to its high power and thermal management requirements. Two more US-based sites will be built in Asheville, North Carolina, and Austin, Texas, with completion slated for the first half of 2024. The final stage will see Cerebras construct an additional six sites, aiming to complete installation of all nine sites by the end of 2024. The systems will be operated under US law and will not be made available to advisory states.
Key takeaways:
- AI company Cerebras has unveiled its Condor Galaxy supercomputer, a distributed cluster that will span nine sites capable of 36 exaFLOPS of combined FP16 performance when complete.
- The first phase of the system has been built for the United Arab Emirates' G42, a multinational conglomerate with an interest in AI research and development.
- The complete Condor Galaxy system will feature 576 CS-2 accelerators and is expected to be installed by the end of 2024.
- While Cerebras will operate and manage the systems, they're owned by G42, which plans to use the systems for its internal workloads. Any leftover resources not consumed by G42 will be made available to both G42 and Cerebras's customers.