Power and cooling limits push 100,000+ accelerator AI clusters past single data centers
Reported by Tom's Hardware1 source ↓
At SC26, Ohio State researchers described scale-across networks that link data centers to address limits on power, cooling and space. They said distance, congestion and communication overhead can reduce performance across distributed clusters.
Key points
- Scale-across networks connect computing resources across buildings, campuses or regions
- Modern AI clusters can include more than 100,000 accelerators and millions of CPU endpoints
- Google introduced its Virgo network architecture to support workloads beyond one data center
- Cisco and Marvell are developing routers, switching chips and optical links for distributed clusters
Details
Ohio State researchers discussed the emerging scale-across approach at Supercomputing 2026. It extends high-performance connectivity between data centers, allowing operators to combine resources across locations rather than fit all accelerators into one facility. The article says power availability, cooling needs and physical space can limit single-site clusters.
The approach faces limits: distance adds propagation delay, and congestion, packet loss and synchronization can reduce performance. These issues matter for AI training, which involves frequent exchanges of model parameters, gradients and intermediate results.
Written from coverage as of
Images
- Tom's Hardware
- Tom's Hardware