Ohio State researchers call for 'scale-across' networking as AI clusters outgrow data centers
At the SC26 conference, researchers from The Ohio State University said that AI clusters with over 100,000 accelerators are exceeding the power and physical capacity of single data centers, requiring workloads to span multiple facilities and regions. They argued that beyond the established scale-up and scale-out interconnect approaches, a new 'scale-across' networking domain is needed to connect geographically separate compute resources efficiently.
GoKawiil's interpretation of the reporting above, not reported fact.
As accelerator compute performance grows faster than network communication capacity, the researchers suggest that interconnect efficiency—not raw chip speed—will increasingly determine how much real-world performance massive AI systems can extract from their hardware. This could push major chipmakers and networking vendors to develop new standards for linking data centers across distances, beyond current technologies like NVLink, UALink or Ultra Ethernet.
- AI clusters now exceed 100,000 accelerators, straining single data center power and space limits.
- Researchers propose a new 'scale-across' interconnect category alongside existing scale-up and scale-out networks.
- Network efficiency, not just accelerator speed, is becoming the key bottleneck for large-scale AI performance.
Source: tomshardware.com — Anton Shilov, 2026-11-20
Published there as: “Power and cooling limits push 100,000+ accelerator AI clusters past single data centers”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.