SIGNAL
AI inference is splitting into two specialized jobs — and the seam between them is becoming a product. One chip class handles the prompt, reading a long input in a single high-throughput pass. A different chip class handles the reply, generating tokens one at a time with minimal latency. Forcing both onto the same silicon wastes the trade-off. Nvidia has been pushing the same division internally; hyperscalers have been quietly splitting inference in their own stacks for months. What changes with a new AMD-Cerebras pairing announced this week is that the seam is now a co-engineered reference design, expected to ship through Cerebras Cloud in the second half of 2026.
The announcement pairs AMD's Helios rackscale platform — a rack of AMD Instinct GPUs — with the Cerebras Wafer-Scale Engine, a single chip roughly the size of a silicon wafer purpose-built for fast token generation. AMD and Cerebras claim the combined workflow delivers up to five times more tokens per second per watt than running the whole job on either side alone, a number worth treating as theirs until someone else measures it. The joint solution will deploy AMD Helios in Cerebras data centers, making the disaggregated workflow available first through Cerebras Cloud.
Reported by Sky for Type0, from AMD and Cerebras Announce Industry-Leading Ultra-Low-Latency and High Throughput AI Inference Solution. Read the original: stockhouse.com