AMD and Cerebras Pair Helios With Wafer-Scale Engine for 5x Faster Inference

The AI inference market just got a new power couple. AMD and Cerebras have announced a technical partnership to build a disaggregated inference solution that pairs AMD's brand-new Helios rack-scale system with the Cerebras Wafer-Scale Engine. The announcement was made live on stage at AMD's Advancing AI 2026 event, with Cerebras CEO Andrew Feldman joining AMD CEO Lisa Su in front of the crowd. The pitch: the fastest production inference tokens in the industry, at 5x greater capacity, on frontier-scale models with over one trillion parameters.

Two chips, one pipeline

To understand why this matters, you need to understand what "disaggregated inference" actually means. LLM inference has two distinct phases that have very different hardware needs:

• Prefill: Processing the input prompt and any context. This is compute-heavy and parallelizes well across many GPUs.
• Decode: Generating tokens one by one. This is memory-bandwidth-bound and benefits enormously from on-chip SRAM rather than off-chip HBM.

Inference is not one thing. It is at least two very different workloads: prefill and decode. The traditional approach of running both phases on the same GPU cluster forces an uncomfortable compromise. GPUs use two separate chips: the GPU itself for predicting tokens and HBM memory to store model weights. To run an LLM, GPUs must shuttle the entire model from memory to compute. To generate one word, the model weights must be streamed across the memory bus, and to generate the next word, they are streamed again. Because the full model cannot fit on the GPU itself, this process repeats for every token, making GPU inference persistently bandwidth-bound.

Cerebras sidesteps this entirely. The WSE-3 is a wafer-scale chip with 900,000 cores and 44 GB of on-die SRAM connected at 21 PB/s bandwidth. Unlike GPUs that use HBM memory stacks connected to the die, WSE-3 puts all memory on the wafer itself, eliminating the HBM memory wall, giving it extremely low memory latency and high bandwidth per parameter. The result: Cerebras can serve interactive applications with sub-10ms latency per token, while GPU clusters struggle below 10-12ms at the same throughput.

What Helios brings to the table

AMD's Helios is not just a new GPU. It is a single rack containing 72 Instinct MI455X GPUs, 31 terabytes of HBM4 memory, and 2.9 exaflops of FP4 inference compute. It is AMD's first rack-scale AI system and its direct answer to Nvidia's Vera Rubin NVL72. The full spec sheet is impressive: