AMD announced a partnership with Cerebras to integrate its Helios rack‑scale solution with Cerebras’ Wafer‑Scale Engine, claiming up to five times higher token throughput per watt than rival designs.

NVIDIA acquired Groq earlier this year after its Language Processing Unit showed strong inference efficiency. The LPU architecture clusters specialized chips with large SRAM and a compiler that schedules operations at the nanosecond level, eliminating traditional branch prediction and scheduling hardware. Cerebras’ Wafer‑Scale Engine places an entire supercomputer’s worth of compute and memory on a single silicon wafer, allowing models to reside fully in on‑chip SRAM and supporting both training and inference workloads.

AMD’s Helios platform provides high‑throughput processing for large context windows and prompt handling. Cerebras’ Wafer‑Scale Engine accelerates token generation with low latency and massive on‑chip memory bandwidth. By linking the two engines in a single workflow, the companies say the combined system can deliver up to five times more tokens per second per watt. AMD noted that Helios offers ultra‑high throughput while Cerebras adds ultra‑low latency for memory‑intensive token generation, creating a differentiated platform for ultra‑low‑latency inference without sacrificing scale.

NVIDIA’s Groq 3 LPX rack‑scale offering includes 256 interconnected LPU accelerators, liquid cooling, and 315 PFLOPS of inference performance. While NVIDIA markets the Groq 3 LPX as a rigid, high‑density solution, AMD’s integrated approach is positioned as more flexible for diverse workloads.

The partnership highlights a competitive push in the AI accelerator market, where differing architectural choices aim to balance throughput, latency, and energy efficiency for next‑generation AI applications.