AMD and Cerebras team up to challenge Nvidia's Groq LPUs
New collaboration aims for faster and more efficient AI inference
1A New Era in AI Inference
AMD and Cerebras Systems have partnered to develop a new platform for AI inference workloads. This collaboration combines AMD's Instinct GPUs with Cerebras' SRAM-based accelerators to deliver ultra-low-latency inference. The solution, positioned as a competitor to Nvidia's Groq LPUs, is specifically optimized for interactive agentic workloads.
2Cerebras' Technological Edge
Cerebras' wafer scale engines (WSE) use on-chip SRAM instead of HBM4, unlike traditional GPUs. This design provides significant memory bandwidth advantages, enabling output speeds exceeding 2,000 tokens per second. The technology has made Cerebras one of the world's fastest inference providers.
3Balancing Performance and Efficiency
The joint solution will process compute-heavy prompts on AMD GPUs while offloading memory-intensive token generation to Cerebras accelerators. This disaggregated architecture promises higher interactivity without compromising throughput. The companies estimate this combination could improve tokens-per-second per watt by up to 5x.
4A Competitive Alternative to Nvidia
Cerebras accelerators serve a similar purpose to Nvidia's Groq 3 LPUs but require far less hardware. For trillion-parameter models like Kimi K2.5, where Nvidia needs thousands of Groq LPUs, the AMD-Cerebras system can deliver comparable performance with just a few dozen accelerators.
5Open Ecosystem Strategy
AMD CEO Lisa Su emphasized that the Cerebras collaboration is part of the company's open ecosystem strategy. Su stated they will continue working with various accelerator technologies for different workloads. This approach positions AMD as a flexible and inclusive player in the AI hardware market.
6Intensifying Competition in AI Hardware
The AMD-Cerebras partnership highlights the fierce competition in AI hardware. Such collaborations challenging Nvidia's dominance are driving innovation across the industry. This competition is expected to bring better performance and cost advantages to consumers, particularly in inference optimization.