Nvidia paid $20 billion for SRAM decode - AMD just partnered for it instead (www.techradar.com)

AMD has announced a strategic partnership with Cerebras Systems to develop a disaggregated inference solution that enhances efficiency in AI workloads. By integrating AMD's Helios rackscale system with Cerebras' Wafer-Scale Engine (WSE), both companies claim to achieve up to five times the tokens per second per watt (TPS/W) compared to a standalone WSE configuration. This collaboration aims to tackle efficiency challenges associated with the WSE, particularly in prompt processing, ultimately streamlining AI inference operations. This development is significant for the AI/ML community as it represents an alternative approach to Nvidia's recent $20 billion licensing deal for SRAM decode technology, providing AMD a cost-effective means to enhance its capabilities without such financial outlay. The partnership leverages the strengths of both companies; while the WSE struggles with the 'prefill' stages of processing, AMD's hardware complements it effectively. However, the validity of the performance claims rests on a specific model, leading to the necessity for further testing across various AI workloads to fully understand its strengths and weaknesses. As the AI market evolves, partnerships like this highlight the competitive dynamics of the industry and the emphasis on efficiency and innovation in AI chip design.