Cerebras hands prefill to AMD GPUs to get 5x more tokens per megawatt Cerebras Systems has put AMD Instinct GPUs on prompt prefill and its CS-3 wafer-scale systems on decode in a production system it says delivers 5x more tokens per megawatt, first on Cerebras Inference Cloud and then in partner data centers. The split-workload design, described by Cerebras CEO Andrew Feldman, pairs AMD GPUs for prompt processing with Cerebras CS-3 hardware for token generation. Cerebras hands prefill to AMD GPUs to get 5x more tokens per megawatt The production system puts AMD Instinct GPUs on prompt processing and Cerebras CS-3 systems on decode, first on Cerebras Inference Cloud and then in partner data centers. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Andrew Feldman built Cerebras Systems around a chip so large it uses most of a silicon wafer. The company's newest production system divides the work between that chip and AMD GPUs.…