SemiAnalysis: Kimi K3 reduces KV transmission bandwidth, AI network demand remains unchanged

SemiAnalysis states that Kimi K3 uses KDA technology to reduce KV cache transmission bandwidth by up to 10 times, but AI network demand will not shrink. The Kimi K3 has 2.8 trillion parameters and requires approximately 1.5TB HBM bandwidth for each forward calculation. It still needs to be connected to chips through high bandwidth networks such as GB300 NVL72. WideEP disperses 896 expert models across multiple GPUs, requiring more than 120 forward computations per operation. The scalability requirements brought by large-scale expert models exceed the bandwidth saved by KDA. A more efficient attention mechanism may drive the context length from 1 million tokens to over 5 million tokens, and the efficiency improvement will expand the scale of AI usage and increase network demand.