BlockBeats News, July 19th - Semiconductor and AI research firm SemiAnalysis published an article stating that although about three-quarters of the Kimi K3's network layer uses KDA, compared to a full global attention model, it can reduce KV cache transfer bandwidth by up to 10 times. However, this does not mean that the AI network switch market will shrink significantly.

The Kimi K3 has 2.8 trillion parameters, and even with MXFP4, each forward computation still requires about 1.5TB of HBM bandwidth. To achieve profitable deployment while maintaining a reasonable interaction speed, it is still necessary to connect a large number of chips through high-bandwidth network connections such as GB300 NVL72 and rely on WideEP extended services.

WideEP distributes 896 expert models to multiple GPUs and executes Token distribution and result merging twice at each layer and each forward computation, requiring more than 120 executions per forward computation. In contrast, KV cache transfer between prefilling and decoding occurs only once per dialogue round, so the bandwidth saved by KDA may be much less than the expanded network demand brought by large-scale expert models.

SemiAnalysis believes that a more efficient attention mechanism may also drive the context length from 1 million Tokens to over 5 million Tokens. According to the Jevons Paradox, efficiency improvements may expand AI usage scale, further increasing network demands.