Tigris | 会讲课教授是好老师 · 2026-01-01 · 帖子

TL;DR: A macro-architecture upgrade made trainable, proving the bottleneck is still bandwidth and memory.” DeepSeek just dropped a systems paper, not an “algo tweak.” Think of training as building a long highway. HC turns each layer into a multi-lane interchange: more routing freedom, more power, but it can destabilize deep training. mHC adds strict “traffic rules” so the mixing stays conservative and trainable at scale, while keeping the marginal overhead around ~6.7%. The real signal: the race in training isn’t over, and the bottleneck is still bandwidth + memory (the memory wall).