Community
Looped transformer study profiles HBM weight reuse in Huginn on a B200
Profiling shows a shared 72 MiB weight chain ran 24% faster than identical copies with almost no HBM reads, while the advantage vanished at 288 MiB, larger than L2. A full Huginn Core read about eight times the HBM data over eight iterations.
Every outlet that ran it 1 article
r/MachineLearningCommunityLead
Do looped transformers save HBM traffic? Profiling weight reuse in Huginn on a B200 [P]