Ep. 038 - GLM 5.3 on AMD, NVIDIA, and Chinese Chips From Day One (InferenceX) | Kimbo Chen, Wenyao Gao, Jordan Nanos

Efficient open models like GLM 5.3 don't reduce memory demand — they raise concurrency and sustain HBM/DRAM bandwidth needs.

• AMD briefly matched NVIDIA: On InferenceX, AMD's MI355X reached cost parity with NVIDIA's B200 at 150 tokens/sec interactivity for GLM 5.3, a result that surprised even the SemiAnalysis team.
• Sparse attention cuts bandwidth, not capacity: GLM 5.3 uses DeepSeek-derived sparse attention (lightning indexer + MLA) to slash compute cost per million tokens, but all tokens must still be loaded into HBM at some point, so memory capacity demand doesn't fall — contradicting the 2023 fear that efficient models kill memory demand.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.