Qualcomm's Durga Malladi: HBC vs HBM, Dragonfly AI 250, and Winning on TCO

Qualcomm's HBC near-memory architecture targets the memory-bandwidth bottleneck, aiming for TCO advantages in AI inference by 2027's AI250 launch.

• HBC defined: Qualcomm's High Bandwidth Compute stacks DRAM directly atop a logic "base die" doing near-memory compute, cutting latency and power versus HBM's wide-bus shuttling to a separate accelerator.
• Memory wall problem: GPU compute has grown 2-3x per generation over six years while memory bandwidth lagged, leaving compute idle waiting on data, especially during inference decode.

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.