ChinaHoumo AI says in-memory compute cuts AI core power to a fifth.
Houmo says edge chips now run 120-billion-parameter models without sending data to the cloud.
Inference is limited by memory bandwidth, so Houmo computes inside the memory array, where weights stay on chip.
Houmo says rival 3D stacking designs face limited compute area and hard heat dissipation.
Houmo has mass-produced the M30 and M50 on its 2D route, and its first 3D chip is next on the roadmap.
How each outlet framed it
- Zhidx leans favorable
- ties Houmo's SRAM-CIM dual-path roadmap to the personal-agent boom, arguing local chips must absorb high-frequency tasks while cloud keeps long-horizon reasoning
Sources: Zhidx