SUNMMIO, a chip startup based in Beijing, has completed Pre-A2 and Series A funding rounds to advance three-dimensional stacked inference processors toward volume manufacturing. The company declined to disclose investment amounts.
SUNMMIO develops 3D-stacked chips purpose-built for large-model inference. The financing will be allocated to expanding its research and development team, refining the technology and scaling production of its TokenPU processors, the company said.
The company’s debut A4E chip reached tape-out completion in June. The architecture stacks eight memory layers atop a logic wafer, delivering claimed memory bandwidth of 16 terabytes per second, a metric that targets the constraint now limiting inference throughput.
Inference performance increasingly hits a wall not at compute limits but at how quickly data can reach arithmetic units. Conventional packaging approaches struggle to alleviate this bottleneck. Stacking memory directly above logic shortens signal travel distance and widens the interconnect path between them, which circumvents constraints conventional arrangements impose.
SUNMMIO joins a growing cohort of Chinese chip firms concentrating on inference rather than training. That emphasis mirrors both commercial dynamics and regulatory realities: while access to cutting-edge training silicon remains difficult due to export controls, inference demand is surging as deployed models serve actual users, and the performance requirements differ sufficiently to create openings for novel architectures.
Volume manufacturing will determine whether the design’s advertised bandwidth translates to production reality.
