Zixuan PNM Architecture Delivers 30TB/s Bandwidth, Crushing HBM and B200

New Purple Light Group has announced a memory architecture called Zixuan PNM that uses 3D DRAM and a 3.5D heterogeneous integration approach. The company claims the architecture achieves a storage bandwidth of 30TB/s, which it says surpasses HBM3e's 1.2TB/s and NVIDIA B200's 8TB/s.

Architecture uses 3D DRAM and 3.5D integration

The PNM architecture supports near-memory computing, which the company says reduces memory access latency by up to 18 times. In simulations, the architecture reportedly delivers 1.5 to 2 times higher token throughput than NVIDIA's B200 series at equivalent compute power.

The technology is based on a domestic supply chain and is said to be scalable for mass production without being subject to foreign technology restrictions. Purple Light Group has not announced a production timeline or availability date for the architecture.

The Zixuan PNM architecture targets AI workloads where memory bandwidth and latency are critical. By integrating 3D DRAM with a 3.5D heterogeneous package, the design aims to address bottlenecks in current memory subsystems.