Alibaba Unveils Zhenwu M890 AI Chip With 144GB HBM3 Memory And 0.6 PFLOPs FP16 Compute Performance

Alibaba Cloud unveiled the Zhenwu M890 AI chip and Qwen3.7-Max large language model at an event in Hangzhou on May 20, 2026.

New AI chip with custom PPU architecture and PCIe 5.0 support for training and inference workloads.

The M890 processor features a custom Parallel Processing Unit architecture with 144 GB of HBM3 memory and delivers 0.6 PFLOPs of FP16 compute performance. The chip supports PCIe 5.0 x16 connections and handles FP32, FP16, FP8, and FP4 formats for AI workloads.

Spec comparison

Alibaba Cloud introduced the ICN Switch 1.0 interconnect chip to support the M890, providing 25.6 Tb/s of bandwidth with a peer-to-peer delay under 150 nanoseconds. The company will integrate up to 128 M890 accelerators into the Panjiu AL128 Supernode Server rack.

Alibaba Cloud claims the Zhenwu M890 offers three times the compute performance of previous generation offerings and matches the A100 from NVIDIA in raw compute speed. The company also states the chip runs three times faster than the Hopper H20 solution, though these figures rely on internal testing without independent third-party verification.

The Zhenwu M890 will become available globally through Alibaba's Model Studio platform for developers soon. Alibaba Cloud has shipped approximately 560,000 Zhenwu AI chips to date across twenty industries and outlined a roadmap that includes the V900 chip in Q3 2027 and the J900 model in Q3 2028.