Toggle light / dark theme

Alibaba Brags That Its Next-Gen Zhenwu V900 AI Chip Offers More On-Package Memory Than NVIDIA H200, Outlines Plans To Deploy 20 GW Of Compute, And Teases 4–10 trillion parameters For Upcoming Qwen Models

Alibaba is now claiming that the upcoming Zhenwu V900 chip will have 216GB of on-package memory, with an official launch slated for Q1 2027. We can only theorize that the V900 will leverage CXMT’s HBM3E solution, especially given their overlapping volume production timelines. The accelerator will sport chip-to-chip interconnect speeds of around 1.2 TB/s via Alibaba’s ICN Switch fabric, and offer around 3x the performance of Zhenwu M890, replete with native FP8/FP4 support. This means that each accelerator will offer a peak computing power of around 1.8 PFLOPS at FP16, given the ~0.6 PFLOPS that M890 had claimed to offer.

Critically, the ICN Switch can allow around 1,000 Zhenwu V900 chips to function as a single accelerator. However, Alibaba is now claiming that each V900 cluster can scale to 500,000 chips, entailing a whopping 108 petabyte of memory across the entire cluster! It remains to be seen if CXMT can fulfill the entirety of this oncoming demand.

Also, Alibaba is now offering its own bespoke rack-scale solution, replete with Yitian CPUs, Zhenwu V900 GPUs, ICN interconnect, Pangu NICs, and Zhenyue storage controllers.

Leave a Comment

Lifeboat Foundation respects your privacy! Your email address will not be published.

/* */