T-Head Zhenwu M890 NEW

The Zhenwu M890 is T-Head's second-generation AI accelerator, shown for the first time at the 2026 Alibaba Cloud Summit on 20 May 2026 inside the Panjiu AL128 supernode. It carries 144 GB of HBM, eight independent ICN ports with 800 GB/s of inter-chip bandwidth and a PCIe 5.0 x16 host interface, and it is the generation where T-Head added FP8 and FP4 to a range that already covered FP32, BF16 and FP16, so that one part can serve high-precision training and very low precision inference. It is also the generation where the interconnect stopped being only a mesh: an ICN Switch makes 64 cards behave as one flat full-bandwidth domain, and Alibaba packs 128 of them into a single rack with peer-to-peer latency below 150 nanoseconds. T-Head claims three times the performance of the Zhenwu 810E and publishes no figure on either side of that ratio, which is why every precision on this row is listed as supported with no rate.