Huawei Ascend 950DT

The Ascend 950DT is Huawei's decode-and-training oriented member of the Ascend 950 generation, and the chip behind both the air-cooled Atlas 650E server and the liquid-cooled Atlas 950 SuperPoD. Huawei publishes it in three bins by Cube Core count — 36, 32 or 28 cores, with twice as many Vector Cores in each case — giving 2,007, 1,784 or 1,561 TFLOPS of MXFP4, 1,034/919/804 TFLOPS of HiF8, MXFP8 and FP8, the same figures again as INT8 TOPS, 547/486/425 TFLOPS of BF16 and FP16, and 273/243/212 TFLOPS of TF32. The figures here are the top bin, matching the 144 GB configuration; memory comes in either 144 GB or 96 GB at 4 TB/s, and both are real shipping configurations, which is why the Atlas 650E and Atlas 950 SuperPoD both specify 96 GB while this chip is listed at 144. The bins are visible in Huawei's own products: the air-cooled Atlas 650E works out at exactly the lowest bin on all three of its published precisions, while the Atlas 950 SuperPoD's "up to 2 EFLOPS mxFP4" across 1,024 cards is the top bin rounded down. Beyond the tensor units the chip carries 128 MB of L2 cache with sector and cache-hint support, a Linx816 AI CPU subsystem, eighteen Unified Bus ports at 112 Gbps for 2,016 GB/s bidirectional, two 400 Gbps UBoE ports shared with those, and PCIe 5.0 x16. Huawei labels none of these figures dense or sparse and publishes no sparse rate.