Atlas 950 SuperPoD
#2 in AI Training HardwareHuawei · v950 · superpod · since Q4 2026 (kommerzielle Markteinführung); Prototyp-Enthüllung 16. Juli 2026 (WAIC) · 3× · last seen Jul 29, 2026
81
Momentum
The Atlas 950 SuperPoD is Huawei's AI training supernode system based on Ascend 950DT NPUs, interconnected via the proprietary UnifiedBus/Lingqu protocol to link up to 8,192 NPUs into a single logical computer. First shown as a 1,024-NPU prototype at WAIC 2026, its full configuration delivers 8 ExaFLOPS FP8 / 16 ExaFLOPS FP4 compute with 256TB of globally addressable unified memory. Commercial launch is scheduled for Q4 2026, targeting training and inference of trillion-parameter large language models. The Ascend 950DT chips are manufactured on SMIC's N+3 process (roughly 7nm-equivalent).
Momentum trend
01.05.30.07.
Features
| Manufacturing Process (nm) | SMIC N+3 process (roughly 7nm-equivalent) for Ascend 950DT chips |
| License | CANN (Compute Architecture for Neural Networks) fully open-sourced since August 2025; supports PyTorch and Triton |
| Platform | Ascend 950DT NPUs with UnifiedBus 2.0 interconnect (Lingqu protocol), CANN software stack |
| Price | No official list price published; Huawei positions it as a lower-cost alternative to Nvidia |
| Compute Performance (FLOPS/TOPS) | 8 ExaFLOPS FP8 / 16 ExaFLOPS FP4 in full configuration (8,192 NPUs) |
| Release Date | Commercial market launch Q4 2026 |
| Memory | 256TB global unified memory address space (full configuration) |
| Availability | Prototype publicly shown July 2026 (WAIC); commercial availability from Q4 2026, incl. planned South Korea market entry |
