Huawei Ascend 960
Product Overview
Huawei Ascend 960 is the fifth-generation Ascend AI chip, officially announced on September 17, 2026 by Rotating Chairman Wang Tao at HUAWEI CONNECT 2026 (Shanghai). The biggest difference from its predecessor: the 960 is no longer a roadmap teaser, but a complete commercial combination of dual versions (960DT training / 960PR inference) + dual SuperPods (Atlas 860 air-cooled / Atlas 960 liquid-cooled), and both chips are ready ahead of schedule — the 960DT three quarters early, the 960PR one quarter early.
Also announced: the world's first SuperPod using NPO (near-package optics) technology — the Ascend 960 SuperPod: 4096 interconnected cards per node, 8 EFLOPS FP8, 1PB HBM, built on the "UnifiedBus + Hi-ONE optical engine". Huawei also confirmed the one-chip-per-year cadence: Ascend 970 in 2028, Ascend 980 in 2029.
⚠️ Information note: single-chip specs on this card follow the official HC2026 disclosures. In August 2026, the Digital China Summit reported that "the Atlas 960 SuperPoD supports 15,488 Ascend cards"; in September the official figure was 4096 cards per SuperPod (a revision of the earlier 15,488 claim; this card follows the official press release).
Core Specifications
The following are the Ascend 960DT (training version, officially disclosed values):
| Parameter | Value |
|---|---|
| Architecture | Da Vinci v6 (Ascend 6th generation) |
| FP8 Compute | 2 PFLOPS |
| FP4 Compute | 4 PFLOPS |
| Memory | 288 GB (in-house HBM, max configuration) |
| Memory Bandwidth | 9.6 TB/s |
| TDP | 700 W (presumed, not officially disclosed) |
| Debut | 2027 Q2 |
📌 Compared with the 950 series (FP8 1 PFLOPS / FP4 2 PFLOPS), every key metric of the 960 doubles: FP8 2→x2, FP4 4→x2, memory 288GB vs 144GB (950DT), bandwidth 9.6TB/s vs 4TB/s. Huawei calls this continuous doubling path "Tao's Law".
⚠️ TDP is a third-party estimate; the official figure has not been published. The existence of an air-cooled version (Atlas 860) implies per-card power must stay within what air cooling can handle.
960DT vs 960PR Dual Versions
| Dimension | 960DT (training) | 960PR (inference) |
|---|---|---|
| FP8 Compute | 2 PFLOPS | Not separately disclosed |
| Ready | 2027 Q1 (3 quarters early) | 2027 Q3 (1 quarter early) |
| Launch | 2027 Q2 | 2027 Q3 |
| Paired SuperPod | Atlas 860 (air-cooled) | Atlas 960 (liquid-cooled) |
| Positioning | Ten-trillion-parameter model training + high-concurrency inference | Large-scale inference deployment |
Ascend 960 SuperPod (world's first NPO SuperPod)
| Metric | Ascend 950 SuperPod | Ascend 960 SuperPod |
|---|---|---|
| Cards per node | 1024 cards (shown at WAIC 2026) | 4096 cards |
| FP8 Compute | 8 EFLOPS (8192-card full config) | 8 EFLOPS (single node) |
| FP4 Compute | 16 EFLOPS (8192-card full config) | 16 EFLOPS (single node) |
| Total HBM | 1152 TB (8192-card full config) | 1 PB (4096 cards) |
| Interconnect RTT | 3 μs | 2 μs (domain-wide D2D) |
| Optical interconnect | UnifiedBus 2.0 | NPO optical engine Hi-ONE (industry's first mass production) |
NPO (Near-Package Optics) Key Figures
- Hi-ONE optical engine: 7.2T transmission capacity per engine, the industry's first mass-produced NPO product, the highest current transmission capability, and the only one with a built-in light source
- 5500 Hi-ONE units replace 48000 800G optical modules, cutting power by more than 550 kW
- System mean time between failures doubled, overall availability at 99.8%
- Why NPO over CPO: NPO keeps optical engines independent while enabling short-distance opto-electric handoff, avoiding CPO's reliability, manufacturability, and serviceability problems, and preserving the existing optical module industry ecosystem
- MFU gain (Huawei Markov Lab simulation): in a 100k-card cluster of 4K SuperPods, MFU improves 2.75x versus 8-card server networking — in traditional architectures, intra-cluster communication takes over 40% of training time
Cluster Scaling Path
- Multiple 960 SuperPods interconnected via UnifiedBus or RoCE: up to 512,000 cards (two-tier CLOS four-plane fabric)
- With multi-rail topology: up to 1 million cards in an Ascend SuperPod cluster
- Kunpeng SuperPod upgraded in step: all-optical networking up to 4096 nodes, 256TB unified memory pool
Annual Cadence Roadmap (Tao's Law)
| Year | Chip | Status |
|---|---|---|
| 2026 | Ascend 950DT / 950PR | In volume commercial use |
| 2027 | Ascend 960DT / 960PR | Ready ahead of schedule, Q2 / Q3 launches |
| 2028 | Ascend 970 | Announced, compute specs continue to double |
| 2029 | Ascend 980 | Announced, memory bandwidth / capacity / interconnect bandwidth all up sharply |
Ecosystem and Deployment Status
- Deployment base: over 1000 Ascend 910C SuperPods deployed and the Ascend 950 SuperPod in volume commercial use — the 960 climbs from this two-step foundation
- CANN ecosystem crosses the tipping point: external developers now account for 61% (surpassing internal teams for the first time), monthly active developers past 5200, and CANN is fully open-sourced with regular community operations
- Kunpeng + Ascend: over 7.8 million developers gathered, with 20,000+ industry solutions incubated
- Storage companion: the OceanStor M900 cluster (PB-scale KV cache one hop away over the UnifiedBus) released in step
Competitor Comparison
| Metric | Ascend 960DT | NVIDIA B200 | NVIDIA Rubin R200 | AMD MI455X |
|---|---|---|---|---|
| FP8 compute | 2 PFLOPS | 4.5 PFLOPS | 12.5 PFLOPS | 2.3 PFLOPS |
| FP4 compute | 4 PFLOPS | 9 PFLOPS | 50 PFLOPS | 4.6 PFLOPS |
| HBM capacity | 288 GB | 192 GB | 288 GB | 288 GB |
| HBM bandwidth | 9.6 TB/s | 8 TB/s | 16 TB/s | 12 TB/s |
| Process | SMIC domestic (unconfirmed) | TSMC 4NP | TSMC 4NP | TSMC 3NM |
| Launch | 2027 Q2 | Available | Available | Available |
The 960's single-card absolute performance still lags NVIDIA's flagship by a generation (Rubin R200 FP8 is about 6x higher); Huawei's strategy is to close the gap with SuperPod system-level capabilities (unified memory addressing, 2.75x MFU, NPO optical interconnect) — evaluating domestic compute should not come down to a single-chip PFLOPS comparison.
Use Cases
- ✅ Ten-trillion-parameter LLM training (4096-card SuperPod + unified memory addressing)
- ✅ High-concurrency inference (inference latency down ~70% and training throughput up ~2.3x vs the predecessor, official figures)
- ✅ Standard server room deployment (Atlas 860 air-cooled SuperPod, no liquid cooling retrofit needed)
- ❌ Single-card purchase (delivered as SuperPod / cluster)
- ❌ Deployment within 2026 (chips launch in 2027)
Vendor Information
| Parameter | Value |
|---|---|
| Manufacturer | Huawei Technologies Co., Ltd. (HiSilicon) |
| Website | https://www.hiascend.com |
| Announced | 2026-09-17 (HUAWEI CONNECT 2026, Shanghai) |
| 960DT launch | 2027 Q2 |
| 960PR launch | 2027 Q3 |
References
- Huawei website: Launching the world's first NPO-based SuperPod (Ascend 960 SuperPod)
- EEWorld: Huawei Ascend 960 chip to launch ahead of schedule
- Huanqiu.com: Huawei launches the Ascend 960 SuperPod, accelerating ten-trillion-scale model training and inference
Related Products
- Huawei Ascend 950 - Previous SuperPod generation (8192-card plan)
- Huawei Ascend 950DT - Previous training card (FP8 1 PFLOPS)
- Huawei Ascend 950PR - Previous inference card
- Huawei Ascend 970 - Next generation (announced for 2028)
- Full comparison table