● Covered by 1 source · 2 reportsMedium impact2 neutral
Huawei updates AI accelerator roadmap, accelerates Ascend 960, introduces 970/980
Key points
- Huawei updated its AI hardware roadmap at Huawei Connect.
- Accelerated Ascend 960 roadmap, introduced Ascend 970 and 980 specifications.
- Transitioning to new SIMD+SIMT architectures for Ascend NPUs.
- Ascend 950 platform is in commercial use, but production capacity is limited.
- Ascend 900-series AI accelerators will be offered only in China.
- Huawei lacks capacity to satisfy domestic demand for Ascend 900-series.
- Eric Xu, Huawei rotating chairman, confirmed the China-only focus.
- Ascend 960DT and 960PR increase FP8 training to 2 PFLOPS.
- Ascend 960DT and 960PR increase FP4 inference to 4 PFLOPS.
Updated AI Hardware Roadmap
Huawei announced updates to its AI hardware roadmap at the annual Huawei Connect event. The company accelerated the timeline for its next-generation Ascend 960 accelerators and disclosed specifications for the upcoming Ascend 970 and 980 models. This update also included the introduction of the Peerium architecture, based on the UnifiedBus, and an expansion of its vertically integrated AI infrastructure portfolio.
Architectural Transition
Huawei is moving from its decade-old SIMD architectures to new SIMD+SIMT architectures for its Ascend NPUs. This new design combines vector-based processing (SIMD) with thread-level parallelism (SIMT) to enhance hardware utilization and performance across various AI workloads, particularly for data-parallel and branch-heavy operations. The Ascend 950PR (prefill and recommendation) and Ascend 950DT (decoding and training) are the first NPUs to adopt this new architecture.
Ascend 950 Adoption and Challenges
The Ascend 950 platform is gaining traction, with Atlas 950 SuperPoD systems reportedly in large-scale commercial use. Huawei stated that tests of its training-oriented Ascend 950DT have yielded "good results," and it anticipates Chinese AI developers will begin training models on 950DT-based systems next year. However, Huawei acknowledged that its production capacity is currently insufficient to meet domestic demand.
Atlas 950 SuperPoD Configuration
In September 2025, Huawei announced the maximum Atlas 950 SuperPoD configuration, including 2,048 Kungpeng 950 CPUs, 8,192 Ascend 950DT NPUs, 160 cabinets, 8 FP8 EFLOPS, 16 FP4 EFLOPS, and 16 PB/s of aggregate interconnect bandwidth. However, a public demonstration in July 2026 showed a smaller implementation with 256 CPUs and 1,024 accelerator cards, which is below the maximum configuration. The company still describes the architecture as scaling up to 8,192 NPUs, but this is not listed on its website.
Market Adoption and Future Outlook
The adoption rate of the Atlas 950 SuperPoD does not appear rapid, potentially due to supply constraints or the need for significant software redesign for the new architecture. This initial phase with the Atlas 950 SuperPoD is expected to pave the way for future developments and broader adoption of Huawei's AI infrastructure.
Updates
- Ascend 900-series AI accelerators will be offered only in China.
- Huawei lacks capacity to satisfy domestic demand for Ascend 900-series.
- Eric Xu, Huawei rotating chairman, confirmed the China-only focus.
- Ascend 960DT and 960PR increase FP8 training to 2 PFLOPS.
- Ascend 960DT and 960PR increase FP4 inference to 4 PFLOPS.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
How outlets covered it
Huawei will not offer its next-generation Ascend 900-series AI accelerators internationally, citing insufficient capacity to meet demand within China. This decision means the chips, which show performance gains but lag behind Nvidia's top offerings, will primarily serve the Chinese market.
Huawei updated its AI hardware roadmap at Huawei Connect, accelerating the Ascend 960 timeline and introducing Ascend 970 and 980 specifications. The company is transitioning to new SIMD+SIMT architectures for improved AI workload performance, despite current production capacity limitations.