AnswersPublished16 sources
Huawei Switches On 1,024-NPU Ascend 950DT Cluster in Wuhan
The 1,024 card Ascend 950DT SuperPoD was switched on in Wuhan on October 28, 2026, at a Hubei provincial AI conference held during Huawei’s 2026 China roadshow. Launch coverage reports 1 EFLOPS of FP8 compute, 256TB of globally unified memory and a one month project timeline.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- The 1,024 card Ascend 950DT SuperPoD was switched on in Wuhan on October 28, 2026, at a Hubei provincial AI conference held during Huawei’s 2026 China roadshow.
- Launch coverage reports 1 EFLOPS of FP8 compute, 256TB of globally unified memory and a one month project timeline.
- Those specifications are launch claims; published coverage does not provide independent results for sustained performance, power use or workload costs.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What are the key details and significance of the first Huawei Ascend 950DT 1,024-NPU SuperPoD cluster activated in Wuhan on October 28, 2026. Article summary: Huawei’s 1,024-NPU Ascend 950DT SuperPoD was reportedly activated in Wuhan on October 28, 2026, at the Hubei Provincial AI Innovation and Development Conference during Huawei’s 2026 China roadshow—a milestone in deployin. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
Huawei’s 1,024-NPU Ascend 950DT SuperPoD was switched on in Wuhan on October 28, 2026, at a Hubei provincial AI conference held as part of Huawei’s 2026 China roadshow. Reports describe it as China’s first 1,024-card cluster based on the Ascend 950DT. Its published specifications are launch claims, not independent measurements of sustained performance. 4
6
What went live in Wuhan?
Hubei Science and Technology Investment Group invested in and built the cluster, according to launch coverage. The system was switched on at the Hubei Provincial AI Innovation and Development Conference. Reports say the project took one month from its start to switch-on—but that does not mean an entire data centre was built from scratch in that time. 4
6
The installation is described as one liquid-cooled SuperPoD containing 1,024 Ascend NPU accelerator cards, arranged across compute and interconnect cabinets. It is a multi-cabinet system, not a single conventional server. 6
Reported specifications
Launch coverage lists 1 EFLOPS of FP8 compute, 256TB of globally unified memory, terabyte-per-second-level bandwidth between cards and microsecond-level end-to-end communication latency. These are reported system specifications; the available coverage does not provide independent benchmark results for sustained training or inference workloads. 4
6
How UnifiedBus fits in
The system uses Huawei’s Lingqu architecture, also referred to as UnifiedBus. Launch reports say the cards can access one another directly. Huawei describes UnifiedBus as a common interconnect designed to speed data access and reduce communication delays between computing cards. 6
25
The aim is to let many accelerators operate as a closely connected system for large workloads. That does not make them literally one processor or eliminate the need for software coordination and data movement. The practical benefit depends on performance on specific workloads—something the launch figures alone do not establish.
What the cluster is intended to support
Launch coverage says the system is intended to support trillion-parameter model training and AI-agent inference, while providing local computing capacity for Hubei developers and businesses. Reports also describe it as helping address a shortage of high-end computing capacity in the province, but do not quantify that shortage or show how much this installation will resolve. 4
6
The Wuhan activation is a deployment milestone, not the first public appearance of the Ascend 950 architecture. Huawei’s Ascend 950 SuperPoD was publicly shown at the 2026 World Artificial Intelligence Conference in July. Coverage of that event also reported more than 750 commercial deployments of the previous-generation Ascend 384 SuperPoD. 43
What “first” means—and what remains unverified
Reports call Wuhan China’s first Ascend 950DT cluster with 1,024 cards. Separately, another report says an Atlas 950 1,024-card liquid-cooled system was powered on on October 15. The available reporting does not establish whether these claims refer to different configurations or deployment milestones. So “first” is best understood as the claim attached to the Wuhan 950DT installation—not as an uncontested first for every 1,024-card Ascend 950 system. 5
6
9
Launch coverage does not provide measured power draw, sustained utilisation, workload throughput or cost per workload for the Wuhan installation. Its stated FP8 capacity alone cannot establish real-world training speed or energy efficiency. 4
6
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Huawei Switches On 1,024-NPU Ascend 950DT Cluster in Wuhan"?
The 1,024 card Ascend 950DT SuperPoD was switched on in Wuhan on October 28, 2026, at a Hubei provincial AI conference held during Huawei’s 2026 China roadshow.
What are the key points to validate first?
The 1,024 card Ascend 950DT SuperPoD was switched on in Wuhan on October 28, 2026, at a Hubei provincial AI conference held during Huawei’s 2026 China roadshow. Launch coverage reports 1 EFLOPS of FP8 compute, 256TB of globally unified memory and a one month project timeline.
What should I do next in practice?
Those specifications are launch claims; published coverage does not provide independent results for sustained performance, power use or workload costs.
Sources
163.com全国首个华为昇腾950DT超节点集群在武汉上线
ithome.com全国首个华为昇腾950DT 超节点集群在武汉上线
huawei.comHuawei UnifiedBus: Helping thousands of computing cards ...
news.metal.com[SMM Flash] Huawei Ascend 950DT 10,000-card pool + 1,024-card liquid-cooled super node goes live - Shanghai Metals Market (SMM)