Huawei’s chip cadence has moved onto a new track across the board.

After Richard Yu showed off a Kirin chip at the launch of Huawei’s tri-fold phone, the company’s AI computing power chips now have a fresh update.

At Huawei Connect, rotating chairman Eric Xu unveiled what the company calls the world’s most powerful computing power SuperNodes and clusters.

The Atlas 950 SuperPoD and Atlas 960 SuperPoD SuperNodes support 8,192 and 15,488 Ascend cards, respectively.

The Atlas 950 SuperPoD is expected to launch in the fourth quarter of 2026, fully surpassing NVIDIA’s NVL576, which is expected in 2027, and maintaining the world’s top computing power position for the next two years.

Huawei also released what it calls the world’s most powerful SuperNode clusters: the Atlas 950 SuperCluster and Atlas 960 SuperCluster, with computing power scales of more than 500,000 cards and up to 1 million cards, respectively, likewise securing the top global cluster position.

Huawei also laid out the next two years of evolution plans for Ascend chips and Kunpeng chips, including the Ascend 950 series, Ascend 960 series, Kunpeng 950 and Kunpeng 960.

Huawei’s rotating chairman also acknowledged that, because of process node and tape-out constraints, “there is a gap with NVIDIA in single-chip performance in the short term”...

But he said Huawei can use extreme system architecture and interconnect technologies to integrate large numbers of chips into a “supercomputer,” achieving a comprehensive lead at the cluster level.

Ascend 950PR Launches in Q1 Next Year, Using Huawei’s In-House HBM

On chips, Huawei will stick to a cadence of “one generation per year, doubling computing power,” continuously evolving data formats and bandwidth technologies to meet the unlimited growth in demand for AI computing power.

On that basis, Huawei disclosed the roadmap for the Ascend 950 series, Ascend 960 series and Ascend 970 series.

Ascend 950 Series

Highlights of the Ascend 950 chip architecture include:

New support for low-precision data formats

Higher vector computing power

A 2.5-fold increase in interconnect bandwidth

Support for Huawei’s in-house HBM

The specific chips are as follows:

Ascend 950PR

Positioning: an inference-optimized chip for recommendation and prefill scenarios.

Huawei is launching this chip because, as AIGC develops, input contexts are getting longer and consuming more compute resources. In e-commerce, content platforms and social media, recommendation algorithms need higher accuracy and lower latency.

Technical innovation: it uses Huawei’s in-house HBM memory solution, which can significantly reduce costs.

Product form factors: standard cards and SuperNode servers.

Launch timing: first quarter of 2026.

Ascend 950DT

Positioning: for training and deep learning scenarios.

Technical innovations:

Uses in-house HBM: HiZQ 2.0

Memory capacity of up to 144GB, with access bandwidth of 4TB/s.

Interconnect bandwidth increased to 2TB/s.

Supports multiple data formats including FP8, MXFP8, MXFP4 and HF8 to improve training efficiency.

Launch timing: fourth quarter of 2026.

Ascend 960 (Planned)

Positioning: a flagship training chip, with all major specifications doubled from the Ascend 950.

Computing power, memory capacity, access speed and the number of interconnect ports are all doubled.

It supports Huawei’s in-house Hi-F4 data format, currently the industry’s best implementation of 4-bit precision, further improving inference throughput while delivering better inference accuracy than industry FP4 solutions.

Launch timing: fourth quarter of 2027.

Ascend 970 (Planned)

Positioning: a fully upgraded training chip, with major gains across all metrics.

Preliminary specifications: compared with the Ascend 960, FP4 and FP8 computing power are both doubled, while memory access bandwidth increases by at least 1.5 times.

Launch timing: fourth quarter of 2028.

Atlas 950: Expected to Stay World No. 1 for the Next Two Years

Huawei vice chairman and rotating chairman Eric Xu said candidly that Huawei faces constraints in single-chip manufacturing, but that with 30 years of accumulated work in foundational software and system architecture, it can use extreme computing power architecture and interconnect technologies to integrate large numbers of chips into a “supercomputer,” thereby surpassing rivals at the cluster level.

A SuperNode integrates large numbers of compute units, such as chips and servers, through high-speed interconnect technologies into one logically unified giant AI computer that can learn, think and reason like a single computer. It is becoming a new paradigm for AI infrastructure.

Huawei’s existing Atlas 900 A3 SuperNode supports 384 Ascend 910C chips. Its maximum computing power reaches 300 PFlops, and it remains the world’s most powerful AI SuperNode by computing power. Since launch, more than 300 units have been delivered, serving over 20 customers.

Atlas 950 SuperNode

Huawei’s latest release this time is the Atlas 950 SuperNode.

It supports 8,192 Ascend 950DT chips, more than 20 times the scale of the Atlas 900.

It consists of 128 compute cabinets and 32 interconnect cabinets, covers about 1,000 square meters and uses all-optical connections.

Key metrics are as follows:

FP8 computing power: 8 EFlops

FP4 computing power: 16 EFlops

Interconnect bandwidth: 16.3 PB/s, more than 10 times the total bandwidth of the global internet

Memory capacity: 1,152 TB

Total training throughput: 4.91 million TPS, 17 times higher than the Atlas 900

Total inference throughput: 19.6 million TPS, 26.5 times higher than the Atlas 900

Compared with NVIDIA’s NVL144, due to launch in 2025, its card scale is 56.8 times larger, total computing power is 6.7 times higher, memory capacity is 15 times larger and interconnect bandwidth is 72 times higher.

The Atlas 950 SuperNode is scheduled to launch in the fourth quarter of 2026. It is expected to maintain the world’s top computing power position for the next two years.

Atlas 960 SuperNode

Huawei also released the Atlas 960 SuperNode.

It is based on the Ascend 960 and Ascend 950DT chips, supports up to 15,488 cards, and uses all-optical interconnects across cabinets.

Key metrics double again from the Atlas 950:

FP8 computing power: 30 EFlops

FP4 computing power: 60 EFlops

Memory capacity: 4,460 TB

Interconnect bandwidth: 34 PB/s

Large-model training and inference performance is 3 to 4 times higher than the Atlas 950.

It is expected to launch in the fourth quarter of 2027.

Huawei Creates Lingqu, an Interconnect Protocol for SuperNodes

SuperNode technology is not only for AI. It is also reshaping general-purpose computing.

Huawei released the Kunpeng 950 and Kunpeng 960 chips and their corresponding SuperNodes.

Kunpeng 950 processor:

Versions: 96 cores/192 threads; 192 cores/384 threads

Features: supports confidential computing and adds four layers of security isolation.

Based on the Kunpeng 950, Huawei has built the TaiShan 950 SuperNode.

It will be the world’s first general-purpose computing SuperNode. It supports up to 16 nodes, 32 processors and 48TB of maximum memory, while also supporting pooling of memory, SSDs and DPUs.

The GaussDB read-write architecture built on the TaiShan 950 SuperNode does not require distributed transformation of the database, and boosts performance by 2.9 times.

It can ultimately replace traditional databases on mainframes and minicomputers smoothly. TaiShan 950 plus distributed GaussDB will become the terminator of all kinds of mainframes and minicomputers, fully replacing mainframes, minicomputers and Oracle’s Exadata database servers across application scenarios.

Beyond core database scenarios, the TaiShan 950 SuperNode also performs strongly in broader use cases: for example, memory utilization in virtualization environments rises by 20%, while real-time data processing time in Spark big data scenarios is shortened by 30%.

Launch timing is the first quarter of 2026.

The biggest challenge in building a 10,000-card SuperNode lies in interconnect technology. Through systematic innovation, Huawei said it has solved two major problems:

First, how to achieve long distance and high reliability. Large-scale SuperNodes have many cabinets and long connection distances between cabinets, and current electrical and optical interconnect technologies cannot meet the requirements. Current electrical interconnect technology has short connection distances at high speeds and can support at most two-cabinet interconnects, while current optical interconnect technology can connect multiple cabinets over long distances but cannot meet reliability requirements.

Second, how to achieve high bandwidth and low latency. Current cross-cabinet inter-card interconnect bandwidth is low, with a gap of five times versus SuperNode requirements; cross-cabinet inter-card latency is high, with the best current interconnect technologies reaching only about 3 microseconds, still 24% short of the Atlas 950/960 design requirement. When latency is already as low as 2 to 3 microseconds and close to physical limits, even an improvement of 0.1 microsecond is a major challenge.

On this basis, Huawei created the Lingqu (UnifiedBus) interconnect protocol and formally opened the Cloudqu 2.0 technical specification, inviting industry partners to build the ecosystem and advance the SuperNode industry.

Finally, Huawei unveiled another major move: the Atlas 950 SuperPlus cluster.

It consists of 64 Atlas 950 SuperNodes connected in parallel, integrating 520,000 Ascend 950T chips.

Total computing power reaches 524 EFlops. It supports both UBOE and RoCE networking protocols, with UBOE offering advantages in latency, reliability and cost.

Launch timing is the fourth quarter of 2026.

The Atlas 960 SuperPlus cluster is also in planning.

It will reach the million-card scale, with total FP8 computing power of 2 ZFlops and FP4 computing power of 4 ZFlops. Launch timing is the fourth quarter of 2027.

Finally, Xu emphasized that Huawei will use SuperNodes and clusters based on Lingqu to keep meeting the rapidly growing demand for computing power, driving the continued development of artificial intelligence and creating greater value.