Huawei Charts Ambitious Three-Year AI Chip Roadmap, Challenging Global Incumbents
Huawei Unveils Ambitious Three-Year AI Chip Roadmap, Challenging Global Incumbents
Huawei has officially announced a comprehensive three-year roadmap for its Ascend artificial intelligence (AI) chip series, signaling a determined effort to bolster China's domestic AI capabilities and directly challenge the dominance of established global players, most notably Nvidia. This strategic announcement, made at the company's annual All-Connect Conference, underscores Huawei's commitment to advancing its AI hardware development amidst a complex geopolitical landscape and increasing demand for powerful computing solutions.
Ascend Chip Series: A Phased Rollout for Enhanced AI Performance
The core of Huawei's announcement is the planned release of four advanced AI chips by 2028. The roadmap commences with the Ascend 950 series, which will feature two distinct variants: the Ascend 950 PR and the Ascend 950 DT. The Ascend 950 PR is slated for a first-quarter 2026 launch, targeting prefill inference and recommendation services. Following closely, the Ascend 950 DT is scheduled for a late 2026 release, designed to excel in decoding and training workloads. Both 950 variants will share the same Ascend 950 Die, promising substantial upgrades over the current Ascend 910C.
Key advancements in the 950 series include support for new low-precision data formats, such as FP8, MXFP8, MXFP4, and Huawei's proprietary HiF8 format. These innovations are expected to deliver significantly higher training efficiency and inference throughput, with theoretical peak performance reaching 1 PFLOPS for FP8, MXFP8, and HiF8, and 2 PFLOPS for MXFP4. Furthermore, the 950 series will boast enhanced vector processing capabilities, refined memory access granularity reduced to 128 bytes, and an impressive 2 TB/s interconnect bandwidth, representing a 2.5-fold increase compared to the Ascend 910C.
Huawei is placing a strong emphasis on its proprietary High-Bandwidth Memory (HBM) technology. The Ascend 950PR will integrate HiBL 1.0, a proprietary, cost-effective HBM solution designed to boost inference prefill and recommendation performance. In contrast, the Ascend 950DT will feature HiZQ 2.0 HBM, offering a substantial 144 GB of memory and 4 TB/s of memory access bandwidth, crucial for demanding training and decoding tasks.
The roadmap continues with the Ascend 960, scheduled for the fourth quarter of 2027. While specific details are still under development, Huawei has indicated that the Ascend 960 will double the compute power, memory capacity, and interconnect capacity compared to the Ascend 950. It will also introduce HiF4, a proprietary 4-bit precision format aimed at enhancing inference throughput beyond standard FP4 solutions.
Culminating the three-year plan, the Ascend 970 is targeted for release in the fourth quarter of 2028. Although still in the early stages of development, the objective is to double the FP4 and FP8 compute power relative to the Ascend 960, double the interconnect bandwidth, and increase memory bandwidth by at least 1.5 times. Huawei has committed to a one-year release cycle for its Ascend chips, with each iteration aiming to double compute performance.
Supernodes and Superclusters: Scaling AI Infrastructure
Beyond individual chip advancements, Huawei is also making significant strides in large-scale AI computing infrastructure with its "Supernode" and "SuperCluster" concepts. The Atlas 950 SuperPoD, expected in late 2026, is designed to integrate over 8,000 Ascend 950 DT chips, creating a powerful, unified computing system. Huawei asserts that this system will outperform Nvidia's NVL144 in key metrics such as compute, memory, and bandwidth. These SuperPoDs can then be interconnected to form even larger "SuperClusters," representing a significant expansion of AI computing capacity.
Huawei's strategy centers on providing robust ICT infrastructure, with a particular focus on these high-performance computing systems. This approach aims to meet the escalating and long-term demand for AI computing power, both within China and globally. The company views these large-scale clusters as a critical differentiator and a source of confidence in its ability to compete.
Strategic Context: Domestic Production and Market Dynamics
This ambitious roadmap is intrinsically linked to China's broader drive for technological self-sufficiency. U.S. sanctions imposed in 2020 significantly curtailed Huawei's access to advanced chip manufacturing capabilities, particularly from Taiwan Semiconductor Manufacturing Co. Ltd. (TSMC). This has necessitated a greater reliance on domestic production and innovation. The recent unveiling of the high-end Kirin 9020 processor in Huawei's Mate XTs smartphone, the first such chip in four years, signals progress in overcoming these production bottlenecks.
The announcement also arrives at a time when U.S. chipmakers, including Nvidia, are facing increased pressure and regulatory scrutiny in China. Reports indicate that Chinese cloud providers are actively exploring domestic alternatives to U.S.-made AI chips due to security concerns and antitrust investigations. Huawei's Ascend roadmap is strategically positioned to capitalize on this market shift, offering a viable alternative for Chinese enterprises seeking advanced AI hardware.
Ecosystem Development and Future Outlook
To foster wider adoption and accelerate the development of its AI ecosystem, Huawei plans to open-source its CANN (Compute Architecture for Neural Networks) platform by the end of 2025. This move is intended to make its AI hardware more accessible to developers and encourage the creation of applications and solutions built upon Huawei's infrastructure.
While acknowledging that gaps still exist between Huawei and industry leaders like Nvidia in terms of individual chip performance and ecosystem maturity, Huawei's leadership remains confident. Their strategy emphasizes leveraging their strengths in ICT infrastructure and developing integrated, large-scale computing solutions. The company's commitment to a consistent, performance-doubling release cycle for its Ascend chips, coupled with its focus on proprietary HBM and advanced interconnect technologies, positions it as a significant contender in the rapidly evolving AI hardware landscape.
AI Summary
Huawei has announced an ambitious three-year roadmap for its Ascend artificial intelligence (AI) chips, signaling a significant push to enhance China's domestic AI computing power and directly confront the global leadership of companies like Nvidia. The company plans to release four advanced AI chips by 2028: the Ascend 950 PR in Q1 2026, the Ascend 950 DT in late 2026, the Ascend 960 in late 2027, and the Ascend 970 in late 2028. This roadmap represents a strategic effort to overcome bottlenecks in local chip production, exacerbated by U.S. sanctions that cut off Huawei from key manufacturers like TSMC in 2020. The new Ascend chips are set to introduce significant performance improvements, including support for new low-precision data formats such as FP8, MXFP8, MXFP4, and Huawei's proprietary HiF8, promising higher training efficiency and inference throughput. Specific enhancements include a doubling of compute power with each annual release, increased interconnect bandwidth, and advanced High-Bandwidth Memory (HBM) technologies. The Ascend 950 series, for instance, will feature 2 TB/s interconnect bandwidth, a 2.5x improvement over the Ascend 910C. The 950PR variant is optimized for inference and recommendation services, utilizing proprietary HiBL 1.0 HBM for cost-effectiveness, while the 950DT is geared towards training and decoding workloads, boasting 144 GB memory and 4 TB/s memory access bandwidth. The subsequent Ascend 960 and 970 chips are designed to further double compute, memory, and interconnect capacities, with the 970 targeting a 1.5x increase in memory bandwidth. Beyond individual chips, Huawei is also focusing on integrated systems, unveiling plans for "SuperPoDs" and "SuperClusters." The Atlas 950 SuperPoD, expected in late 2026, will link over 8,000 Ascend 950 DT chips, aiming to outperform Nvidia's NVL144 in compute, memory, and bandwidth. These SuperPoDs can be combined to form SuperClusters, creating immense computing power. Huawei's strategy emphasizes providing robust ICT infrastructure, with a core focus on these large-scale computing systems. The company is also making its CANN (Compute Architecture for Neural Networks) platform open-source by the end of 2025 to encourage broader adoption. This initiative comes at a time when U.S. firms like Nvidia face increasing scrutiny and pressure in the Chinese market, with Chinese cloud providers actively seeking domestic alternatives. Huawei's push for self-sufficiency in AI hardware is a critical component of China's broader national strategy to reduce reliance on foreign technology, particularly in advanced semiconductor manufacturing.