[News] Huawei Unveils Ascend 950 with In-House HBM in 2026, Touts SuperPoD to Rival NVIDIA

According to Mydrivers, at today’s Huawei Connect 2025 conference, the company unveiled a multi-year roadmap for its Ascend AI chips, with steady releases planned through 2028. The rollout will start with the Ascend 950PR in Q1 2026 and the 950DT in Q4 2026, followed by the Ascend 960 in Q4 2027 and the Ascend 970 in Q4 2028.

The report also highlights that a key feature of the 950 chip is Huawei’s self-developed HBM, along with support for low-precision data formats, stronger vector performance, and a 2.5-fold boost in interconnect bandwidth.

As indicated by the report, Huawei’s Ascend 910C entered mass production in the first quarter of this year, integrating two Ascend 910B chips with advanced packaging. It delivers about 800 TFLOPS of FP16 performance and 3.2 TB/s of memory bandwidth—roughly 80% of NVIDIA’s H100. Alibaba, Baidu, and Tencent are reportedly among its first customers.

Huawei’s SuperPoD Ambitions: Challenging NVIDIA in Computing Power

Alongside these chip-level advances, Huawei is pushing system-level performance with the Atlas 950 SuperPoD, unveiled at the conference. Equipped with 8,192 Ascend NPUs, it was described in the company’s press release as the world’s most powerful SuperPoD to date. The system is scheduled to debut in Q4 this year, with the company claiming it could even outperform NVIDIA’s NVL576 system planned for 2027, as Guancha points out.

Looking ahead, the next-generation Atlas 960 SuperPoD, with 15,488 Ascend NPUs, is scheduled for release in Q4 2027. The company also unveiled the world’s first general-purpose computing SuperPoD, the TaiShan 950 SuperPoD, slated to launch in Q1 2026, Mydrivers adds.

Scaling with SuperClusters: Huawei Bets on Local Chips

Beyond the SuperPoD, Huawei introduced the Atlas 950 SuperCluster, featuring more than 500,000 Ascend NPUs, and the Atlas 960 SuperCluster, with over one million NPUs. Both are large-scale computing systems formed by linking multiple Huawei SuperPoDs, as its press release highlights.

As noted by South China Morning Post, citing Huawei deputy chairman and rotating chairman Eric Xu, the company is developing a “superPoD + cluster” computing solution based entirely on China’s chip manufacturing processes to meet rising compute demands. While a single Huawei chip still trails NVIDIA, Xu stressed that sustained investment in interconnect technology enables its systems to handle very high workloads, as highlighted by Guancha.

Read more

• [News] Huawei’s Bold Return: Mate XTs Debuts Kirin 9020 After Four Years of Chip Silence
• [News] China Reportedly to Triple AI Chip Output Next Year With Fabs Serving Huawei, Cutting NVIDIA Reliance

(Photo credit: Huawei)