Chinese tech giant Huawei has unveiled its once-secretive AI chip development program, an announcement that could reshape expectations across the fast-moving, high-stakes global industry as the firm poses as direct rival to Nvidia.
On September 18, Huawei chairman Eric Xu announced a three-year roadmap for the company’s Ascend series of processors at the Huawei Connect 2025 conference in Shanghai.
Coming soon after the Chinese government decided to block Chinese companies from using Nvidia processors, Huawei’s announcement will heighten competition between Chinese and American AI efforts while signaling a significant reduction in China’s dependence on American technology.
About The Ascend Chips
Huawei launched the Ascend 310 chip in 2018 and the Ascend 910 in 2019. By 2025, the Ascend 910C has become widely recognised in the industry, particularly with the scaled deployment of the Atlas 900 A3 SuperPoD.
The company now plans to continue the evolution of the Ascend series to strengthen the foundation of AI computing power, in China and worldwide. Over the next three years, it plans to release three new chip series: the Ascend 950, Ascend 960, and Ascend 970.
Ascend 950 Series
The Ascend 950 series introduces two models:
-
Ascend 950PR – optimized for prefill and recommendation.
-
Ascend 950DT – optimized for decode and training.
Both are built on the same Ascend 950 Die, offering fundamental improvements over their predecessors:
-
Expanded data format support: FP8, MXFP8, MXFP4, and Huawei’s proprietary HiF8 format (precision close to FP16, efficiency similar to FP8). Performance reaches 1 PFLOPS in FP8/HiF8 and 2 PFLOPS in MXFP4.
-
Stronger vector processing: More compute allocated, a combined SIMD–SIMT design, and finer memory access granularity (128 bytes vs. 512).
-
Higher interconnect bandwidth: 2 TB/s — 2.5x that of the Ascend 910C.
-
Custom memory options: Two proprietary HBMs — HiBL 1.0 (cost-effective, for prefill/recommendation) and HiZQ 2.0 (high-performance, for decode/training).
-
The Ascend 950PR will launch in Q1 2026, available in card and SuperPoD server form factors.
-
The Ascend 950DT will launch in Q4 2026, with 144 GB of HiZQ 2.0 HBM and 4 TB/s memory bandwidth.
Ascend 960
Launching in Q4 2027, the Ascend 960 will double the compute, memory, and interconnect capacity of the Ascend 950 series. It will also introduce HiF4, a proprietary 4-bit format delivering superior precision over existing FP4. Training and inference performance will rise to a new level, with 2 PFLOPS in FP8 and 4 PFLOPS in FP4.
Ascend 970
Planned for Q4 2028, the Ascend 970 will double compute power again, delivering 4 PFLOPS in FP8 and 8 PFLOPS in FP4. It will also double interconnect bandwidth (4 TB/s) and increase memory access by at least 1.5x.
