The physical appearance of Huawei's new - generation Ascend 950 AI accelerator has come to light, revealing the packaging style of its self - developed chip and self - developed High Bandwidth Memory (HBM). This chip seamlessly integrates the first - generation self - developed HBM and a new - generation AI acceleration unit, specifically designed for large - scale computing clusters. Rather than directly competing with rivals such as NVIDIA by relying solely on single - chip performance, it aims to gain an edge by boosting system scale and cluster density.

The Ascend 950 series is slated for launch in early 2026, with at least two models in the pipeline: the 950PR and the 950DT. The 950PR version incorporates 128GB of Huawei's self - developed HBM, boasting a bandwidth of around 1.6TB/s. In contrast, the 950DT version ups the ante with a capacity of 144GB and a bandwidth approaching 4TB/s. Both chips are engineered to deliver single - card performance of 1 PetaFLOPS in FP8 and 2 PetaFLOPS in FP4, making them ideal for large - scale model inference and training scenarios.

Huawei places a strong emphasis on enhancing computing power and interconnection efficiency at the cabinet and data center levels. This is achieved through high - density packaging and efficient interconnection networks, effectively compensating for any gaps in single - chip performance. From a manufacturing process perspective, industry experts anticipate that the Ascend 950 will adopt SMIC's N + 3 process (a 5nm - level node). This process has already achieved mass production without the need for EUV equipment.

In terms of design, the Ascend 950 adopts a multi - chip packaging approach. Its core is composed of two computing chip dies, which are paired with additional I/O and networking - related chip dies to form a multi - chip module (MCM). The I/O and networking chips play a crucial role in connecting the accelerator card to large - scale clusters. They enable high - bandwidth interconnection for hundreds of thousands of cards through a new - generation "Lingqu" interconnection protocol and optical interconnection technology.

The overall design approach of the Ascend 950 bears a resemblance to that of high - end GPUs like NVIDIA's Blackwell. However, Huawei puts a greater emphasis on a "scale - based victory" strategy. By leveraging dense packaging, multi - card interconnection, and super - cluster solutions, it offers an alternative solution for the data center and AI cloud computing markets. At the same time, it strengthens the autonomy and controllability of the local supply chain.