Huawei Develops 4-Chiplet Ascend 910D Accelerator — Competitor to Rubin Ultra and an alternative to CoWoS from TSMC

Huawei may soon introduce its first quad-chip AI accelerator, the Ascend 910D, aimed at competing with NVIDIA Rubin Ultra. According to a new patent application, the company is developing own chip packaging technology, which could compete with TSMC's CoWoS and Intel's Foveros. This could allow Huawei to bypass US sanctions and get closer to the performance of top AI-GPU without access to modern technological processes.

The patent describes bridge architecture for inter-chip communication, similar to EMIB or CoWoS-L, with the ability to integrate HBM memory via an interposer. Assuming that the Ascend 910D is indeed based on four Ascend 910B dies (665 mm² each), the total chip area could exceed 4000 mm²., which corresponds to 5 EUV reticles. This makes the project technically extremely ambitious, especially considering TSMC's plans to launch similar packages only in 2026.

Preliminary data suggests that the 910D will be able to outperform the NVIDIA H100 in single-device performance., and the key way to achieve this will be aggressive packaging scaling and HBM memory usage. At the same time, Huawei remains on old technological processes due to sanctions, but uses chiplet architecture as a way to compensate for lithography lagThe Ascend 910D is likely to form the basis of a new CloudMatrix cluster that is more powerful than the GB200 but consumes 4x more power.