The Huawei Ascend 920 AI processor is slated for an official launch in the second half of 2025.

Huawei has reportedly developed the Ascend 920 AI chip as the successor to its Ascend 910C, with mass production slated for the second half of 2025. According to DigiTimes, this next‑generation processor is timed to fill the gap left by restricted access to NVIDIA’s H20 accelerators in China.

Production Timeline and Strategic Importance

• H2 2025 Mass Production: Huawei plans to ramp up manufacturing of the Ascend 920 in late 2025, aiming to meet domestic demand for high‑performance AI silicon.

• Reducing NVIDIA Dependence: Industry insiders believe the Ascend 920 could become China’s go‑to AI chip, drawing major customers like ByteDance and Tencent away from NVIDIA.

Key Specifications

• 6 nm N+3 Node by SMIC: Built on SMIC’s 6 nm process, the Ascend 920 offers modern fabrication in the face of export controls.

• 900 TFLOPS BF16 Compute: Delivers substantial AI compute power comparable to—or potentially exceeding—NVIDIA’s H20.

• 4,000 GB/s Memory Bandwidth: Utilizes HBM3 modules to sustain data‑intensive workloads.

• PCIe 5.0 & Next‑Gen Interconnects: Supports high‑throughput links for efficient cross‑node collaboration and resource scheduling.

Architectural Enhancements

• Chiplet Design: Retains the dual‑chiplet layout of the Ascend 910C but claims a 30–40% boost in training efficiency.

• Enhanced Tensor Accelerators: Optimized specifically for Transformer and Mixture‑of‑Experts (MoE) models to accelerate large, complex AI workloads.

Outlook

While real‑world performance data have yet to emerge, the combination of advanced process technology, upgraded tensor units, and robust interconnect support suggests that the Ascend 920 will not only rival NVIDIA’s H20 in raw power but also offer improved efficiency over previous Ascend generations. As U.S. export restrictions tighten, Huawei’s latest AI silicon could play a pivotal role in revitalizing China’s AI chip ecosystem