At the Huawei China Partner Conference 2026, Huawei officially launched the Atlas 350 accelerator card, powered by the new Ascend 950PR processor. This marks the first hardware product in the Ascend 950 series, targeting high-performance AI inference workloads.

The Atlas 350 is built for three core AI scenarios: recommendation inference, multimodal generation, and LLM inference. Seven ecosystem partners—including Kunlun, Huakun Zhenyu, and Shenzhou Kuntai—simultaneously released systems built around the new card.

Performance numbers stand out. According to Huawei, the Atlas 350 delivers 2.87 times the single-card compute of NVIDIA’s H20. It is also the only domestic inference accelerator supporting FP4 precision, filling a critical gap in low-precision inference.

Memory specs are equally aggressive. The card packs 112GB of HBM, 1.16x the capacity of H20, which boosts multimodal generation speed by 60%. Memory access granularity has been reduced from 512 bytes to 128 bytes, improving small-operator access efficiency by 4x.

Key specs include 1.56P FP4 compute, 1.4TB/s bandwidth, and 600W power consumption, positioning the Atlas 350 for dense AI inference deployments in data centers.

ICgoodFind: Huawei’s Atlas 350 sets a new bar for domestic AI inference. With 2.87x H20 performance, exclusive FP4 support, and a growing ecosystem, it’s a serious challenger to overseas accelerators.