Huawei's new AI chips are catching up to NVIDIA fast. The Chinese tech giant's CloudMatrix 384 beats NVIDIA's GB200 NVL72 in benchmarks, even though it needs more power. Huawei currently uses Ascend 910C chips that deliver strong results but aren't as efficient. They plan to fix this with their upcoming Ascend 920 family built on SMIC's 6 nm process.
The fresh Ascend 920C will push past 900 TeraFLOPS for BF16 calculations, according to DigiTimes. It upgrades to HBM3 memory offering 4,000 GB/s bandwidth, much faster than the older model's HBM2E setup. The current Ascend 910C maxes out at 780 TeraFLOPS with chip-to-chip connections running at 400 GB/s.
Huawei will keep using chiplets but improve the tensor engines for Transformer models. Internal tests show the 920C might perform 30-40 percent better than its predecessor. This narrows the gap with competitors. The new chips will support PCIe 5.0 with better interconnect protocols to reduce delays between nodes in large systems.
These chips should enter mass production during the second half of 2025, which could mean they'll be available just months from today.
The fresh Ascend 920C will push past 900 TeraFLOPS for BF16 calculations, according to DigiTimes. It upgrades to HBM3 memory offering 4,000 GB/s bandwidth, much faster than the older model's HBM2E setup. The current Ascend 910C maxes out at 780 TeraFLOPS with chip-to-chip connections running at 400 GB/s.
Huawei will keep using chiplets but improve the tensor engines for Transformer models. Internal tests show the 920C might perform 30-40 percent better than its predecessor. This narrows the gap with competitors. The new chips will support PCIe 5.0 with better interconnect protocols to reduce delays between nodes in large systems.
These chips should enter mass production during the second half of 2025, which could mean they'll be available just months from today.
