Huawei unveils extendable Ascend chip strategy - three-year initiative sets bold target for on-site HBM delivering up to 1.6 TB/s data transfer rate
In an attempt to rival Nvidia's market-leading AI compute clusters, Huawei has announced its first official long-range Ascend chip strategy. The Chinese tech giant plans to release four new Ascend chips over the next three years, with the Ascend 950PR, an ai chip, scheduled to ship in Q1 next year.
The Ascend 950PR chip features 128GB of in-house High-Bandwidth Memory (HBM), delivering up to 1.6 TB/s of bandwidth. Huawei is also planning to deploy 'supernodes' that will house thousands of Ascend chips, aiming to break the bottlenecks that have been a challenge so far.
One of the key components of the supernodes will be the Atlas 950 system, part of the supernodes, which is expected to debut in Q4 this year. Atlas 950 is positioned as a next-generation ai compute cluster, with up to 15,488 Ascend accelerators in a single system.
However, Huawei faces significant challenges in matching Nvidia's performance in training, efficiency, and model throughput. Nvidia's advantage lies in its NVLink and optimized software stack. To succeed, the Huawei platform must match Nvidia in these critical areas.
To address this, Huawei has not disclosed details about the manufacturing process, packaging, or foundry for its in-house HBM or the Ascend chips. The factory likely involved in the production of Huawei's Ascend chips for the next three years is SMIC (Semiconductor Manufacturing International Corporation), as Huawei plans to rely on domestic fabrication after the current supply of chips produced abroad by TSMC runs out within the next nine months.
The Chinese government is demanding more domestic silicon production from Huawei, and the company is under pressure to produce more domestic silicon. This pressure is further intensified by U.S. sanctions, which bar Huawei from accessing TSMC's advanced nodes and CoWoS packaging lines.
Huawei's current offerings do not match Nvidia in training, efficiency, and model throughput. The Atlas 950 and 960 systems are designed to rival Nvidia's GB200 NVL72 configurations in deployment scale, but plans alone are not sufficient to break the performance gap.
In conclusion, Huawei's Ascend chip strategy is a significant step towards challenging Nvidia's dominance in the ai compute cluster market. However, the success of this strategy hinges on Huawei's ability to match Nvidia's performance in training, efficiency, and model throughput, as well as its ability to overcome the challenges posed by U.S. sanctions and domestic silicon production demands.
Read also:
- AI-Generated Humor Spreads on Gemini Nano Banana: Light-hearted Modifications Spark Concerns over User Privacy
- China is, unlike the United States, embracing technological progress rather than attempting to restrict it.
- Social media spat between Elon Musk and Sam Altman features their confrontation; discord revolving around business rivalry in relation to Apple
- Video game franchise F1® debuts its 25th installment today in early access, boasting state-of-the-art graphics technology with Path Tracing, next-gen AI-powered DLSS 4 with Multi Frame Generation, and DLSS Ray Reconstruction.
