Huawei Advances the Ascend 960DT to Early 2027

Key Takeaways

1Huawei is advancing the launch of the AI chip Ascend 960DT to the first quarter of 2027

2The company is introducing the Peerium Computing Architecture to massively connect its accelerators

3An analyst points out a gap between past scaling ambitions and the size of the announced system

4The project is part of the technological rivalry between the United States and China, despite American restrictions

Why it matters : This acceleration of the timeline and the ambition to aggregate thousands of chips illustrate Huawei's desire to compete with Nvidia and strengthen China's technological autonomy in the face of international pressures.

Huawei promises the Ascend 960DT by the first quarter of 2027 and presents an architecture designed to connect tens of thousands of accelerators. However, an analyst points out a contrast between previous goals and the size of the system mentioned this week, as the technological competition between the United States and China intensifies.

Aggregation Targets Questioned Compared to Previous Announcements

Chinese analyst Rui Ma reminds us that Huawei had previously introduced an Atlas 960 SuperPoD that was supposed to scale up to 15,488 Ascend 960 chips. She highlights that, in comparison, the most recent announcement refers to a system limited to 4,096 chips. Furthermore, Huawei claims that an Atlas 950 SuperCluster can connect up to 256,000 accelerator cards.

Chip Timeline Advanced to Early 2027 and Announcement at Huawei Connect

Huawei has advanced the launch of its AI chip Ascend 960DT to the first quarter of 2027, whereas it was initially scheduled for the third quarter of the same year. A spokesperson clarifies that the 960DT should be ready by this deadline and asserts that the Ascend 960 chips are being launched ahead of schedule, with a doubling of performance. David Wang, rotating and interim president of Huawei, announced this new timeline at the Huawei Connect conference. The stated goal is to compete with Nvidia.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Proprietary Architecture to Assemble a Very Large Number of Accelerators

Huawei aims to aggregate hundreds of thousands, if not millions, of AI chips to create a vast computing ensemble. This approach, dubbed Peerium Computing Architecture, relies on UnifiedBus technology developed by Huawei to connect processors, memory, storage, and network hardware. Eric Xu, rotating president of Huawei, indicates that this architecture aims to build larger AI computing systems suitable for both training and inference. The Atlas 950 SuperPoD and SuperCluster systems are the first to utilize this architecture.

Context of Technological Rivalry and American Constraints

Huawei continues its ambitions in chip development despite American restrictions on China's access to advanced semiconductor technologies. Rui Ma believes that these measures are unlikely to stop China in its pursuit of a domestic industry, given the stakes of self-sufficiency. Donald Trump rejected requests from AI leaders to slow down progress on security grounds, stating that the United States must maintain its lead over China. For his part, Eric Xu asserts that China must intensify its AI development to catch up with the United States, and that Chinese companies need to advance first before they can fully understand and manage the risks associated with more advanced AI systems. This sequence unfolds as a meeting between Trump and Xi Jinping is scheduled in Washington on September 24.