Huawei Moves Ascend 960DT Launch To Q1 2027
Huawei is answering US restrictions on advanced semiconductor manufacturing by pushing system-level scale, linking many processors through UnifiedBus, rather than relying on the performance of individual chips.
Reporting from 1 source: GIGAZINE.
At Huawei Connect in Shanghai on September 17, Huawei said it will launch its next-generation AI chip Ascend 960DT in the first quarter of 2027, earlier than the third quarter previously planned. The Ascend 960PR is set for the third quarter of 2027. Huawei also outlined the Peerium Computing Architecture and UnifiedBus connection technology, a plan to link up to 1 million processors, the Atlas 950 SuperCluster with 256,000 compute cards, the Atlas 960 system, and the Atlas 960E SuperPoD.
Huawei gave no detailed specifications or processing performance for the Ascend 960DT at the event. The chip is the first piece of a wider cluster plan: the Peerium Computing Architecture, built around the proprietary UnifiedBus connection, is meant to link processors, memory, storage, and network equipment through a common mechanism. Huawei says the Atlas 950 SuperCluster, connecting 256,000 compute cards, is already being deployed, and the Atlas 960E SuperPoD links chips by optical communication, with up to 4096 AI processors in a single system.
- Ascend 960DT: planned for the first quarter of 2027
- Ascend 960PR: planned for the third quarter of 2027
- Peerium Computing Architecture: operates many processors like a single large computer
- UnifiedBus: connects processors, memory, storage, and network equipment
- Atlas 950 SuperCluster: connects 256,000 compute cards, already being deployed
- Atlas 960: under testing, planned to use a new connection method with optical communication
- Atlas 960E SuperPoD: connects chips using optical communication, up to 4096 AI processors per system
- Ascend 970: planned for 2028
- Ascend 980: planned for 2029
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.