Huawei's next-generation AI accelerator, the Ascend 950DT, is set to launch on Huawei Cloud in August 2025. The chip brings a significant memory bandwidth upgrade, more than doubling performance over its predecessor.
Self-developed HBM delivers 144GB capacity
The Ascend 950DT is a high-performance AI chip designed for cloud data centers. It succeeds the earlier Ascend 950PR and targets large-scale AI training and inference workloads.
Specifications
- Memory Capacity: 144GB
- Memory Bandwidth: 4TB/s
- Interconnect Bandwidth: 2TB/s
- Supported Data Formats: FP8, MXFP8, MXFP4, HiF8
The chip features self-developed HBM memory with 144GB capacity and 4TB/s bandwidth. Interconnect bandwidth reaches 2TB/s. It supports FP8, MXFP8, MXFP4, and HiF8 data formats.
DeepSeek V4.2 may deploy on the chip
According to Huawei Vice President Chen Lin, memory bandwidth increased from 1.6TB/s to 4TB/s, more than doubling performance. The chip will be available on Huawei Cloud starting in August 2025.
DeepSeek V4.2 may deploy on the Ascend 950DT in August 2025, according to sources. This deployment is not confirmed by DeepSeek.
The Ascend 950DT is confirmed to launch in August 2025 on Huawei Cloud with the specified memory and interconnect specs. Pricing has not been announced.
Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.