The Information reported DeepSeek is actively helping shape the design of Huawei's future AI chips; DeepSeek also released open-source tools for Ascend and helped optimize a 128-chip Ascend 950 supernode.
Following·Day 180·3 sources
DeepSeek and Huawei Deepen Chip and Software Co-Design Collaboration
The gist
The collaboration represents a concrete step toward China building a full domestic AI stack covering models, software, and silicon without dependence on Nvidia or CUDA. Hardware-software co-design between DeepSeek and Huawei is producing chips and software optimized together from the start, reducing the performance gap that typically emerges at launch.
The full picture
DeepSeek is actively helping shape the design of Huawei's future Ascend AI chips, according to The Information. DeepSeek also released open-source programming tools for the Ascend chip line and helped optimize a supernode built around 128 Ascend 950 chips. The collaboration extends to software: Huawei's Ascend 950PR and 950DT chips achieved what Huawei called 'day zero' adaptation to DeepSeek's V4 model at launch, with Huawei's CANN software prepared ahead of release with fused operators and optimizations for the chip's dual-core design. DeepSeek V4 is a 1.6 trillion parameter Mixture of Experts model with 1 million token context support. DeepSeek acknowledged V4 will face throughput constraints until Ascend 950PR supernodes ship at scale in the second half of the year. Huawei's CANN platform serves as its alternative to Nvidia's CUDA.
How it developed
Huawei announced 'day zero' adaptation of Ascend 950PR and 950DT to DeepSeek V4 at model launch, citing close pre-launch hardware-software co-design.
Sources
Milk Road AI Twitter
huaweicentral.com
Related
- Split fromDHS's DeepSeek federal exclusion order
The daily email
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free