Report Says DeepSeek and Huawei Are Incorporating Training Requirements Into Next-Generation Ascend Chip Design

AT A GLANCE

The report says DeepSeek and Huawei are collaborating on next-generation Ascend chip design and plan to initiate pre-training with 950DT as early as year-end.

Article

DeepSeek and Huawei Join Forces to Challenge Nvidia AI Computing-power hegemony: participating in next-generation chip design, racing to begin pre-training on domestic chips by year-end

Beating News AI breaking news: Last year, DeepSeek attempted to train using Huawei Ascend chips R2, but ultimately failed. The Financial Times reported at the time that there were issues with chip stability, inter-card communication, and software support. Even with Huawei engineers providing on-site assistance, DeepSeek ultimately switched back to Nvidia chips to complete training. 1 More than a year later, the two sides' collaboration has further deepened into hardware design. The Information now reports that Huawei engineers have been stationed at DeepSeek's offices for extended periods, directly incorporating practical model-training requirements into the design of next-generation Ascend chips, including how to schedule computing tasks and transfer data between chips.

This year, DeepSeek September recently open-sourced the Ascend version of TileLang, as well as underlying tools for matrix computing, MoE expert communication, and more, reducing the development work required to migrate models to domestic chips. But getting the software running is still a long way from full pre-training. Tens of thousands of chips must work continuously in coordination for weeks; a failure in any one chip, or a delay in data transmission, could slow the entire cluster. The two sides are now addressing these issues from the chip-design stage.

Huawei plans to prioritize most of the supply of Ascend 2027 expected to be ready in the first quarter of 960DT for DeepSeek, rather than distributing it among other model companies. DeepSeek, in turn, is preparing to begin pre-training with 950DT as early as the end of this year. Relevant testing has currently used more than 1000 chips; full pre-training may require tens of thousands of chips, and some cluster-management software has yet to be completed. Meanwhile, DeepSeek is still seeking more Nvidia chips.

Original link https://m.theblockbeats.info/flash/372713

Key points

Last year, DeepSeek attempted to train using Huawei Ascend chips R2 but failed, ultimately switching back to Nvidia chips to complete training.
The Financial Times reported that there were issues at the time with chip stability, inter-card communication, and software support.
The Information reported that Huawei engineers have been stationed at DeepSeek's offices for extended periods, incorporating model-training requirements into next-generation Ascend chip design.
This year, DeepSeek September open-sourced the Ascend version of TileLang, as well as underlying tools for matrix computing, MoE expert communication, and more.
The report says Huawei plans to prioritize DeepSeek for most of the supply of Ascend 2027 expected to be ready in the first quarter of 960DT.
DeepSeek is preparing to begin pre-training using 950DT as early as the end of this year; relevant testing has used more than 1000 chips, while full pre-training may require tens of thousands of chips.
Some cluster-management software has yet to be completed, and DeepSeek is also still seeking more Nvidia chips.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

The two sides are attempting to make domestic Ascend chips better suited to training large models: DeepSeek provides practical training requirements, and Huawei incorporates those requirements into chip design. Open-source tools can reduce model-migration work, but large-scale pre-training also requires large numbers of chips to cooperate reliably over extended periods, and some supporting software has yet to be completed.

Why it matters to readers

If the reported design collaboration, chip supply, and pre-training plans progress smoothly, Ascend chips will be tested in larger-scale training scenarios; however, DeepSeek is still seeking Nvidia chips, so it cannot currently be confirmed that replacement has been completed.

Beginners can follow whether the two sides formally confirm this, as well as 950DT the scale and stability of testing, progress on cluster software, and 960DT actual delivery status; plans should not be treated directly as results already achieved.

Risks and unknowns

Whether Huawei and DeepSeek have formally confirmed joint chip design, 950DT pre-training, and 960DT priority supply is not specified in the material.
950DT Whether it can support full pre-training has not yet been confirmed.
Ascend 960DT Whether it can be ready and actually delivered on schedule has not been confirmed.
More than 1000 chips' test results, stability, and specific progress have not been disclosed.

Report Says DeepSeek and Huawei Are Incorporating Training Requirements Into Next-Generation Ascend Chip Design

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center