SemiAnalysis measured that Ascend 950PR matrix calculation speed was approximately H100 of 43%

AT A GLANCE

A report citing testing said that Ascend 950PR had an actual speed of approximately H100 of 43%，DeepSeek is improving efficiency through software adaptation.

Article

Huawei Ascend 950PR Matrix calculations only H100 of the speed,DeepSeek is squeezing performance out of the software

According to a Beating AI news brief, last year,DeepSeek tried using Huawei Ascend chips to train R2, but even with Huawei engineers providing on-site assistance, it still failed to run successfully. Problems existed with chip stability, cross-chip communication, and software support. This year, the two sides continued to deepen their cooperation.September At the end of the year,DeepSeek publicly released six sets of Ascend adaptation code;October there were also reports that Huawei was adjusting its next-generation chip design according to the model requirements of DeepSeek.

The semiconductor research firm SemiAnalysis recently published measured results for Ascend 950PR. Researchers selected 1055 groups of matrix calculations commonly used by large models and found that Huawei's own low-level programs had already enabled 950PR to achieve 86%of its attainable computing power, higher than Nvidia H100's 75.5%. However, the actual computation speed was still only H100 of 43%, and compared with the newer B300 it was only 21%. These results did not yet use the operators newly ported by DeepSeek.DeepSeek is reducing data transfers and waiting time through software. For example, when running attention calculations, two data sets that originally had to be read separately can be retrieved at once after FlashMLA adjusted the storage method. Another tool,TileLang, enables the chip to calculate while preloading the next batch of data. SemiAnalysis measured that automatic scheduling reduced the time required for some small computational tasks by up to approximately 40%。DeepSeek also ported DeepEP, which handles communication among MoE experts, but communication and computation still compete for memory bandwidth, and Huawei plans to continue improving the firmware.

The 950PR tested this time mainly targets inference input processing, while DeepSeek is preparing to use 950DT for large-scale pretraining as early as the end of the year. The two sides had previously conducted testing at the -chip scale, but enabling tens of thousands of chips to continuously complete one round of large-model training will require resolving more complex coordination and stability issues.

Original link https://m.theblockbeats.info/flash/373937

Key points

According to measured results published by SemiAnalysis, in 1055 groups of matrix calculations commonly used by large models, Ascend 950PR had an actual computation speed of H100 of 43%。
Huawei's low-level programs enabled Ascend 950PR to achieve 86%of its attainable computing power, higher than H100's 75.5%。
FlashMLA can adjust the data storage method,TileLang and can preload the next batch of data during computation.
The report cited measurements saying that automatic scheduling reduced the time required for some small computational tasks by up to approximately 40%。
DeepSeek has ported MoE DeepEP, which handles communication among experts, but communication and computation still compete for memory bandwidth.
Ascend 950PR mainly targets inference input processing, while Ascend 950DT is planned for large-scale pretraining as early as the end of the year.
The two sides had previously conducted testing at the -chip scale, but completing one round of large-model training continuously with tens of thousands of chips still faces coordination and stability issues.

AI-assisted interpretation

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

In simple terms, 950PR's low-level programs can already utilize a relatively high proportion of its theoretical computing power, but its speed on common large-model computations in actual operation is still lower than H100。DeepSeek has attempted to reduce waiting by modifying data reads, preloading data, and optimizing expert communication, but stable completion of large-scale training has not yet been demonstrated.

Why it matters to readers

This shows that a chip's actual performance depends not only on its theoretical computing power but also on software, communication, and system coordination; software optimization may improve efficiency, but the report still noted that large-scale training faces challenges.

When evaluating chip performance, one cannot look only at theoretical computing power or individual tests; software adaptation, cross-chip communication, memory bandwidth, and long-term stability must also be considered.

Risks and unknowns

The original article did not provide SemiAnalysis's complete test report, hardware configuration, or testing methodology.
The report was not independently confirmed by Huawei or DeepSeek.
The headline says matrix calculations run at H100 of the speed, while the measured figure given in the body is 43%, so the two descriptions are not completely identical.
Ascend 950DT The specific timetable for using it in large-scale pretraining has not been independently confirmed.

SemiAnalysis measured that Ascend 950PR matrix calculation speed was approximately H100 of 43%

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center