Test: DeepSeek matches Nvidia on code, but not on speed

DeepSeek has released two new software libraries called DeepGEMM-Ascend and DeepEP-Ascend. They work as a kind of translator: AI developers can take code they have written for Nvidia's chips and run it directly on Huawei's Ascend chips without rewriting it.

That is a big advantage for Chinese tech companies, which because of US export rules have difficulty getting access to Nvidia's newest chips. With DeepSeek's software, they can switch to Chinese hardware without months of reprogramming.

Speed still lags

But the software doesn't solve every problem. When you measure how quickly the chips can exchange data with each other (the so-called dispatch bandwidth, meaning how much information can be moved around per second), it looks fine in smaller setups: Here, 373-375 gigabytes per second is measured.

The problem is that the very largest AI models run on far more chips at the same time. In that type of large-scale setup, speed drops to 313-320 gigabytes per second, a decline of around 15 percent. In other words: The chips perform best in tests that don't resemble the reality in which the largest AI models actually run.

There is also an important detail about how the tests were carried out. DeepSeek tested the software on a prototype set of hardware that the company was given access to and that is manually configured. The commercial version of the chip, Atlas 850E, is not expected to be ready until mid-October 2026. This means other companies cannot yet verify the results under normal conditions.

The conclusion, then, is that DeepSeek has solved a real compatibility problem, so Chinese companies can more easily use domestic chips instead of American ones. But claims that the performance can match Nvidia only hold under special test conditions that are not yet available commercially.