But it’s unclear with which ones
The Moore Threads company currently produces the most powerful video cards, if we talk about Chinese developments. According to its own tests, its MTT S4000 adapter can rival Nvidia cards in AI tasks.
MTT S4000 was introduced late last year, and is a solution initially aimed at the data center and AI. The adapter's performance in FP32 mode is 25 TFLOPS, which is slightly lower than that of the RTX 4070, and in FP16/BF16 mode, which is more important for AI, it already reaches 100 TOPS. In INT8 mode we are talking about 200 TOPS.
But this is theoretical performance. In practice, when training the MT-infini-3B language model with three billion parameters, the system based on Moore Threads accelerators showed results comparable to Nvidia products.
True, there are important nuances. Firstly, the comparison was made when training different language models, which is no longer particularly representative. Secondly, there are no details about which Nvidia accelerators were used for comparison, in what quantities, and even regarding the Chinese system there are no such details. Of course, the authors could have made the comparison as correct as possible, to the extent they could, but we cannot know about this.
However, for Moore Threads itself, perhaps, opposing Nvidia accelerators is not so important, since now the Chinese market is aimed at completely replacing foreign products with its own developments.