I don’t get it. How is this an alternative to NVidia, it just sounds like a way to connect more machines reliably (but those machines still have NVidia hardware)?
They use their own Ascend NPUs. They’re weaker individually than Nvidia’s chips, so they emphasize connecting many more of them together to compensate. It’s not about replacing individual pieces; the product is basically the complete system (think rack-scale AI supercomputer). It’s an alternative to Nvidia’s system level products like NVL144, not necessarily in terms of performance but also because it’s produced 100% domestically in China.
They’re not using Nvidia hardware, they developed their own Ascend chips and they figured out a way to connect them which drops latency meaning that even though individual chips are slower, the cluster as a whole is competitive with Nvidia.
