Not only this, they secured a future for Chinese GPU manufacturing like Moore Threads by substantially decreasing the target they have to reach to be competitive with the graphics mafia at Nvidia. Specialized AI chips are a lot simpler than a full GPU with full CUDA compatibility. All you need is compatibility with PyTorch which is fully open source, and a GPU that does matrix multiplication well possibly with low precision support for inference. If precision flexibility features are too hard you can make specialized chips for training and specialized chips for inference. USA really shot its own companies in the foot with this one.
For sure, it’s always easier to make specialized chips for a specific use case than general purpose ones. And once you have that nailed down, you can expand functionality from there as you go. The whole idea that US could just freeze China out of chip development has always been so obviously absurd, it’s frankly shocking to me that anybody thought it could work.
Not only this, they secured a future for Chinese GPU manufacturing like Moore Threads by substantially decreasing the target they have to reach to be competitive with the graphics mafia at Nvidia. Specialized AI chips are a lot simpler than a full GPU with full CUDA compatibility. All you need is compatibility with PyTorch which is fully open source, and a GPU that does matrix multiplication well possibly with low precision support for inference. If precision flexibility features are too hard you can make specialized chips for training and specialized chips for inference. USA really shot its own companies in the foot with this one.
For sure, it’s always easier to make specialized chips for a specific use case than general purpose ones. And once you have that nailed down, you can expand functionality from there as you go. The whole idea that US could just freeze China out of chip development has always been so obviously absurd, it’s frankly shocking to me that anybody thought it could work.