Frankly, a few American companies buying up and blocking the entire global supply of memory and chips is going to mean the Chinese fabs only have to compete on price and quantity, not quality, to dominate much of the world’s chip sales.
A lot of the world will accept cheap Chinese hardware with large quantities of memory over the anemic overpriced offerings Nvidia is putting to market for 10x as much.
ETA: A Fenghua 3 with 112GB HBM is rumoured to be about $1-3k for H100 like performance, which is 1/10th the price.
What info have you heard about Fenghua 3? I’d last read that it’s not strictly an AI accelerator but can actually do graphics tasks, which is neat. Would make it more of a competitor to a professional workstation card like an RTX PRO 6000.
I’m most curious about their CUDA compatibility claim. I would expect that to cause a pretty significant performance hit since when writing high-performance CUDA kernels, you generally need to specialize the kernel to the individual GPU (an H100 kernel will look quite different compared to a 4090 kernel, for example). But if in spite of that it can achieve H100 performance, that’d be cool.
This was only a matter of time.
Frankly, a few American companies buying up and blocking the entire global supply of memory and chips is going to mean the Chinese fabs only have to compete on price and quantity, not quality, to dominate much of the world’s chip sales.
A lot of the world will accept cheap Chinese hardware with large quantities of memory over the anemic overpriced offerings Nvidia is putting to market for 10x as much.
ETA: A Fenghua 3 with 112GB HBM is rumoured to be about $1-3k for H100 like performance, which is 1/10th the price.
What info have you heard about Fenghua 3? I’d last read that it’s not strictly an AI accelerator but can actually do graphics tasks, which is neat. Would make it more of a competitor to a professional workstation card like an RTX PRO 6000.
I’m most curious about their CUDA compatibility claim. I would expect that to cause a pretty significant performance hit since when writing high-performance CUDA kernels, you generally need to specialize the kernel to the individual GPU (an H100 kernel will look quite different compared to a 4090 kernel, for example). But if in spite of that it can achieve H100 performance, that’d be cool.