As long as TSMC fab capacity is the bottleneck TFLOPS/mm² with different floats. probably the most important performance factor (adjusted for yield if we want

As long as TSMC fab capacity is the bottleneck TFLOPS/mm² with different floats. probably the most important performance factor (adjusted for yield if we want to be exact). TFLOPS/MTr works also. Beating H100 with 1.4x - 1.7x when Nvidia is introducing Blackwell B200 at the same time with claimed 4.0x performance does not sound like winning microarchitecture, unless Gaudi has much smaller die area.