It's really technical that is fair! If you have any questions I would be happy to try to explain!
Exactly - standardize the datatypes so that everyone can use the same number format and build hardware the supports it.
CUDA/Nvidia don't loose at all! In my opinion - they gain as much as...
This is for both training and inference! You end up with a small gap using MX4 compared to FP32, but that might be acceptable for your use case. MX6 is on par with FP32 training.
Not exactly - the element data type for MXFP4 for example is 2 exponent bits and 1 mantissa bit. These are grouped...