FP4 quants on nvidia are different

Summary

The transcript discusses performance gains in quantizing from FB8 to FB4 on Nvidia's hardware, highlighting unexpected speed improvements and efficiency. Nvidia reports a 5x speed increase, which is more significant than the expected 2x due to complementary formats and optimizations in Hopper and Blackwell architectures. The most striking revelation is the dramatic 25-50x reduction in power consumption, suggesting major advancements in computational efficiency for AI and machine learning technologies.