Comment by tgtweak

I think AMDs shady marketing where they claimed 1.4x over H100 is enough to just steer clear of the hype and wait for results.

Summary: they cherrypicked legacy nvidia sdk's and used llama batch sizes that are not used often in production...

https://twitter.com/karlfreund/status/1735078641631998271

https://developer.nvidia.com/blog/achieving-top-inference-pe...