Good to see this analysis. Now that both nr. 1 supercomputers in the US and EU are basically AMD machines. Nvidia needs to step up their game if they want to stay relevant in the HPC space

There was a hackernews post yesterday about running CUDA unmodified on AMD GPUs. In this article, we show that this is generally a bad idea. If you want to get good performance on AMD you have to specifically optimize the code towards AMD GPUs. Perhaps surprisingly, the reverse is not true, taking code that has been optimized for AMD to Nvidia is possible with little performance penalty.

This paper on bringing auto-tuning to AMD's HIP shows some intriguing results! The performance boost on AMD GPUs is 10x compared to Nvidia's 2x. It highlights the unique tuning challenges and the need for tailored solutions. Anyone else excited to see more competition in the GPU space pushing for optimized performance across different architectures? Also, any thoughts on how this might impact the broader software development landscape?