mcdow

In my experience even the best software projects have a few skeletons in their closet, blemishes on an otherwise well-built project.

At the end of the day, we all have to build things that simply work and provide business value. Striving for perfect code is not the goal. But it does make me wonder: does perfect software even exist? If not, what's the gold standard?

Nvidia has an effective monopoly on the GPUs used for AI/ML. AMD GPUs only seem to be useful for gaming. The common sentiment is that the difference between the two GPUs is only the software; the hardware is comparable. Specifically, many people say there are issues with the drivers crashing.

I'd like to hear from those familiar with writing and using GPU software.

1. Is it really only a software issue?

2. What are all of the issues? Crashing? Interface compatibility with CUDA?

3. What is the scale of work that would need to be done to get AMD GPU software to the level of Nvidia GPU software? Does this require a complete revolution of the culture at AMD, or is this a task that could be accomplished by a highly motivated and competent vanguard?

I ask because the common sentiment seems silly to me. How hard is it really to write some good drivers? I'd like to challenge my held beliefs.