Many users who are happy with "I want a GPU to drive my display and maybe game" are pleased with the current state of AMD - and that's fine. However, that population rarely goes beyond "how many FPS am I getting?". That population regularly complains about the Nvidia closed-source/binary driver situation. A lot of that criticism (from their point of view) is fair - although I've never had issues with the binary driver and as seen in this thread I'm not alone.
However, if they were to expect even just a little bit more - ML, reliable and usable hardware video encoding/decoding, etc they'd quickly realize what we have. The closed-source Nvidia driver is remarkable capable and consistent (from a functionality perspective) supporting every single feature and function of any and every card that says "Nvidia" on it across multiple platforms. From nearly decade-old laptop GPUs to the latest and greatest H100 datacenter GPU and everything in between. This is before you even get to the dozens of libraries Nvidia supports and provides (like cuDNN, TensorRT, etc) that have the same consistent, reliable, and cross-platform functionality.
Again, I understand their perspective and I don't look down on them or their simple use case (gaming on desktop) but in 2023 that's an extremely low bar. There's also a lot of criticism of the Nvidia price point - but those of us who are making full utilization of our hardware and supported features balk at that criticism. What's also interesting is the more buggy and less supported also closed-source (to my knowledge) AMD Pro driver that's required to do anything more than game. So at this point you're still not going to please the open source community and the functionality and experience is worse in every aspect.
My 4090s will not only beat AMD top of the line on just about every gaming benchmark, in seconds (or simultaneously) I can be training an ML model, encoding video, upscaling, doing realtime noise suppression, etc. The functionality fundamentally supported by that driver is amazing. When you look at it from that standpoint an Nvidia card being (at most) 3x the price point is perfectly reasonable. I'm thrilled to spend $1000 more if it means I can actually use the card for what I need and I save many, many hours of my time. In terms of total value it's not even close.
The increase in value, utility, and time returned is astounding and (to many of us) more than justifies the driver situation and cost. I support competition and have tried several times over the years to get ROCm to perform the most basic ML tasks - it's a blackhole of wasted time, energy, and effort (starting with the AMDPro driver). AMD clearly just doesn't care - and even if they decided to throw everything they have at it Nvidia has a decade-plus head start.
I truly wish the AMD software situation was better but I've been let down too many times to be hopeful I will be making anything approaching full use of any AMD GPU anytime in the near future.
Many users who are happy with "I want a GPU to drive my display and maybe game" are pleased with the current state of AMD - and that's fine. However, that population rarely goes beyond "how many FPS am I getting?". That population regularly complains about the Nvidia closed-source/binary driver situation. A lot of that criticism (from their point of view) is fair - although I've never had issues with the binary driver and as seen in this thread I'm not alone.
However, if they were to expect even just a little bit more - ML, reliable and usable hardware video encoding/decoding, etc they'd quickly realize what we have. The closed-source Nvidia driver is remarkable capable and consistent (from a functionality perspective) supporting every single feature and function of any and every card that says "Nvidia" on it across multiple platforms. From nearly decade-old laptop GPUs to the latest and greatest H100 datacenter GPU and everything in between. This is before you even get to the dozens of libraries Nvidia supports and provides (like cuDNN, TensorRT, etc) that have the same consistent, reliable, and cross-platform functionality.
Again, I understand their perspective and I don't look down on them or their simple use case (gaming on desktop) but in 2023 that's an extremely low bar. There's also a lot of criticism of the Nvidia price point - but those of us who are making full utilization of our hardware and supported features balk at that criticism. What's also interesting is the more buggy and less supported also closed-source (to my knowledge) AMD Pro driver that's required to do anything more than game. So at this point you're still not going to please the open source community and the functionality and experience is worse in every aspect.
My 4090s will not only beat AMD top of the line on just about every gaming benchmark, in seconds (or simultaneously) I can be training an ML model, encoding video, upscaling, doing realtime noise suppression, etc. The functionality fundamentally supported by that driver is amazing. When you look at it from that standpoint an Nvidia card being (at most) 3x the price point is perfectly reasonable. I'm thrilled to spend $1000 more if it means I can actually use the card for what I need and I save many, many hours of my time. In terms of total value it's not even close.
The increase in value, utility, and time returned is astounding and (to many of us) more than justifies the driver situation and cost. I support competition and have tried several times over the years to get ROCm to perform the most basic ML tasks - it's a blackhole of wasted time, energy, and effort (starting with the AMDPro driver). AMD clearly just doesn't care - and even if they decided to throw everything they have at it Nvidia has a decade-plus head start.
I truly wish the AMD software situation was better but I've been let down too many times to be hopeful I will be making anything approaching full use of any AMD GPU anytime in the near future.