>>105991533
Intel and AMD are competing fine in inference at certain pricepoints. But at the very top, yes, it's Nvidia or nothing. It's a meme that you can't use either for your image/video genning needs though. There is a reason the A770 is not below $250 USD on used markets. It's the cheapest GPU that can run AI with BF16 support with 16GB of VRAM, which means it's better than even Volta.
>>105991637
The datacenter cards are "good" if you can get good prices for them. Gaudi puts up good enough results on MLPerf but it's ASIC specific inherited from Habana and the software it uses is a dead end. Intel's Ponche Vecchio which is their GPU line that actually has a future hasn't done enough to make them compelling. Furthermore, being outdated, they have not been written them down on their balance sheet to make that happen. Geohotz wrote about this economic dilemma this year. He never worked on their GPUs because of what they wanted for their cards.
https://geohot.github.io//blog/jekyll/update/2025/03/24/tragic-intel.html