Interesting AI GPUs for Beginners: Intel Dual Arc Pro B60 and AMD AI Pro R9700
[17:14 Mon,26.May 2025 by Rudi Schmidts]
Nvidia is known to have the undisputed top position in current AI hardware. Hardly any AI data center is currently not based on Nvidia hardware, which gives the US company hefty profit margins and exorbitant billions in profits. But if you look closely, you will also notice that the technology behind the most expensive data center GPUs is based on the same hardware DNA as current gamer graphics cards. The differences between data center-expensive and gaming-cheap are mainly in the memory expansion and the networking options (i.e. fast interconnect solutions via NVLink).
For smaller AI experiments or local LLMs, however, a GPU is often sufficient, which is why you can in principle use any Nvidia graphics card, provided it offers enough memory for your own requirements. But here too, Nvidia charges its customers dearly as soon as the usable memory exceeds the typical amounts for gamers. At Nvidia, this is 32 GB (which also cost over 2000 euros here with the RTX 5090).
AMD and Intel are now jumping into this breach by offering GPUs with more memory significantly cheaper than Nvidia - in the hope of inspiring cost-conscious users to jump out of the CUDA ecosystem.
In fact, Intel has succeeded in creating the biggest surprise, as nobody had expected this player so far. With the MaxSun Dual Turbo will soon be launching a dual GPU card with double the memory (2×24 GB, 2x 456 GB/s) a dual GPU card is coming onto the market that fits into a 16x PCIe slot, but is seen and treated by the system as two 8x PCIe 4.0 cards.
The crucial thing is that many AI models can be distributed well across multiple cards during execution, as long as they fit into the available GPU memory overall. For many models, this should be the cheapest 48 GB card in the near future, which, however, should also be associated with some fiddling in order to actually get the desired models running smoothly.
Things are certainly a little better with AMD, whose ROCm interface already enjoys a much wider distribution. Although it has not yet reached CUDA level, many models, especially from the open source area, can be executed with just a few mouse clicks under a ROCm-compatible AMD GPU. And the supported card types are actually increasing more and more.
With the AMD AI Pro R9700, AMD is sending a 32 GB RDNA 4 graphics card into the race, which is otherwise practically the same as a Radeon RX 9070 (XT) with 32 GB. Since this is much slower than an RTX 5090 with the same memory configuration, this card should hardly be offered for more than 1000 euros on the market in order to find relevant attention in AI circles. Video editors without an AI focus, on the other hand, are more likely to use Nvidia cards with 16 GB of memory, as memory there is also much faster thanks to GDDR7. An RTX 5070 Ti for 800 euros should also clearly beat an AMD AI Pro R9700 in all areas in which 16 GB of GPU memory is sufficient.
;)
;)