So, what about Intel B65 GPUs, are they any good? Has anyone any experience with them?
Seems to be a cut down version of the B70 that still keeps all 32G of VRAM at somewhat lower price point.
I’m thinking of buying one to put in my server for some homelab grade LLM and VM passtrough workloads. Does this make any sense or should I just go for the full B70?
if you do not need sr-iov, go with a used cuda card. if you cannot buy used, go with r9700
the b50 i have bought is ok because it can run a llm and a guest but tooling was bad and there is no reason to get a b65 - except maybe price
I would just get the B70. The B65 can be like half of a B70 in some benchmarks. But can be a lot worse too.
Not a bad option. The cards are fine to run MoE models where an expert is smaller than 6b or so. The cards are bad at running Dense models like Qwen 3.x 27b. The higher the active parameter count, the slower it will be. Plenty good ones to pick around the pareto line:
The further left, the faster it should run on your hardware. Don’t make the mistake of dismissing models based on one or two Intelligence Index points, this is more a matter of “regions“. Also, don’t be afraid to offload a few layers to CPU with MoEs when the 32GB just isn’t enough.
So in that sense the card is great, you get models to run for a low price and you have options in terms of what models to run. I run Ling 3.0 Flash on a Strix Halo, it has more VRAM, but far less memory bandwidth, so i have to look at the same type of models. Ling 3.0 Flash is not for you, but G9v3 and Qwen 3.6 35B are. As well as maybe Nemotron 3.5 Lightning if you don’t code.
Well, i bought a B65 and will be finding out in practice (I was about to get a R9700 but then the price jumped significantly and I went for the B65 in anger).
SR_IOV functionality is also interesting for me.
Thanks for the replies and info so far.
