Different gpu mixed node

Posted by Force88@reddit | LocalLLaMA | View on Reddit | 10 comments

Hi, Currently my nodes have 2x 5060ti 16gb, today I saved enough to add another card to my node...But when I checked the price of 5060ti 16gb has jumped $100, make it very close to the 5070 12gb, which is just $55 more. Also the cheapest 9070xt also have a sale today, which is only $100 more than the 5060ti. What card should I buy ? Is there going to be issue when mixing gpus between vram size (5060ti + 5070), or even mixing vendor (5060ti + 9070xt) So my case is,

Reply to Post

10 Comments

[-]

Ok-Measurement-1575@reddit

Buy another 5060Ti 16GB.

[-]

Force88@reddit (OP)

Thanks, I'll try to look for better listing till the end of the sale. $100 increase hurts lol

[-]

tecneeq@reddit

The $100 might give you pain this month, but 4GB more VRAM will last you a lifetime. https://preview.redd.it/hdu2aujmcc1h1.jpeg?width=760&format=pjpg&auto=webp&s=500cb7cc8af4faca6a7afdabf2fbca828c7cbfee

[-]

tecneeq@reddit

I think VRAM is more worth than the fastness of the 5070, so i would get another 5060 Ti 16GB. That would bring you to 48GB VRAM, pretty nice for qwen 3.5 27b Q6 and full context.

[-]

wbulot@reddit

I don't know about the choice, but regarding compatibility, there is no issue. I run a mixed setup with an RTX 5060 Ti 16GB + AMD 6800 XT 16GB, using CUDA and Vulkan, running Qwen 3.6 27B Q6 on it. 300 tokens prompt, 15-20 t/s generation, no problem.

[-]

Force88@reddit (OP)

Thanks! Do you notice slower speed than normal, compared to single card ?

[-]

wbulot@reddit

If I test with a smaller model that can fit on a single GPU, I don't see any difference between running it on one GPU or two GPUs.

[-]

Formal-Exam-8767@reddit

If the choice is between 5060ti 16GB and 5070 12GB, it makes no sense to go with less VRAM since you want really see any speed benefits that 5070 has over 5060ti. To me, it only makes sense to consider 5060ti 16GB or 5070ti 16GB.

[-]

Force88@reddit (OP)

Thanks, I was shock by the price increase of 5060ti 16gb, but in the end it should still be the most stable choice huh.

[-]

chickN00dle@reddit

In my opinion, you can't go wrong with either option. But if you want to use the AMD card, I might expect a bit more tinkering to get it working the way you want it to. If that's worth the extra performance to you, then yes I'd go for it. But I don't have the knowlegde to tell you if vulkan (because mixed) would be as performant as a pure CUDA setup alone, assuming the model and all context fit in VRAM.