I have a PC with a RX 9700 XT (16GB) that I mostly use for games, but I have an opportunity to replace it with an AI PRO R9700 (32GB) for a reasonable mount of money, with the idea of using it for casual agentic work in personal projects.
My main concern is the tradeoff between game and AI performance. Just from the reduction in size, even if the GPU core is the same, I suspect game performance is going to be impacted in the AI PRO version. This should be fine up to a point, if I can get decent AI performance. I only have experience with frontier models, so I don’t really know how big of a model I can run locally on 32GB of VRAM, and how they perform in real world tasks. I want to use it mostly for debugging, and maybe the occasional one-off script, help with project settings, and that sort of task.
My questions:
Does anybody have any gaming performance comparisons between the two?
What LLMs can I reasonably run on the AI PRO card, and what level of “competency” can I expect from them?
Any ideas, suggestions, and pointers to further information are welcome
and the difference is really up to what you feel comfortable with financially. The LLM difference of 16 vs 32GB memory is great for capacity, you can comfortably fir Qwen 3.8-27b at 4 bit quant, which is one of the most popular models used today. (Qwen 4 is releasing this month with further gains.)
The pro cards are generally louder, but the gaming performance is basically the same.
I really like local models, but you have to use them as a supplement to frontier models. For context, my agentic AI runs on dual B70’s and a RTX2000 hosting helper models- so 80GB of VRAM-and it’s nowhere near what you get with even a $20 Claude subscription.
To truly replace frontier for primary workloads, you need something like 4X RTX6000 or multiple unified devices or something, so keep your expectations realistic.
The biggest risk of buying a 9700PRO or Intel B70- is that you’re going to want another one!
So look at this card as a chance to learn to build a working ecosystem, one that can handle some workloads
Thanks for the answer! If we’re still at a level where local AI can’t really do anything valuable for me, then it’s definitely not worth switching. I get very little value out of it, enough to pay an extra K for a card to speed-up debugging, but not a subscription; much less the card and a subscription.
I am surprised it can keep gaming performance in that form factor, I’d expect to hit issues with thermals, and a drop in core clock speed when game runs for any meaningful amount of time. That’s good to know
I think you mean “9070 XT”
Indeed, the two models are RX 9070 XT (gaming GPUs), and AI PRO R9700 (AI GPUs), both Navi 48 architectures. I mistyped
Yeah, local AI is really cool, but I think a lot of us have gotten so caught up in the excitement that we fail to properly communicate just how big the difference is between a $2000 local device and a $20/month service.
The best way to learn whether a 32GB card is right for you is to rent a server with one- their are lots of sites where you can do this and to test out some workloads. Renting a virtual private server is VERY cheap- like a few bucks an hour- and a great place to start.
In short, the R9700 performed about the same as an 9070 XT I had for testing (wasn’t mine). The R9700 was within 2-5% (slower) overall FPS (which you can’t tell while gaming). When using the same FSR and RT settings between the two (my tests were Cyberpunk 2077 on an 120 Hz 1440p widescreen - which is about the same as 4K pixel fill).
However, the R9700 was pulling almost twice the watts at over 400W! You must artificially limit the R9700 to 210-250W, which you should be doing anyways for LLMs so not to waste power.
All tests are with FSR (frame gen) because at 1440p widescreen, you’re pushing almost 4k pixels. All on Linux with just the latest kernel. No driver install needed. 9950X3D2 behemoth with 64GB DDR5 6000.
With that said, the R9700 absolutely crushes Ray Tracing with the heavy Path Trading enabled - especially in Cyberpunk. MechWarrior 5 Clans showed significant improvements as well. It must be the additional GPU horsepower because it was like twice the framerates and very playable at 120 FPS. But again, at like 350-370W of power without being limited.
As a comparison, my 4070 Super 12 GB performs just 10% slower with FSR and normal RT (no Path Tracing). And only pulled 190W for the GPU.
I ended up putting the 4070 Super back in as the R9700 was combined with a second R9700 and went into a 4U Threadripper build.
Search the forums here. Plenty of comparisons.
IME, as a professional software engineer for 3 decades, you really need two R9700s to get any real meaningful long term usage out of them. A single R9700 combined with 64GB DRAM can absolutely get by with Qwen3.8 27B @ Q4. But running things at FP8 with dual R9700s is a significant upgrade you can tell.
I state that because I am an extremely heavy user of my local Ai with harnesses. Consider me as a worse case scenario for review where most others don’t run as much.
Not to mention the work the work Radiance is doing with Swift 1.5 Qwen3.8 Next Flash, especially if you have 64GB of DRAM with dual R9700s.
It must be the additional GPU horsepower because it was like twice the framerates and very playable at 120 FPS.
As far as I know, both boards pack the exact same GPU (a Navi48), same memory interface, and almost the exact same clock configuration (R9700 has the same base clock frequency, but slightly slower running and boost clocks).
I’m in a similar boat, software engineer with close to 25 years of experience between industry and academia. But I’m afraid I can’t justify spending an extra $1800 on a second card. At half the price, it’d be reasonable swapping my current RX9070XT with a R9700… if it can help with debugging (which is the only thing I’ve found AI agents can do better and faster than me, and in non-disruptive ways), and it still serves its primary purpose.
Likewise, subscriptions are completely out of the question. I want fewer subscriptions, not more of them. I specially don’t want to pay a subscription that I know is going up the moment the hype dies down, and VC stops funding the service.
