Nah, my 7900XT is more or less about the same as a 6950XT, maybe 10% faster in 4k gaming? Not enough to justify the markup (I spent $750 and got it with starfield, so it was a deal for me). The 6900XT actually performs BETTER in AI related tasks right now due to AMD dragging their feet on ROCM (less gay CUDA) support.
Same/similar. Got a 3060 for CUDA and it's been surprisingly alright on Linux.
I don't think it's a really bad card, feels like an accidental refresh of 6800XT with 7000 series features at the same/similar street price.
Idk. It's really up to you to decide how serious you are about the stuff. People do AI stuff on AMD cards. It won't be as fast or efficient as Nvidia, but it's not like it just won't do it.
Just a couple of weeks ago AMD announced that they've made huge improvements for SD.
I’ve done Stable Diffusion on my 6900XT since the release and it’s worked fine. It’s slower than a comparable Nvidia card, but it works and has decent size RAM. I’d still recommend getting Nvidia if this is what you’re after because AMD support is kind of shaky, I have to run a special ROCM version of PyTorch in docker just to load the AI, which is a lot more hassle than the basically just works of Nvidia.
By the name that’s probably a windows exclusive. Typical.
L1 Techs have a few reviews on different models and a general opinion piece on it:
This is what I was doing. In fact, @snov helped me get started. But now I'm just using DirectML (specifically I'm using ComfyUI) and producing images pretty quickly. Whilst my 7900XT probably isn't quite as fast technically as Nvidia like for like, it has 20GB VRAM which I couldn't afford with an Nvidia card, is excellent for gaming and the software side is getting better all the time. Just a few months ago I was jumping through all manner of technical hoops to get Stable Diffusion working on Windows with AMD. A couple of weeks ago I literally just downloaded ComfyUI from its repo and typed python main.py --directml into the Windows terminal and I was up and running. I guess I already had Python installed if you want to count that but you'd need to with Nvidia as well.
Does this apply to all cards or just the 7xxx series?
Yeah that's what concerns me, I already seen this happening in the IoT sector with rpi were basically those SBCs became the standard so everything its tailored to those and all other boards even if better priced with better specs are incompatible with those solutions or barely runs at all with too many bugs to be considered reliable at all, and if its for industrial operations you just can't risk it.I’ve done Stable Diffusion on my 6900XT since the release and it’s worked fine. It’s slower than a comparable Nvidia card, but it works and has decent size RAM. I’d still recommend getting Nvidia if this is what you’re after because AMD support is kind of shaky, I have to run a special ROCM version of PyTorch in docker just to load the AI, which is a lot more hassle than the basically just works of Nvidia.
By the name that’s probably a windows exclusive. Typical.
As fun as SD is to use I rather move to other stuff with more professional potential like GPT alternatives like LLaMA, tho I was just reading some stuff and it turns LLaMA is more CPU than GPU dependent? now I don't know if I'm better off with my Ryzen or should've gone with one of those used $100 Xeons which have more cores than a redhead kid has freckles, even if the Ryzen wipes the floor with it in single core perf.
This is kinda baffling to me, that there is 0% performance uplift and 0% value improvement. What is even the point of releasing a new card at that point?
From an MSRP standpoint there is a value improvement. It's $500.
7700 XT was the first nearing acceptable bump in new price/performance since 3060 Ti almost 3 years ago, and 7800 is a bit better than that.
It's probably more like a 5% uplift (I think 6800 XT was going for $480 so almost same value). It could have been intended to be more, but RDNA3 is a flawed architecture that did not hit its original performance targets.
I haven't played around with LLMs yet though I hope to. But my early investigations of it suggest that it wants very high RAM. I.e. 30GB+. I think only a few can you hope to run in normal domestic machines. Training might be out of the question. Happy to be corrected though - I believe there are some newer ones that can be run on high spec home machines...
I thought I was clear. It's a big jump forward from anybody on a 5700, 580 or 480. Most people don't buy cards every year and given the insane prices during the mining craze and the fact they hadn't sunk that much after it, there's a large number of people holding off. They hold off because there is uncertainty - will the new cards be wildly better, will they be similar, will they be cheaper, on and on. Lots of reasons to hold and wait. Now the waiting is over - people can see what the new cards are like, the pricing, etc. They have certainty - so all the people who were putting back an upgrade can now buy without feeling they're about to find out they screwed themselves by not waiting. And when they upgrade, what are they going to buy - an older generation or the newer one with a little better performance and nice new features? The latter, generally.
I was on a 5700xt until just recently, and it feels underwhelming that the 7000 series doesn't beat the price/perf value the 6000s have offered for the last 6 months or so. I'm glad I didn't wait. Say hello to the new card, same as the old card.
I have no doubt it will sell, the 6800 offers the same value proposition at present and it sells just fine. I guess I just have this old fashioned expectation of generational improvement.
