There are many different niches of ML. 99% of hobbyist would use consumer grade hardware. It’s quite frankly more than good enough.
Even in commercial usage, consumer GPUs provide better value unless you need to do something that very specifically require a huge vram pool. Like connecting multiple A100 GPUs to have hundreds or tens of thousands of gigabyte vram. Those use cases only come up if you’re making base models for general purpose.
If you’re using it for single person use case, something like 4090 is actually the best hardware. Enough ram to run almost anything and it’s higher clock speed than enterprise GPU means your results come back faster.
Even training doesn’t require that much vram. Chat models are generally more vram heavy but if you’re doing specific image training like stable diffusion for how to render your face, or some specific fetish porn, you only really need like 12GB of vram to do it. There are ways to even do it at lower like 8GB but 12 is sweet value spot where even 3060 or 4060ti can do. Consumer GPUs will get that trained in like 30min to 24hrs depending on settings and model.
GPU’s aren’t in a shortage like they were. The majority of new GPUs are just regular people selling them. I wouldn’t personally call it scalping if it’s below MSRP.
That and there just hasn’t been much gains in performance in recent years, so it makes sense to not upgrade for a while. And a lot of people upgraded all at once during the pandemic, so there are less people on the market for a new GPU.
I got a prebuilt like a couple years ago after getting a chunk of money and it still does me ok. There’s a 6800xt in it and it still handles current games ok. I’m in no rush, the only thing I would like is better ray tracing but that’s not enough of a reason for me to spend £700+ on a new card.
At this point, emulating or using a wrapper for 3dfx is not gonna make any game that needs it run bad. Don’t really need the “full speed” of native support anymore.
I know it’s a small set, but for gaming and is honestly king. Unless you want the absolute “I’m willing to pay double the cost for 5% more performance” top of the line, amd is just great.
For AI and compute… They’re far behind. CUDA just wins. I hope a joint standard will be coming up soon, but until then Nvidia wins
The only thing it’s missing is dedicated video decode hardware (which is mostly a convenience) and an equivalent to shadow play. Otherwise it’s a great alternative to a 4080/S
You can even skip the whole suite if you don’t need the AMD per game driver tweaks. OBS now come with direct AMD av1 support and also can record HDR content.(which relive can’t do.)
For AI and compute… They’re far behind. CUDA just wins. I hope a joint standard will be coming up soon, but until then Nvidia wins
I got a W6800 recently. I know a nvidia model of the same generation would be faster for AI - but that thing is fast enough to run stable diffusion variants with high resolution pictures locally without getting too annoyed.
The completely different software stack is a killer. It’s not that you can’t find versions of a model to run, but almost everything that hits the GPU for compute is going to be targeting CUDA, not RocM. From a compatibility standpoint alone this killed AMD for me. I just do not want to spend my time fighting the stack to get these models running.
Yeah, it’s not that I like this state of affairs, but right now the vendor lock-in is so one-sided that it’s hard to say there’s a viable alternative to CUDA. I hope that changes one day.
Admittedly I’m just toying around for entertainment purposes - but I didn’t really have any problems of getting anything I wanted to try out with rocm support. Bigger annoyance was different projects targetting specific distributions or specific software versions (mostly ancient python), but as I’m doing everything in containers anyway that also was manageable.
Add comment