I’m currently still using my trusty old Sapphire RX 580 8GB Nitro card on my upgraded system, which is a Ryzen 7950x, 64GB RAM on a ROG STRIX X670E-F GAMING WIFI motherboard. I think the GPU is now my bottleneck in processing time. Also, the hardware is no longer officially tested for amd opencl drivers and rusticl is not quite there yet.
Now I’m wondering how much you can gain in processing speed by changing GPU. I already tried an ARC 770 card which should have triple FP32 performance but I honestly couldn’t tell much difference. I returned the card because of the fan noise. So where should I look next, or should I look at all? Some of the opencl bugs are really nasty and although @hannoschwalm really kicks the bugs’ butts, working drivers would be really nice. Also, I don’t want to switch to prorietary NVidia stuff.
Without nvidia you could wait for new intel cards, which should have higher end models with better cooling.
I find it strange that the fans were even on in the Arc 770 as modern GPU’s have their fans turned off unless they heat up, which doesn’t happen with “light” work (which darktable qualifies, at least from personal experience). Iirc there were some problems with the fan control under linux, maybe that was the issue?
Either way, Arc Battlemage is scheduled for 2024 so that seems like the only option if you don’t want to go with a newer amd card.
It was a Acer Bifrost card and the fans were running low but turning up a bit and down constantly creating a really bad sound I could not ignore. Maybe Intel’s own cards were better in that regard.
Completely understand you, replaced the AMD stock CPU cooler because of that… It’s easy to tune out constant volumes but other than that it’s very nerve racking.
I’m not familiar w/ AMD stuff, but there are tools to monitor GPU usage for Nvidia. Have you installed anything to look at GPU load / mem-usage while editing? I’m assuming you’re on nvme disks or equivalent
Ran a quick test with an RTX 3080, by power limiting it to see what the performance impact would be. I realize that this only lowers clock speed, but it’s an indicator of how a lower performing card would respond, with the same memory amount:
100W: pixel pipeline processing took 14.122 secs (30.190 CPU)
200W: pixel pipeline processing took 4.052 secs (19.058 CPU)
350W(Default): pixel pipeline processing took 3.738 secs (19.459 CPU)
A quick glance at nvidia-smi and I saw power peak at 270W when unlimited, so it’s not surprising that limiting at 200W didn’t have much of an impact. Plus 3080’s are overtuned from the factory, and there’s very quick diminishing returns after 250W. Such a shame we can’t lower voltage under linux though, and only set power limits.
I never have trouble with the nVidia proprietary driver on Arch. 95 times out of a hundred, it is updated right along with the Linux kernel. On the few occasions when it isn’t, I just recreate the kernel image with “mkinitcpio -p linux”.
In the end, if you’re buying a new card I’d go for nVidia, if only because AFAIK most of the devs don’t have AMD or intel cards, so darktable is probably tested better on nVidia.
I would simply try to get a current model RTX for as much money as I felt like putting into. My PC is now a few years old, and I’m still satisfied with the 6 GB RTX-2060 that it came with, so I’m not really familiar with the latest ones.