Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
Do you have a method for ensuring or increasing your chances at getting a piece of hardware, let's say a GPU, from an online store when supply is short?
During the darkest crypto/pandemic/scalping days, it seemed that only those with scripts or some other method were able to get one directly.
I need a new one soon, might be able to wait for the 4000 Super series, and I'm expecting them to be torn off the digital shelves quickly at launch and then prices will rise.
Go to eBay, buy a 3090. Today. Honorable mention to the 3080 12GB (Ti or not), and if you only care about games, the 7900XT/XTX.
They're cheaper or the same price as new cards of equivalent compute power, EVGA ones still usually have a year's worth of warranty on them, and 24GB of VRAM means it's exactly as capable as a 4090 is for ML workloads (sure, it'll be half as fast for about 40% the price, but it can run models that the 4070 and 4080 will never be able to).
Anything else that isn't either a 4090 or a 100-dollar 1080 is a waste of your time and money. The 3060 and 4060 are barely outperforming 1080s for 3x the price, 3070s are just 2080Tis but with less VRAM, the 4070 and Ti are just a 3090 with half the VRAM, and the 4080 is just too expensive to not have 24GB of VRAM (the reason why it doesn't is because, if it did, it would cannibalize 4090 sales to prosumers who are buying them for ML workloads) and if you're only looking for gaming performance AMD's 7900 series is better in the gap between the 3090/4070Ti and 4090.
Price per FPS is not coming down significantly any time soon thanks to TSMC having a monopoly on all advanced nodes (which they will enjoy far into the decade)- sure, the Super series might result in the eBay 3080s dropping in price by 10%, but the days of faster silicon for the same price are definitively over.
I have a 3060 Ti currently. I want a solid upgrade on it, performance and vram-wise. I don't want to buy used, so 3090 is out. I don't want AMD because they suck at the 3d rendering I like to play with. It's a shame that the 4080 only has 16gb, and that the Super won't increase the vram. Would have been great with 20+ gb. Also it's pretty insanely expensive, but the 4090 costs +50% and gives only +27% performance, and would necessitate a new PSU too, so that's not very tempting either. That's why I'm eyeing the 4080 Super, depending on price.
More options
Context Copy link
EVGA's transferable warranties are nuts; it's a pity they've gotten out of the market. I'll caution that in addition to the risk of getting a bad card, you're also just going to get an older one. Be prepared to replace fans in a year or so, and recognize that you're probably going to get an earlier or lower-gaming-performance model like a 1080 SC than a 1080 Ti FTW3.
That said, there is a performance benefit to some less eye-watering intermediate upgrades that won't show up in the simple frequency check or G3DMark score, even without raytracing, especially when comparing to the simpler model 1080 SCs. It's not a huge difference, especially at the price point, but for many newer and especially VR games it's the difference between a great experience and a moderate or unplayable one.
Agreed that AMD has a number of more cost-effective options for purely gaming performance points, and at more than just the gap between the highest ends.
There are also some ML models out there that can fit into 12GB VRAM but not 8GB, though I expect anyone with that use case is already aware of them, and some rendering use cases where the difference between a 30- or 40- series card and a 10-series card is pretty (bizarrely) large.
4060 Ti has 16gb of vram for 13b/20b LLM models with certain quantization. I can’t find a peer card that isn’t used with the same capacity.
Yeah, that's fair. Most of the stuff I've seen with LLMs either pushes for tiny models at low quantization for speed, or goes all the way to 65b on CPU for intelligence, but I'm sure there's a lot of use cases in the middle.
More options
Context Copy link
More options
Context Copy link
True, but that's kind of the thing: unless you can get to a 3080, I'd argue upgrading to anything else (other than a bargain bin 10/20 series if your card is somehow even older than that) is pointless outside of those specific niches. The 3060 and 4060 are so hideously awful from a price/performance standpoint compared to everything else that I'm amazed they even exist at all, but I guess it's just a tax on people who really aren't paying attention to the reviews and/or don't trust buying used hardware.
Seriously- it's twice the price for a 20% gain (between used 1080 and new 3060/4060), but for 3-4x the price the gain is 100% (between used 1080 and used 3080). Something something Boots Theory; the high-performance cards really are that good, and that's the entire problem with them (and the reason reviewers bitch and complain about every new release that isn't the card everyone actually wants to buy, which is "a 3080 for 400USD"). Unfortunately for them, I figure next-gen consoles are probably going to stick with 4060-tier performance (like they've done for the past 10 years with their respective GPU generations) since that's the only way they'll meet their 500-dollar target price point, so the prices on the highest end cards are probably going to stay high for a long time (inb4 the 5090 is another doubling in performance over the 4090, but this time for 3500USD).
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link