Hacker News Reader

Comment by colpabar | Hacker News Reader

Ask HN: I am still using a 980. What should I get? I do not care at all about having the latest/greatest, I just want an upgrade.

View on HN

I think you should wait until late November for dust to settle(budget segment won't change that much because of it if you don't consider buying used). By then both Nvidia and AMD will release their cards(it is likely that Intel too, but Alchemist is a joke), mining cards will hit the market in large quantities and pricing for everything will change drastically.

Go down this list until you reach performance level you'd like:

https://www.videocardbenchmark.net/gpu_value.html

Alternatively look at https://www.logicalincrements.com/

Depending on how much you want to spend, a 1080ti, 2080ti and 3080ti are all good upgrades (roughly doubling in price with each step). All of them come with 11GB or more memory, making them very usable for not only gaming but also "average" ML (e.g. stable diffusion runs great on my 2080ti, and it's very usable for developing tensorflow models).

M1 runs SD in a few seconds. Is 2080ti faster than that?

It is hard to tell since you don't provide exact numbers. I've seen SD generating 50 samples 512x512 image on 2080 Ti in 8 seconds, but it is almost certain that there is a faster version out there.

A 2080 Ti, which requires a supporting platform, seems like a really bad deal compared to a 16GB M1, which you can get factory refurbished for $1000.

Stable Diffusion is but 1 project. M1 gonna be a huge bottleneck for anything that takes time. That "Ultra" for $4k is about 5 times slower than a single 3090. It will be nearly 10 times slower than 4090.

A used 2080TI goes for around $500, with the remaining $500 I can get a computer with a 6x3.6GHz Ryzen 5 and 16GB RAM. Which is something completely different than a Macbook Air, so that's kind of an apples to oranges comparison, but I don't see the big price advantage. And presumably GP already has the "supporting platform", since they are upgrading, making the 2080TI upgrade half the price of a switch to an m1?

It is VASTLY slower than 2080Ti. There is no competition at all. I don't know exact performance numbers, but I would estimate that M1 takes more than a minute to generate one image.

UPD: asked a friend to generate image on m1 with diffusion bee. It took 4 minutes 30 seconds to generate one 25 sample image. It is not even close and much worse than I thought.

Must have been an 8GB model. On my 16GB it takes a few seconds.

Could you provide the exact settings you use for generating the image? Settings matter a lot in terms of it/s. Also please share the exact it/s you're getting :)

It was on 16GB model. Can you provide details how do you run them and how fast it can run.

I just tried it, on my 2080ti I get a 50 sample 512x512 image in 7.0 to 7.5 seconds (using PLMS, DDIM or k_lms). Or at 256x256 with 8-image batches about 1.15 seconds per image.

I've seen 30-60 seconds for M1 / M1 Pro. Is that what you mean by a few seconds?

For ML you need a card that supports bfloat16, which rules out 1080Ti, and probably 2080ti.

The 2080ti supports bfloat16 in the sense that it (mostly) runs, but to get any speedup over fp32 you need a RTX3000-series GPU.

I upgraded from a 970 to a 2080 a while back and have been very happy with it. Still plays modern games at good frame rates, though not with max settings.

Obviously your choice depends greatly on budget + availability of hardware, as well as your desired resolution, but I'd suggest the 2070 or 2080 (maybe 2080ti?) as a good starting point.

It depends on your resolution. If you have a 1080p monitor, look for a great 1080p card. same for 1440p, etc.

https://www.tomshardware.com/reviews/gpu-hierarchy,4388.html

I upgraded my wife's 2015 workstation from 980ti to 2080sc of some sorts. The limit is pci bandwidth of the MB, anything later made no sense. My kids use that as gaming PC and the happiness index on the latest titles (valorant, halo) went way up. It cand compare to the 3080 that I have in the latest gaming PC, but it does great. Bonus is that I sold the 980 to someone for 70 dollars, I was going to trash it but apparently it is still worth something

I have a 980, too, and it's great. It plays most stuff on Steam without any problems. I'm only considering an upgrade because I want to play Flight Simulator and Cyberpunk on max graphics. I'll be getting a 3080-- they're fairly mainstream and a great value. I love ultra-quiet hardware so I'll be getting the Asus 3080 Noctua, which is the quietest version by a huge margin.

I think your local prices matter the most. Here in Hungary, right at the moment, looking at one specific large retailer, the 2060 makes the most sense, if we look at price/value. Compared to the 980, it offers <50% extra speed. It really depends on your circumstances, to judge whether that works for you or not.

IMO, get the best 3000-series card you can afford.

You might be able to pick up a card that was used for mining super cheap.

I've considered this, but I'm not sure. Is there any concern about cards used for mining being too "beat up"? They say you shouldn't buy a used car from someone you know was using it for racing. Does that apply to graphics cards used for mining?

Apparently mining cards are ok because they suffer a constant load, sometimes run at a capped peak performance (?), vs games which can be very spikey.

Whether this is just misdirection thrown around by crypto miners looking at a hole of sunk assets or not, I could not say. Personally I'd just buy new, but I have also probably never spent over 300usd on a card.

My previous card (Radeon 7970HD) was used for BTC mining in a gaming PC and lasted until about 2018 (So about 6 years). My current GPU is a 1080Ti that used to be in a mining rig and shows no signs of damage at about 4 years of age.

I'm looking forward to buying a used crypto mining card when the current one dies.

The major concerns would be the thermal paste degrading or the fan dying. Both are easy and cheap to remedy, and are probably worth preventively fixing when you buy the card.

A week or so ago 3090s were going for < $1000, which is a steal if you need the memory for ML or other intensive apps.

perhaps a 3080 , maybe 4080 12gb when the price drops a bit ? I'm envious I went form a 970 to 1660ti I'm not sure what my next move is Im usually more comfortable finance wise with 70s than 80s

Get the 4090 and forget about upgrading for a long while.

Or even 3090

I would not get a 3090 for gaming. It's only 10-20% faster than a 3080, yet is significantly more expensive. Games won't use the extra VRAM.

Parent didn’t mention gaming, so I assumed general use, hobbyist ML etc.

Even for gaming though wouldn’t 3090 be more future-proof? 4K is really the way to experience games in 2021 and beyond

10-20% faster isn't a lot of future proofing. Certainly not worth paying double.

You certainly don't need a 3090 for gaming in 4K. nVidia's claim of the 3090 being a "4K" card is marketing bullshit. If a game can't be run in 4K on a 3080, then the 10-20% performance increase of a 3090 isn't likely to change that.

Whats your use case?

Reply on news.ycombinator.com