24 comments

[ 2.1 ms ] story [ 35.0 ms ] thread
(comment deleted)
32GB at $2000 seems sensible given the current market. I was wondering when RAM would increase.

I don't understand AI TOPS. Per NVidia:

* 5090: 3352

* 4090: 1321

* 3090 Ti: 320

* 3090: 285

* 3060 Ti: 130

* 2080 Ti: 114

source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-serie...

What's increased tenfold since Ampere? Can someone explain this to me and how it impacts real-world performance?

These are probably int4 numbers.
Yeah 100% its 4 bit vs 8 bit. They did the same thing for the DC units last year to claim a 2x performance uplift.
$2K, straight from nvidia and retailers. But after the scalpers, LLM companies and crypto miners, it will be $4K or more
This is a guess, but I wonder if higher memory speeds led to a bigger jump in performance than ops/sec would suggest.
yup, seems sensible because we have been getting gouged to death with high prices for a minute. Being that 4090's are going for $2000. I suspect we are going to see 5090's easily settle at $3000. The demand from the AI space is intense.
The current market being a monopoly where your next best choice are HPC cards, again by nvidia? $2000 gives ludicrous profit margins and the mid and especially low end 5000 series will suffer. Hopefully AMD eats them up in the non-compute market.
Oh fuck me. Okay, I’d better sell off my 4090s. This is way more sick.
Likely too late, Nvidia's claiming 4090 Perf @ $549, shipping this month.
Well, haha hopefully I can dump them before availability. Thank you. Oh good, not enough RAM on lower end. Good.
Ah, looks like the 5070 = 4070 is not anywhere close to fair. It requires generating multiple frames per game engine frame and running the older DLSS on the 4070.

The real world differences with the same settings looks to be MUCH slower. Seems like 5xxx = 4xxx * 1.3 or so. 3rd party benchmarks that are MUCH fairer should be out later this month.

Despite the flak Nvidia gets for VRAM sizes, 5090 seems to be a decent offer. Can’t say the same about the rest of the line.

FWIW, it appears the best homelab set up would still be 2-4x 3090 if you want VRAM for LLMs, but a single 5090 would likely be the best in class for prosumers on any less VRAM heavy tasks such as image / video generation or deep RL research

I'm curious, what flake are they getting? I assumed they limit gaming GPU VRAM specifically to avoid having them abused by AI people.
Locking DLSS-4 to 50xx only is unfortunate. I don't expect it to rely on any new superpowers. Otherwise - looks like a big step with the AI.
3000 TOPS meaning 3000 terraops per second? That’s 3 peta ops
Wow cool features. Will these translate into higher frame rates, with less heat, and better prices?
What will these prices be if Trump tariffs go into effect?
I just want to see some proper competition in the high end GPU market, not Nvidia launching absurdly priced 5090s.
These vram amounts are fucking depressing,

was expecting 32/24/24/16 not 32/16/16/12 like wtf