Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

theverge.com

Post reply on HN
Print view View original

561–570 of 776 posts

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#561

Post by janalsncm »

I have trained transformers on a 4090 (not language models). Here’s a few notes.

You can try out pretty much all GPUs on a cloud provider these days. Do it.

VRAM is important for maxing out your batch size. It might make your training go faster, but other hardware matters too.

How much having more VRAM speeds things up also depends on your training code. If your next batch isn’t ready by the time one is finished training, fix that first.

Coil whine is noticeable on my machine. I can hear when the model is training/next batch is loading.

Don’t bother with the founder’s edition.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#562

Post by magicalhippo »

janalsncm wrote: post #561

I have trained transformers on a 4090 (not language models). Here’s a few notes. You can try out pretty much all GPUs on a cloud provider these days. Do it. VRAM is important for maxing out your batch size. It might make your training go faster, but other hardware matters too. How much having more VRAM speeds things up also depends on your training code. If your next batch isn’t ready by the time one is finished trai…

Thanks for sharing your insights, was thinking of upgrading to a 5090 partially to dabble with NNs.

> Don’t bother with the founder’s edition.

Why?

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#563

Post by boroboro4 »

boroboro4 wrote: post #172
lemoncookiechip wrote: post #159

I have a feeling regular consumers will have trouble buying 5090s. RTX 5090: 32 GB GDDR7, ~1.8 TB/s bandwidth. H100 (SXM5): 80 GB HBM3, ~3+ TB/s bandwidth. RTX 5090: ~318 TFLOPS in ray tracing, ~3,352 AI TOPS. H100: Optimized for matrix and tensor computations, with ~1,000 TFLOPS for AI workloads (using Tensor Cores). RTX 5090: 575W, higher for enthusiast-class performance. H100 (PCIe): 350W, efficient for data cente…

H100 has 3958 TFLOPS sparse fp8 compute. I’m pretty sure listed tflops for 5090 are sparse (and probably) fp4/int4.

And just for the context RTX 4090 has 2642 sparse int4 TOPS, so it’s about 25% increase

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#564

Post by wruza »

t-writescode wrote: post #450
williamDafoe wrote: post #402

Earlier quoted context omitted.

1. Because you shoot at puddles? 2. Because you play at night after a rainstorm? Really, these are the only 2 situations where ray tracing makes much of a difference. We already have simulated shadowing in many games and it works pretty well, actually.

Yes, actually. A lot of games use water, a lot, in their scenes (70% of the planet is covered in it, after all), and that does improve immersion and feels nice to look at. Silent Hill 2 Remake and Black Myth: Wukong both have a meaningful amount of water in them and are improved visually with raytracing for those exact reasons.

https://www.youtube.com/watch?v=cXpoJlB8Zfg

https://www.youtube.com/watch?v=iyn2NeA6hI0

Can you please point at the mentioned effects here? Immersion in what? Looks like PS4-gen Tomb Raider to me, honestly. All these water reflections existed long before RTX, it didn't introduce reflective surfaces. What it did introduce is dynamic reflections/ambience, which are a very specific thing to be found in the videos above.

does improve immersion and feels nice to look at

I bet that this is purely synthetic because RTX gets pushed down the players throat by not implementing any RTX-off graphics at all.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#565

Post by pshc »

danudey wrote: post #462
numpy-thagoras wrote: post #389

Similar CUDA core counts for most SKUs compared to last gen (except in the 5090 vs. 4090 comparison). Similar clock speeds compared to the 40-series. The 5090 just has way more CUDA cores and uses proportionally more power compared to the 4090, when going by CUDA core comparisons and clock speed alone. All of the "massive gains" were comparing DLSS and other optimization strategies to standard hardware rendering. Som…

> All of the "massive gains" were comparing DLSS and other optimization strategies to standard hardware rendering. > Something tells me Nvidia made next to no gains for this generation. Sounds to me like they made "massive gains". In the end, what matters to gamers is 1. Do my games look good? 2. Do my games run well? If I can go from 45 FPS to 120 FPS and the quality is still there, I don't care if it's because of f…

DLSS artifacts are pretty obvious to me. Modern games relying on temporal anti aliasing and raytracing tend to be blurry and flickery. I prefer last-gen games at this point, and would love a revival of “brute force” rasterization.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#566

Post by ahartmetz »

m3kw9 wrote: post #515
throwaway48476 wrote: post #466

Earlier quoted context omitted.

Fake frames, fake gains

The fps gains are directly because of the AI compute cores, I’d say that’s a net gain but not a the traditional sense preAI.

Kind of a half gain: smoothness improved, latency same or slightly worse.

By the way, I thought these AI things served to increase resolution, not frame rate. Why doesn't it work that way?

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#567

Post by shock-value »

UltraSane wrote: post #528
mdre wrote: post #508

Earlier quoted context omitted.

Yes.

how so?

P and B frames are compressed versions of a reference image. Frames resulting from DLSS frame generation are predictions of what a reference image might look like even though one does not actually exist.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#568

Post by leetcrew »

laweijfmvo wrote: post #511
numpy-thagoras wrote: post #389

Similar CUDA core counts for most SKUs compared to last gen (except in the 5090 vs. 4090 comparison). Similar clock speeds compared to the 40-series. The 5090 just has way more CUDA cores and uses proportionally more power compared to the 4090, when going by CUDA core comparisons and clock speed alone. All of the "massive gains" were comparing DLSS and other optimization strategies to standard hardware rendering. Som…

I started thinking today, when Nvidia seemingly keeps just magically increasing performance every two years, that they eventually have to "intel" themselves, where they haven't made any real architectural improvements in ~10 years and just suddenly power and thermals don't scale anymore and you have six generations of turds that all perform essentially the same, right?

it's possible, but idk why you would expect that. just to pick an arbitrary example since steve ran some recent tests, a 1080 ti is more or less equal to a 4060 in raster performance, but needs more than double the power and a much more die area to do it.

https://www.youtube.com/watch?v=ghT7G_9xyDU

we do see power requirements on the high end parts every generation, but that may be to maintain the desired SKU price points. there's clearly some major perf/watt improvements if you zoom out. idk how much is arch vs node, but they have plenty of room to dissipate more power over bigger dies if needed for the high end.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#569

Post by bb88 »

lukeschlather wrote: post #545
epolanski wrote: post #539

Earlier quoted context omitted.

> Increasing performance per watt means that you can get more performance using the same power. I'm currently running a 150 watt GPU, and the 5070 has a 250 TDP. You are correct. I could get a 5070 and down volt it to work in 150ish range e.g. and get almost the same performance (at least not significantly different to notice in game). But I think you're missing the wider point of my complain: it's been from Maxwell…

> But I think you're missing the wider point of my complain: it's been from Maxwell that Nvidia hasn't produced major updates on the power consumption side of their architecture. Is this true or is it just that the default configuration draws a crazy amount of power? I wouldn't imagine running a 5090 downvolted to 75W is useful, but also I would like to see someone test it against an actual 75W card. I've definitely…

Today's hardware typically consumes as much power as it wants, unless we constrain it for heat or maybe battery.

If you're undervolting a GPU because it doesn't have a setting for "efficiency mode" in the driver, that's just kinda sad.

There may be times when you do want the performance over efficiency.

Re: Nvidia announces next-gen RTX 5090 and RTX 5080 GPUs

#570

Post by sempron64 »

UltraSane wrote: post #495
dragontamer wrote: post #473

Earlier quoted context omitted.

Because if two frames are fake and only one frame is based off of real movements, then you've actually lost a fair bit of latency and will have noticably laggier controls. Making better looking individual frames and benchmarks for worse gameplay experiences is an old tradition for these GPU makers.

DLSS 4 can actually generate 3 frames for ever 1 raster frame. When talking about frame rates well above 200 per second a few extra frames isn't that big of a deal unless you are a professional competitive gamer.

If you're buying a ridiculously expensive card for gaming you likely consider yourself a pro gamer. I don't think ai interpolation will be popular in the market
Post reply on HN