NVIDIA 2025 International CES Keynote: Liveblog

[igormp]

Yeah, on the footnote it says: Flux.dev FP8 on 40 Series, FP4 on 50 Series.

Why NVIDIA

Fucking marketing is annoying

40 series does not support FP4, it's a new feature from blackwell (it was already known given the GB200 is already in production)

 

[AusWolf]

I would assert the opposite, I found AMD's one worse and straight up malicious.

I don't think it is. Of course all CPUs are equal in a GPU-limited scenario. This is just common knowledge, they didn't really invent the wheel there.

But this is like comparing at different graphics settings purposefully to make the new product look better than it is. It's straight up lying.

 

[Visible Noise]

View attachment 378672

The RTX 5070 and Ti comparison charts are unfair. They should be comparing it to the SUPER successors, especially since for both they were significant performance jumps from the non-SUPERs.

Are you really crying “unfair” about marketing slides?

Lol, don’t be a sucker for AMD. They aren’t your freind.

 

[oxrufiioxo]

It's like comparing with different graphics settings. It's false data, there are no two words about it. I'm disgusted to the core.

I agree it does not belong on a benchmark graph... They are improving the 3 core technologies for all RTX owners though so that was probably my favorite slide lol. DLSS SR, DLAA, DLSS RR.


Still I want to see it in action and if it's better than OG Frame generation I just don't think I will care if I am getting 140fps with frame generation at 35ms or 240fps with Multi frame generation at 35ms lol it's going to feel the same and just look slightly smoother.

Why do you have to bring up AMD here?

I mean they both do this so it shouldn't be a surprise amd tries to fake it slightly less but the 5900XT/5800XT was especially bad trying to pass them off as equal to rocketlake in gaming. I am going to take Geralts stance on this one... Evil is evil. Lesser, greater, middling, it's all the same

 

[Dr. Dro]

I used the SVG markup on Nvidia's website to calculate their advertised gen-on-gen uplift.

Comparison	Far Cry 6 improvement	A Plague Tale: Requiem improvement
5090 vs. 4090	27.5%	43.2%
5080 vs. 4080	33.2%	35.1%
5070 Ti vs. 4070 Ti	33.2%	35.1%
5070 vs. 4070	31.3%	40.1%

Bar graphs for 5080 and 5070 Ti have exact same sizes for these two games. And who knows if any of these bars are just an "artist's rendition of performance improvements". But Nvidia won't put real numbers in their presentation so this is the best we've got...

W1zzard will have reviews ready in 3 weeks, don't beat yourself up over it

 

[evernessince]

I was really only here for GeForce and all we get is a price and TOPS numbers? And a promise from the man who disappointed us last gen about how the 5070 "matches the 4090" and the 5090 is "double the performance of a 4090"? In what? In F#&*%^G WHAT? Room heating?! Avant garde sous vide recipes?! Generating an 8K image of a cat in a spacesuit?! I wanted specifics and all I got was vague ideas. I don't know how you can somehow become worse at presenting performance metrics than AMD or god forbid Intel. Christ.

There is zero chance you are generating an 8K cat image on a 5090 with only 32GB of VRAM unless you are using tiled rendering. FLUX Dev takes 24GB to do 1024x1024.

The focus this gen is firmly on AI. The updates to DLSS and FG are purely switches to newer models that Nvidia could easily do as a side project to it's AI R&D.

RTX Mega Geometry is supposed to be the solution to that.
A 4090 is cpu limited a lot of the time. Nvidia knew this and are making changes to make sure it’s not a large issue for the 50 series. They really do think ahead.

The problem is that the 4090 is CPU limited a lot regardless of whether RT is enabled or not. Mega Geometry isn't supposed to solve the overhead problem in general, just tame some of it in regards to RT. Mega Geometry itself is just batching BVH updates on the GPU.

 

[Visible Noise]

Why do you have to bring up AMD here?

Comparing features is relevant.

 

[oxrufiioxo]

W1zzard will have reviews ready in 3 weeks, don't beat yourself up over it

eh got nothing better to do but it is interesting if the 3 lower tiers come close to matching farcry in every game pretty damn good but the 5090 would be mildly disappointing. Which would be interesting because going in the 3 lower cards looked meh with only the 5090 looking awesome...

My guess is the 3 lower cards will be slightly lower than that overall with the 5090 being slightly higher...

Comparing features is relevant.

Not when We want to see actual performance improvements and not how much frame generation X3 improves stuff if all they were announcing was MF Frame generation and showing us how it is different I agree but these are actual gpu's and this is just one thing they do and to focus on that one thing would make anyone thing they are hiding actual RT/Raster gains gen on gen.

 

[wolf]

I don't think it is.

It's off topic so I'll leave it at agree to disagree, because I absolutely can't and won't concede that said marketing slides weren't absolutely heinous and worse than what I'm seeing here.

The common ground between both? Lets see what reputable reviews get with their 'here and now' benchmark suites of expected titles.

 

[theglaze]

Pricing is make believe, like performance:

 

[AusWolf]

I agree it does not belong on a benchmark graph... They are improving the 3 core technologies for all RTX owners though so that was probably my favorite slide lol. DLSS SR, DLAA, DLSS RR.


Still I want to see it in action and if it's better than OG Frame generation I just don't think I will care if I am getting 140fps with frame generation at 35ms or 240fps with Multi frame generation at 35ms lol it's going to feel the same and just look slightly smoother.

Exactly my point. 140 or 240 FPS doesn't matter. 20 or 35 would matter, but that's not a possibility.

I mean they both do this so it shouldn't be a surprise amd tries to fake it slightly less but the 5900XT/5800XT was especially bad trying to pass them off as equal to rocketlake in gaming. I am going to take Geralts stance on this one... Evil is evil. Lesser, greater, middling, it's all the same

Well, I could use a 1050 Ti with my 35-Watt Haswell i7, or with my 7800X3D, it wouldn't matter, would it? Low-end GPUs don't care what CPUs you're using. It's just common knowledge. Trying to sell it as useful info is idiotic, but laughable at best. Trying to sell a product based on false data, though, is malicious and unethical towards the buyer base.

 

[igormp]

There is zero chance you are generating an 8K cat image on a 5090 with only 32GB of VRAM unless you are using tiled rendering. FLUX Dev takes 24GB to do 1024x1024.

That's at FP16, at Q4 you need less than 8GB of VRAM.
A 5090 with its 32GB and FP4 support could do 4096x4096 blazing fast. If you don't give a damn about quality, Q2 should be able to do 8k.

 

[Wasteland]

I'm guessing 15% up and down the stack give or take 5% and 35-45% for the 5090... Besides the 5090 all of them are cheaper though the 5080 is almost 300 usd cheaper when accounting for inflation than the 4080 was...

I have mixed feelings. On the one hand Blackwell pricing seems surprisingly sane. AMD's 9070s will have a very hard time competing.

On the other hand you're still paying a minimum of $750 for a 16 GB frame buffer, and it looks like the perf uplift on non-90 cards will be largely inconsequential. Framegen 4x is firmly, at least to me, in the "cool but not useful" category. It does make one wonder what Nvidia will do about their first-party benchmarks going forward, though--are they just going to keep increasing the framegen multiplier to juice the charts? Framegen 8x in two years? Framegen 16x in four years? At some point, the rollercoaster has to end, doesn't it? They can't keep finding ways to spin 100+% gen-on-gen gains out of nothing ... right? Right?

Anyway, it's been a fun thread. I think it's time for me to go enjoy my super-top-secret futuristic zero-input 24x Framegen tech, though. (It's a movie.)

 

[AusWolf]

Comparing features is relevant.

So you don't care about Nvidia's products. You only care that they're better than AMD. Ok.

 

[oxrufiioxo]

It's off topic so I'll leave it at agree to disagree, because I absolutely can't and won't concede that said marketing slides weren't absolutely heinous and worse than what I'm seeing here.

The common ground between both? Lets see what reputable reviews get with their 'here and now' benchmark suites of expected titles.

The only difference in my book is that if you Turn on DLSS4 you will get the % they claim but as I said above showing one thing a gpu does that a previous generation doesn't do makes it look like they are hiding real performance gains.

Well, I could use a 1050 Ti with my 35-Watt Haswell i7, or with my 7800X3D, it wouldn't matter, would it? Low-end GPUs don't care what CPUs you're using. It's just common knowledge. Trying to sell it as useful info is idiotic, but laughable at best. Trying to sell a product based on false data, though, is malicious and unethical towards the buyer base.

It's technically not false if you turn it on and it does what it shows though. False would be them showing 1000 fps but when reviewers get the product it maxes at 200 fps.....

 

[Visible Noise]

So you don't care about Nvidia's products. You only care that they're better than AMD. Ok.

How did you get here from what I wrote?

 

[Cheeseball]

Are you really crying “unfair” about marketing slides?

Lol, don’t be a sucker for AMD. They aren’t your freind.

Brother look at NVIDIA's 50-series comparison chart and the varying settings between the 40 and 50 series in the footnote.

I don't know why you keep mentioning AMD here. NVIDIA is not a friend too, especially with the misleading marketing.

40 series does not support FP4, it's a new feature from blackwell (it was already known given the GB200 is already in production)

Is FP8 not doable on the RTX 50 due to architectural changes? Why can't they just run the RX 50 series in FP8 quantization for a fair comparison instead?

 

[AusWolf]

It's technically not false if you turn it on and it does what it shows though. False would be them showing 1000 fps but when reviewers get the product it maxes at 200 fps.....

Nobody reviews with FG on because it makes the data incomparable. But that's not a problem here, apparently. I wonder why they don't turn all graphics settings to low on the 50 series while they're at it. It would make them look even better.

How did you get here from what I wrote?

We're trying to discuss the Nvidia keynote here, and just go "AMD this, AMD that, Nvidia is so much better". There's no need for that kind of penis measuring contest.

 

[wolf]

I wonder why they don't turn all graphics settings to low on the 50 series while they're at it. It would make them look even better.

I think we get it man, how much can you just beat the same drum over and over? you've made your point several times now in this thread to the point where frankly seeing you reply it over and over is becoming boring.

 

[oxrufiioxo]

Nobody reviews with FG on because it makes the data incomparable. But that's not a problem here, apparently. I wonder why they don't turn all graphics settings to low on the 50 series while they're at it. It would make them look even better.

I think we are in agreement that they should not use frame generation in comparison benchmarks. Especially a 2x vs 4x because in that scenario they are not showing how good the cards are but how much they've improved frame generation.

I'm not a fan. That doesn't make the data false but it's only showing potentially how much better frame generation is not how good these cards are.

I think we get it man, how much can you just beat the same drum over and over? you've made your point several times now in this thread to the point where frankly seeing you reply it over and over is becoming boring.

I agree with him for the most part though this shouldn't be a thing unless there only goal was to show us how FG improved generation over generation..... If this is the best thing the new cards do that kinda sucks after 2 years. I don't think most people are going to be great we got better frame generation now with 4x the frames vs 2x the frames.... You have to remember the average consumer is going to think 240fps native is the same as 240fps framgen and Nvidia knows that.

 

[AusWolf]

I think we get it man, how much can you just beat the same drum over and over? you've made your point several times now in this thread to the point where frankly seeing you reply it over and over is becoming boring.

Ah so you arguing your point to death is fine, but me replying is boring. Sure, then, have it your way, everything Nvidia does is for the consumer, frame generation is amazing and lying is not a bad thing if you make money on the hype generated by it. Better?

 

[nguyen]

So DLSS4 means that there will be new DLLs for Super Res, Ray Reconstruction, Reflex 2.0 and Frame Generation to improve visuals and performance for previous RTX GPUs.

Meanwhile RTX 5000 takes it up a notch with Multi Frame Generation

Still can't find info about games that will feature Neural rendering though

 

[Visible Noise]

Nobody reviews with FG on because it makes the data incomparable. But that's not a problem here, apparently. I wonder why they don't turn all graphics settings to low on the 50 series while they're at it. It would make them look even better.


We're trying to discuss the Nvidia keynote here, and just go "AMD this, AMD that, Nvidia is so much better". There's no need for that kind of penis measuring contest.

Again, I’m comparing features that were announced in the Nvidia keynote. Just because you don’t like that AMD doesn’t have them doesn’t mean the differences shouldn’t be pointed out.

Of course people review with FG. I hate to tell you but it’s a fundamental gaming graphics technology now. UE5 is literally built around it and image reconstruction. Sorry if your team is behind, but those are the facts.

 

[igormp]

What is Nvidia basing the AI TOPS value off of? FP8?

FP4 I think but I could be wrong.... Sad we only have Farcry to give us some idea of performance lmao fing Nvidia...

It's INT4. The standard for TOPS in most places is INT4.

Is FP8 not doable on the RTX 50 due to architectural changes? Why can't they just run the RX 50 series in FP8 quantization for a fair comparison instead?

IMO it's pointless to compare at FP8. When tensor cores came out everyone started doing mixed-precision because it was so much faster and used half the memory.
Same goes for smaller data types.

Anyhow, Blackwell's FP4 rate is double the value of its FP8 perf, and 4x of the FP16 rate, so you can just figure those out by yourself.

 

[oxrufiioxo]

Ah so you arguing your point to death is fine, but me replying is boring. Sure, then, have it your way, everything Nvidia does is for the consumer, frame generation is amazing and lying is not a bad thing if you make money on the hype generated by it. Better?

Bottom line we have 1 game without frame generation to go by and it looks promising 30% up and down the stack although that would only make the 5070 like 10-15% faster than the 4070 super.... so meh but other than that not bad.

We all know they improved frame generation and for it's fans good for them.

So DLSS4 means that there will be new DLLs for Super Res, Ray Reconstruction, Reflex 2.0 and Frame Generation to improve visuals and performance for previous RTX GPUs.

Meanwhile RTX 5000 takes it up a notch with Multi Frame Generation

Still can't find info about games that will feature Neural rendering though

It's going to take time unless Developers can do it the old way and new way without any extra work most games will still be developed on PS5 first afterall and then PS6.... I found it to be the most interesting thing.