Discussion Nvidia Blackwell in Q1-2025

[Win2012R2]

5090 vs 4090

Only Far Cry compared without DLSS and with RT and such a tiny increase in perf! That just can not be real, considerably larger chip and like 20% increase in game that scales pretty well as long as you've got 3D CPU?

 

[Harry_Wild]

The lower cost 50 series RTX 5070 is around middle $500s and supposed to be more powerful than the current 4090! Coming out in February 2025! I going to be my next GPU for doing e-mail, watching streaming content, internet browsing and YouTubing. This will replace my RTX 3060 Ti Limit Edition.

 

[insertcarehere]

yes matter. in fact all the yes matter.View attachment 114381
They know if too.

"Our architecture is designed to be inefficient, take up a ton of space and power and be bad value to users" says no company marketing, ever.

Coming from the same company 2 years ago that said chiplet GPUs were great for reducing manufacturing costs before promptly backtracking in one generation. LOL.

 

[adroc_thurston]

It's lower than I expected. Only 1.77x membw...

Still a ton. Again, why not flex it. Weird.

"Our architecture is designed to be inefficient, take up a ton of space and power and be bad value to users" says no company marketing, ever.

no it's a very specific thing to say.

Coming from the same company 2 years ago that said chiplet GPUs were great for reducing manufacturing costs before promptly backtracking in one generation. LOL.

you see, 2.5D capacity is not infinite.
It's the reason why N33 was a monodie N6 instead of 1+2 36 or 38CU thingy.

 

[Hulk]

It's not predictive. These are still in between frames. It's holding back a real frame, inject 3 fake ones, and then showing the real one, that it was holding back.

Nothing can change those three fake frames, because they are in the past, and it already has a real frame after them used to generate the steps in between. You can only influence the future after the held back real frame drops.



Example:

Roll a ball across floor.

No FG:

Real frame 1, ball 2 feet away,
Real frame 2, ball 8 feet away,

RTX 40 FG:

Real frame 1, ball 2 feet away,
Fake frame 1, figure out ball is 5 feet away, by interpolating between real frames.
Real frame 2, ball 8 feet away (this frame was held back and used to generate fake frame 1)

RTX 40 FG:

Real frame 1, ball 2 feet away,
Fake frame 1, figure out ball is 3.5 feet away by interpolating more steps between real frames.
Fake frame 2, figure out ball is 5 feet away by interpolating more steps between real frames.
Fake frame 3, figure out ball is 6.5 feet away by interpolating more steps between real frames.
Real frame 2, ball 8 feet away (this frame was held back and used to generate fake frame 1,2,3)

But this would insert lag into the gaming experience? Gaming by nature has to be realtime.
I'm under the assumption that using your analogy it only needs to predict what the screen will look like for 3 frames. So instead of computing those 3 frames, they are predicted. Assuming you are running 100fps, that's 3ms or 1 frame per ms. The 4th frame is then computed. Since only 3 ms has transpired unless there was a complete scene change the difference between the 3rd predicted frame what that actual computed 3rd frame would have looked like will be basically nothing, so then the moving from frame 3 (predicted) to frame 4 (computed) would be smooth.

I would assume that the faster the hardware can natively run the game the more predicted frame can be used. So if the game is running natively at 60fps maybe only 1 frame can be predicted between computed frames. But if it's running 150fps, then perhaps 2 or 3 frames can be predicted.

The super serious competitive gamers will tell us if this technique affects the gaming experience as those high level gamers can detect any "fakery."

 

[pj-]

Any thoughts on the form factor of the 5090? It looks to be dual slot with the PCB not directly attached to the io panel. Would flow through on both ends be enough to allow a small card to dissipate 575w? Or maybe founders edition is lower tdp?

 

[gdansk]

Any thoughts on the form factor of the 5090? It looks to be dual slot with the PCB not directly attached to the io panel. Would flow through on both ends be enough to allow a small card to dissipate 575w? Or maybe founders edition is lower tdp?

Also is the connector angled?
Haven't found good pictures of it yet.
Edit: it is 575W however on Nvidia's site.

 

[Win2012R2]

Either the CPU limit or something in the uarch has scaling oopsies.
Wonder why did they not show any 8k results to showcase 2x membw bump on gb202.

On my 4090/7800x3d Far Cry (all versions) are all GPU limited in 4k.

Surprised they did not show 8k, guess it's a joke of resolution in terms of lack of monitors, plus HDMI 2.2 would really be handy for high refresh rates on those.

 

[adroc_thurston]

Surprised they did not show 8k, guess it's a joke of resolution in terms of lack of monitors, plus HDMI 2.2 would really be handy for high refresh rates on those.

yeah but it would let GB202 flex the membw.

 

[Hulk]

No, what i'm saying is the prediction can incorporate your actions. If the rendering is happening at 50fps with 3 predicted frames for each real one, you get "200 fps". If the game sample input at 200hz, the predicted frames can incorporate your inputs, e.g. changing direction or speed.



There's no indication that it works that way here: https://www.nvidia.com/pt-br/geforce/news/dlss4-multi-frame-generation-ai-innovations/

If it is just interpolation then several of those slides are straight up lies

Predicted frames aren't free. They might be generated 50% faster than computed frames but they still take time to generate, they're just rendered faster than computed frames.

So if you are running computionally at 1 frame/ms or 100fps and every other frame is predicted 50% faster than computed then in 50ms, you can predict 75 frames, or 50% more than 50 frames. So now your frame rate is 125fps instead of 100.

My numbers are guesses to illustrate how I'm conceptualizing this. If you can predict 66 frame out of 100 and the prediction is faster than 50% of computation then frame rates using prediction would be much higher than +25% as in the example above.

If nVidia tells us the increase in performance using this technique and we assume 1 computed for every 3 predicted then we can estimate how much faster on average frames are predicted rather than computed.

 

[branch_suggestion]

I have a couple things to say.
Firstly, I wanna see someone downclock the memory to GDDR6 speeds and check the performance with all the new features. I assume a lot of the extreme perf uplift is membw driven.
Also checking the performance of the ViT model vs CNN on previous generations, the new model is way heavier and so the uplift in frames from upscaling may be a lot lower, whether image quality and stability in motion holds up is another question.
Also clocks look unchanged or maybe even regressed which shows how bulky those SMs are getting with all these ASICs bolted on.

Lastly, CPU bottlenecks are going to be more common, but also harder to detect, lots of fun indeed.
Please Microsoft, update your APIs at something faster than IBM time.

 

[jpiniero]

Wow, Blackwell sucks. No wonder why AMD didn't say anything about RDNA4, they need to see benchmarks now.

 

[Win2012R2]

Well, I am very very disappointed, thought to get 5090 day 1 even from scalpers, but now it seems it might just be 20-25% non-DLSS perf increase over 4090, that's way too low, with 4090 everything is pretty quick already and I could always turn on DLSS 3 to generate fake frames and double it. Getting top end card should not be about those at all.

This brings the question - will 5000 series really last 2 years? If AMD did not stop high end product they could have brought it at the perfect time when 5090 isn't providing that huge increase, what a grand mistake that was.

I assume a lot of the extreme perf uplift is membw driven

Where is that extreme perf uplift? Frame generation does not count.

 

[gdansk]

Wow, Blackwell sucks.

Does it? It might be less than I expected in raster. That remains to be seen. But it's also cheaper than I expected, except the 5090 which was as expected. So overall it seems balanced.

 

[adroc_thurston]

Wow, Blackwell sucks. No wonder why AMD didn't say anything about RDNA4, they need to see benchmarks now.

It's fine but outside of moar interpolated frames even gb202 seems kinda mid I guess?

Does it? It might be less than I expected in raster. That remains to be seen. But it's also cheaper than I expected, except the 5090 which was as expected. So overall it seems balanced.

5090 seems a bit off, given the fattest spec bump of all.
Like with 4090 it was very evident the chungus is more chungus than the usual.

 

[branch_suggestion]

Where is that extreme perf uplift? Frame generation does not count.

The super duper frame gen with the updated models does not seem to give a linear uplift vs the former frame gen, implying some sort of bottleneck.
Outside of that the membw exists for inference and rendering beyond anything else.

 

[jpiniero]

It's fine but outside of moar interpolated frames even gb202 seems kinda mid I guess?

5070 might be slower than 4070 Super in RT, let alone raster.

 

[adroc_thurston]

5070 might be slower than 4070 Super in RT, let alone raster.

No, faster I think. They're fine especially at 549.
It's gb202 that feels a fair bit off.

 

[Win2012R2]

It might be less than I expected in raster.

It does not appear to be much better in ray tracing either - very strange

It's gb202 that feels a fair bit off.

Like WAY off, feels like Titan xp vs 2080 ti - perf increase way too low to bother

 

[poke01]

Wow, Blackwell sucks. No wonder why AMD didn't say anything about RDNA4, they need to see benchmarks now.

can we wait for third party reviews before claiming that…

We all thought ADA was bad until the reviews hit. It’s funny how for Nvidia people determine the outcome at the announcement but for AMD and Intel let’s wait for reviews.

 

[jpiniero]

No, faster I think. They're fine especially at 549.
It's gb202 that feels a fair bit off.

48 vs 56 SMs and Blackwell's clocks are only barely higher.

 

[adroc_thurston]

We all thought ADA was bad until the reviews hit

yea? it was kinda ass until the supers.

48 vs 56 SMs and Blackwell's clocks are only barely higher.

yeah same node, can't do magic.

 

[gdansk]

🤔 Boost clocks are lower than expected, but aren't they rather fictional anyway? My 4090 is often 200 MHz higher than the boost clock. Who knows if they are becoming even more fictional now.

 

[Hulk]

I don't know maybe I'm old school but there is something I don't like about this frame generation. Perhaps I just don't know enough about these new Blackwell GPUs yet.

 

[poke01]

yea? it was kinda ass until the supers.

Nah, the 4090 was great. I’m talking about the architecture itself. Not gimped vram and SMs on lower tiers