Discussion Nvidia Blackwell in Q1-2025

[Heartbreaker]

Thought this was an interesting slide for the 5090. I'm assuming DLSS On also includes Reflex 2?

I think you will find it's just reflex 1.

But remember that's against DLSS completely off.

So it's comparing the latency of running at native 4K with no reflex, compared to DLSS performance (so running at 1080p) with reflex on.

Just more misleading comparisons from NVidia.

 

[StinkyPinky]

I simply cannot wait. Playing games with AI generated NPC dialogue and with fake framerate numbers and playing them on 600Hz monitors that people pretend to notice the difference on. This is the great future we were all promised.

 

[StinkyPinky]

very underwhleming increase.... they used GDDR7 and this is what they could muster. poor RT improvement

There's no incentive for them to do so. Nvidia is like Intel was prior to AMD releasing Ryzen. AMD cannot keep up so Nvidia will just drip feed us generation after generation from this point forward. At least that's my view of it.

 

[MrTeal]

Everything is about equal to shader count increase, except 5070. Maybe they are just pushing 5070 extra hard, because 4070 Super had the biggest gains and if they didn't it would risk underperforming the 4070 Super which would be a tremendously bad look. (runs off to check 4070 super review)

According to HWUB review, 4070 Super is 19% faster than 4070 at 1440p Raster, and According to these benchmarks 5070 is about 20% faster than 4070.

That just makes it equal to 4070 super, so they really had to push the 5070 just to NOT lose. No wonder it got a price cut.

While it fits in with the known numbers, if Blackwell really has very little per shader uplift at ISO clocks without using specific features, it's going to be a boondoggle.

Reviews should be really interesting if that's the case though, I expect a lot of polarization between "This is trash, what a disappointment" and "2x faster than last gen"

 

[SolidQ]

Horizon . You can compare

 

[Saylick]

Thought this was an interesting slide for the 5090. I'm assuming DLSS On also includes Reflex 2?

View attachment 115005

Might just be DLSS2, which uses a lower internal resolution which allows for higher native fps (that is, “real” fps, not interpolated)

 

[poke01]

blackwell SM

 

[Saylick]

While it fits in with the known numbers, if Blackwell really has very little per shader uplift at ISO clocks without using specific features, it's going to be a boondoggle.

Reviews should be really interesting if that's the case though, I expect a lot of polarization between "This is trash, what a disappointment" and "2x faster than last gen"

Reviewers who say this is trash: *Nvidia sends stern email about how they will no longer be receiving review samples for not falling in line with the reviewer’s guide a la HWU.*

Reviewers who say 2x faster and “FG is here to stay, all frames are fake frames anyways”: *Nvidia hands over bag with $ symbol on it*

 

[Saylick]

Density improvements are nil, as expected:


Blackwell is just a scaled up Ada with a ton of AI software on top. Again, with hardware/silicon improvements coming to a standstill, I would not be surprised if Nvidia starts to use even more software segmentation in the future to entice people to upgrade since raw improvements aren’t happening as quick. That goes for raster, RT, and even tensor too, seeing as how without lower and lower precision tensor OPS will be be subject to silicon scaling laws.

 

[coercitiv]

Thought this was an interesting slide for the 5090. I'm assuming DLSS On also includes Reflex 2?

View attachment 115005

For those games it's just Reflex 1, the second version is only coming to Valorant and The Finals for now.

 

[Heartbreaker]

Anyway, get ready for a deluge of videos and articles telling us how important it is to match monitor refresh rates and how unimportant it is to care about lower latency and stuff.

In fact we can start right away

I wonder how many people have super high refresh monitors. NVidia is leaning heavy on showing 300 FPS and over numbers.

But most of the affordable gaming monitors are more like mine ~165Hz.

Even if you can make an argument that matching your refresh is important, 300 FPS with fake frames is literally doing more harm than good with a 165 Hz monitor.

 

[CakeMonster]

Density improvements are nil, as expected:
View attachment 115008

Blackwell is just a scaled up Ada with a ton of AI software on top. Again, with hardware/silicon improvements coming to a standstill, I would not be surprised if Nvidia starts to use even more software segmentation in the future to entice people to upgrade since raw improvements aren’t happening as quick. That goes for raster, RT, and even tensor too, seeing as how without lower and lower precision tensor OPS will be be subject to silicon scaling laws.

They'll do that regardless, but at least there'll be some N3* gains with next gen. Not sure if that node will be cheaper (NV loves their margins) than N4 is now, but possibly. I really don't worry about NVidia's lineup until the gen after that again, but they'll probably have a chokehold for longer than that regardless.

 

[CakeMonster]

Even if you can make an argument that matching your refresh is important, 300 FPS with fake frames is literally doing more harm than good with a 165 Hz monitor.

Theoretically, I think you'll notice the latency less the higher the input/base frame rate. I don't think 120->240, or at least 240->480, will feel as bad as 60->120. While I'm not at all interested in FG right now, I could easily imagine using FG if that reasoning checks out to real world gameplay feel in the future.

 

[PJVol]

Reviewers who say this is trash: *Nvidia sends stern email about how they will no longer be receiving review samples for not falling in line with the reviewer’s guide a la HWU.*

I suspect that's already the case de facto. Just think about how come there's currently a separate benchmark section in every review with so called "RT performance"?

P.S. As always, "focus groups", GPP or whatever, is "nothing personal, just business" )

 

[Win2012R2]

blackwell SM

Hang on, so they walked away from dual issue and only now have FP32 that can also do INT? That's a big change if true - where is IPC uplift then?

 

[CP5670]

One really off-putting feature is the performance uplift is about equal to power consumption uplift. For me, the only game I'd like more performance in is CP2077, 1440p@ultra with RT. Right now I'm getting around 90 FPS pure raster performance and upgrading from a 7800X3D to the 9800X3D didn't change that, so I know it's GPU limited. That said, I'm starting to lean towards keeping my 4090 and wait it out for the next generation. Also, anything below a 5090, I have no interest whatsoever in.

I had the same game in mind, but finished Phantom Liberty and don't see myself going back to it. It runs great without RT, and the visual impact of RT is questionable. It makes a difference in some areas and not others. I found myself often toggling it on and off.

 

[Heartbreaker]

Theoretically, I think you'll notice the latency less the higher the input/base frame rate. I don't think 120->240, or at least 240->480, will feel as bad as 60->120. While I'm not at all interested in FG right now, I could easily imagine using FG if that reasoning checks out to real world gameplay feel in the future.

I agree.

But my point is the fundamental tradeoff of FG is that you are trading away real responsiveness (even if minimal), for some gains in visual smoothness.

If you can't actually see the extra frames because they are beyond your monitors refresh, then you only have have the negative side of that tradeoff, however minimal.

So 300 FPS fake with fake frames is just a negative if you have a more typical 165 Hz monitor.

 

[Golgatha]

These numbers seem to be vs. the original 40 series. That would mean the 5070 is actually slower than the 4070S without AI, the 5070Ti is 5% faster compared to the 4070TiS and the 5080 is 10% faster than the 4080S.

nVidia is basically doing a $50 price cut at every tier of performance below the 5080, giving a small performance bump at the 5080 for the same money, a performance + nearly equal price increase at the 5090 tier, and giving the consumer a few extra nifty software tricks for good measure for purchasing a 5000 series GPU. Really nothing overly exciting to begin with, and if tariffs raise pricing on the whole line-up soon, there will be a lot of stock sitting on store shelves.

 

[blckgrffn]

I agree.

But my point is the fundamental tradeoff of FG is that you are trading away real responsiveness (even if minimal), for some gains in visual smoothness.

If you can't actually see the extra frames because they are beyond your monitors refresh, then you only have have the negative side of that tradeoff, however minimal.

So 300 FPS fake with fake frames is just a negative if you have a more typical 165 Hz monitor.

Buying a 5090 and using a 24" 1080P ultra high refresh monitor for "smoothness" - lol. Not that many super high refresh monitor options out there it seems like.

Also, would the "5070" have been better off with 16GB GDDR6 vs 12GB GDDR7 in terms of a place to live in the lineup?

 

[Golgatha]

I had the same game in mind, but finished Phantom Liberty and don't see myself going back to it. It runs great without RT, and the visual impact of RT is questionable. It makes a difference in some areas and not others. I found myself often toggling it on and off.

I guess I'm of the opinion that I shouldn't need to turn anything down on a $1600 video card, regardless of how incrementally better leaving a feature on or off is.

 

[lixlax]

Using SVG height
5090 vs 4090
Resident evil 4 (RT) 31%
Horizon Forbidden West (dlss) 32%
5080 vs 4080
Resident evil 4 (RT) +15%
Horizon +15%
5070ti vs 4070ti
Resident evil 4 (RT) 19%
Horizon 22%
5070 vs 4070
Resident evil 4 (RT) 20%
Horizon 22%

I expected something similar after CES, but against SUPER variants. Below 5090 the RTX 5000 series seems to be just for expensive software upgrade.
Edit: And these rough values are from first party slidedeck, it could be worse still.

 

[gdansk]

Density improvements are nil, as expected:
View attachment 115008

Blackwell is just a scaled up Ada with a ton of AI software on top. Again, with hardware/silicon improvements coming to a standstill, I would not be surprised if Nvidia starts to use even more software segmentation in the future to entice people to upgrade since raw improvements aren’t happening as quick. That goes for raster, RT, and even tensor too, seeing as how without lower and lower precision tensor OPS will be be subject to silicon scaling laws.

If these are official/accurate then GB205 is bigger than some expected. GB203 should be an improvement in performance at basically the same die size. But I wonder how much of that could be simply due to power limit increase.

 

[Heartbreaker]

I'm curious what other people refresh is in these GPU discussions, so I did a poll (since steam doesn't show these numbers and I couldn't find another poll).

Question - Poll: What is your monitor refresh?

Too bad steam doesn't show this in their survey.

forums.anandtech.com

 

[CP5670]

I guess I'm of the opinion that I shouldn't need to turn anything down on a $1600 video card, regardless of how incrementally better leaving a feature on or off is.

I do it all the time with ultra settings, volumetrics, DOF, shadows, pretty much anything except textures. I would rather have constant 120fps. In fact Cyberpunk the "screen space reflections" setting alone makes a huge difference.

 

[adroc_thurston]

Blackwell is just a scaled up Ada with a ton of AI software on top

Naaa, the SM is just very underwhelming to say the least. I guess this had a higher fmax target?