Discussion Nvidia Blackwell in Q1-2025

[blckgrffn]

You need to be able to navigate through marketing and deception BS.
-When you are GPU limited and maxed out (so 98-99%), latency increases a lot compared to when you are not GPU maxed out.
-DLSS 3 and above, when FG is enabled, automatically activate Nvidia Reflex, which is a technology that fixes this huge latency increase when GPU maxed out by very slightly reducing framerate and technically not being maxed out anymore.

So what's happening in this instance is that DLSS off and DLSS 2 are showing latency numbers without Reflex enabled, whereas DLSS 3.5 and DLSS 4 are numbers with Reflex on. But you can also enable Reflex on DLSS off and DLSS 2 instances, and then the latency would be lower than on DLSS 3.5 and DLSS 4.

The reality is, the way this technology works, makes it inevitable to have around an extra frame of latency, because it holds the future frame along with the previous frame to be able to generate the extra frames between the two.

But tell me more about what it does to the FPS counter. That's what I want hear about!

😉

 

[Keller_TT]

Naturally. But considering the generational uplift, DLSS4 on 50 series shouldn't have more latency than DLSS3 on 40 series. In fact, it should be a bit less, considering the 50 series has better raster.

MFG is going to hit latency over just adding every other frame. Plus DLSS4 itself likely has more overhead and takes that fraction longer, I guess.
So, Reflex 2 is essential for DLSS4 to minimize this. DLSS 3.5 is well optimized at this stage. It remains to be seen if DLSS4 + Reflex 2 can bring that fluidity without perceptible latency over DLSS 3.5/3.7.

 

[SiliconFly]

.

 

[Heartbreaker]

MFG is going to hit latency over just adding every other frame. Plus DLSS4 itself likely has more overhead and takes that fraction longer, I guess.
So, Reflex 2 is essential for DLSS4 to minimize this. DLSS 3.5 is well optimized at this stage. It remains to be seen if DLSS4 + Reflex 2 can bring that fluidity without perceptible latency over DLSS 3.5/3.7.

I'd bet Reflex 2 doesn't work FG/MFG, so they are mutually exclusive features. None of the demos combined them and it would complicate things.

I expect NVidia to separately address:

A) Hard core competitive E-Sports gamers, that won't accept the Latency hit from any FG, with standalone Reflex 2.
B) People insensitive to latency, that want a big number on their frame counter, with MFG.

 

[SiliconFly]

.

 

[CakeMonster]

Having read a lot of speculation and hot takes, I'll be making two unqualified 5090 predictions:

1) CPU limitations will be very obvious and will cause a lot of drama in the discourse (there needs to be at least one major drama with a release)

2) Image quality with DLSS2 will be improved, but there will be so much confusion with FG/DLSS4 that most people won't realize what's what, and it will drown in the drama or praise over FG.

 

[Heartbreaker]

Reflex 1 works well with DLSS3 FG. Same should apply to Reflex 2. It should work well with DLSS4 FG.

Are you extremely confident about that? 😀

I'm only about 90% that they won't work together, instead of 99.9% like I was last time.

We should find out January 24th (5090 Review embargo) if not sooner.

 

[SiliconFly]

.

 

[coercitiv]

I expect NVidia to separately address:

A) Hard core competitive E-Sports gamers, that won't accept the Latency hit from any FG, with standalone Reflex 2.
B) People insensitive to latency, that want a big number on their frame counter, with MFG.

Hmm, that's an interesting point. Reflex 2 works by adjusting the frame that is about to be displayed with up-to-date info from the game engine concerning the camera position. This makes the tech quite valuable for competitive shooters where aim accuracy is paramount, but does not lower latency for other actions within the game.

Nvidia said they're bringing the tech to Valorant and The Finals, two competitive shooters where people clearly favor lowest latency possible and are getting very big FPS numbers anyway. Frame Gen has no place there.

 

[Heartbreaker]

Are you saying with 90% or 99.9% confidence that Reflex 1 doesn't work with DLSS3 FG?

I guess the implication was too subtle.

I was 99.9% confident in our previous discussion, where I said, that NVidia was using interpolation, not extrapolation to generate frames in DLSS 4.

I'm down a small amount to 90% confident that "Reflex 2", won't work with any FG. Reflex 1 will still work...

2) With FG, if 50 series latency is similar to 40 series, they're in the clear. If not, people are gonna have a field day with Jensen.

Clear of what? Latency will be similar to 40 series. Judging by DF video, there will be a much smaller hit for the additional frames, but the big hit is for the first frame, because that is when the current frame is buffered and held back, which is is the source of most of the lag.

 

[SiliconFly]

.

 

[Heartbreaker]

Obviously. If DLSS4 can match DLSS3 in latency, they've already won.

I've just told you it doesn't, more fake frames, more lag. But most of the lag is on the first fake frame.

 

[SiliconFly]

.

 

[Heartbreaker]

I just told you it does. 5090 != 4090. 5090 has better raster (see specs I posted before). Meaning, it can generate a real frame faster with additional headroom left with which it can easily match previous gen latency or maybe even bring it down a bit.

The point of testing MFG is to test the actual technology, and reveal it's characteristics, not setup some kind of weird scenario to make NVidia look good.

So the appropriate comparison will be (all running on 5090, all with Reflex 1 on).

Added Latency of DLLS with 0 Generated Frames?
Added Latency of DLLS with 1 Generated Frames?
Added Latency of DLLS with 2 Generated Frames?
Added Latency of DLLS with 3 Generated Frames?

That will answer the question: how much latency does FG add at each level.

Judging by what DF video showed would could expect something like this, at about ~60FPS real frame rate. But we only have one data point so far, from a restricted source, so there could be some shift.

Added Latency of DLLS with 0 Generated Frames = 0
Added Latency of DLLS with 1 Generated Frames = ~17ms
Added Latency of DLLS with 2 Generated Frames = ~20 ms
Added Latency of DLLS with 3 Generated Frames = ~23 ms

 

[Golgatha]

Having read a lot of speculation and hot takes, I'll be making two unqualified 5090 predictions:

1) CPU limitations will be very obvious and will cause a lot of drama in the discourse (there needs to be at least one major drama with a release)

2) Image quality with DLSS2 will be improved, but there will be so much confusion with FG/DLSS4 that most people won't realize what's what, and it will drown in the drama or praise over FG.

I'm actually pretty concerned with #1. At 1440p or above, the games that benefit most from a 5090 in terms of increased pure raster performance increases may be CPU limited. Thankfully there will be reviews before its available for purchase, so no surprises there hopefully.

 

[SiliconFly]

.

 

[Mopetar]

I have no personal beef with the tech, I mentioned a few pages back that we'll ultimately just end up with a "AI-Tracing" where there won't even have to be rays cast, AI will just insert shadows and lighting etc where it thinks they should be based on the image presented to the viewer.

The sad part is that plenty of people will gladly lap it up and swear up and down how it's better. If you don't believe me you can go look back at the DLSS2/FSR2 threads to find plenty of people saying that it was producing results better than native.

The amount of koolaid drinking and self-delusion in GPU aficionados is just below the HiFi people and their $5000 gold-plated monster cables. I'd bet that if JHH made a GPU that could crawl out a PC case and driver a boot to the groin with enough force to give a person another set of eyeballs that the person who bought it would tell everyone it was the most amazing experience ever.

The emperors clothes are best viewed using interpolated frames of an upsampled ray reconstruction image.

 

[Heartbreaker]

The point of comparing generations is reveal their characteristics, not setup some kind of weird scenario to make the newer generation look bad.

I'll simplify it for you. I'm saying 50 series with DLSS4 + FG will be equally good or better in almost all aspects that 40 series with DLSS3 + FG. Thats it. Something that doesn't go well with you (which is fine by me).

When has NVidia ever released 90/Titan series that wasn't an across the board improvement? That's pretty much a given.

Reputable testers are going to test the percentage differences between 4090 and 5090, at Raster and to a lesser degree, Ray Tracing.

Not which card can generate the most fake frames when MFG only exists on the new generation. That kind of comparison only serves NVidia marketing and nothing else.

The new features will be tested in separate sections of the review, that will all be on the card that has all the new technology,

You can see the same thing when FG was first introduced on 40 sereis. Reputable testers didn't test FG on 4090 vs no FG on 3090, and go: "Wow it's 4x faster". The did Raster/RT comparisons, and then separately tested FG on it's own, only on the 4090.

 

[steen2]

It's a bit more than 2x like for like vs ADA. INT8 and FP16. That's what I remember from the launch. It also adds FP4 precision boost now but the stated 2x TOPS was for INT8 and FP16.

It was rhetorical. You're conflating reduced precision OPs with Nvidia's measure of "AI TOPS". For ADA it's FP8 vs FP4 for Blackwell. Just like cf DLSS3/FG vs DLSS4/MFG for FPS charts. Marketing well done.

 

[moonbogg]

I'm not sure how I feel about the 50 series. It sort of feels like a 40 series refresh with price adjustments where they were needed. People buying the best will always pay more, so they should be charged more. That means 5090 cost $2k+ now. Besides that, it doesn't seem like much has changed besides a $50 discount to the 70Ti class and a 50 series exclusive software feature. The performance improvements seem kind of small, so yeah.

 

[jdubs03]

I'm not sure how I feel about the 50 series. It sort of feels like a 40 series refresh with price adjustments where they were needed. People buying the best will always pay more, so they should be charged more. That means 5090 cost $2k+ now. Besides that, it doesn't seem like much has changed besides a $50 discount to the 70Ti class and a 50 series exclusive software feature. The performance improvements seem kind of small, so yeah.

Yeah, but most chatter is that a 5070 Ti is more like what a 5070 should be, so that price reduction is kind of meh, I would consider that a wash. And one could make that argument for a future 5080 Ti being what a 5080 should be. I still stand by my previous comments that they skimped out on core counts. The ratios are just bad.

But you’re right people will buy the 5090 surely as they have the money and just want the best irrespective of cost; and for people looking to upgrade a 2000 series, and maybe a desperate cohort of 3000 series will upgrade.

The reviews are going to be super important. If the 5090 averages less than 35% more raster than the 4090 then it’s a weak offering; considering most would’ve expected at least 50% if not higher. They showed one game with no DLSS4 shenanigans and it was around that 35% number. What if that was the high end? Hopefully not. But pie on face if so.

 

[coercitiv]

I'm not sure how I feel about the 50 series. It sort of feels like a 40 series refresh with price adjustments where they were needed.

Yes, it's the 40 Super Super edition, with the exception of the 5090 which is bigger and testing the limit of everything (area, power, wallets, board design ingenuity). Performance estimates for the cards are a bit all over the place though, as the faster memory is likely a strength for the compute in modern titles. The better way to estimate performance is by looking at the asking price for each card.

 

[amenx]

The sad part is that plenty of people will gladly lap it up and swear up and down how it's better. If you don't believe me you can go look back at the DLSS2/FSR2 threads to find plenty of people saying that it was producing results better than native.

The amount of koolaid drinking and self-delusion in GPU aficionados is just below the HiFi people and their $5000 gold-plated monster cables. I'd bet that if JHH made a GPU that could crawl out a PC case and driver a boot to the groin with enough force to give a person another set of eyeballs that the person who bought it would tell everyone it was the most amazing experience ever.

The emperors clothes are best viewed using interpolated frames of an upsampled ray reconstruction image.

DLSS4 and MFG will be awesome. I'm paying for my 5090 with cash. 3 out of 4 notes will be photo-copied.

 

[SiliconFly]

.

 

[beginner99]

The single most important reason for more frames is for increased responsiveness. If you don't have it, then you might as well not have it.

And the irony is it actually decreases as each frame is held a bit longer so that they can generate the in-between fake frames.