News Benchmarking Blackwell and RTX 50-series GPUs with Multi Frame Generation will require some changes, according to Nvidia

[DougMcC]

I really hope most benchmarking sites don't go along with this. I don't want fake frames packed with artifacts disconnected from anything actually happening in the game. I want to know how many real FPS my card can deliver, and if the 5090 has nothing to offer in that dimension maybe it's not the card I want to buy.

 

[m3city]

Dont benchmark it. Measure natively generated frames at correct resolution, compare and thats it. Add a side notes after conclusions about fake frames etc. Excitment about it fuels efforts to work on this crap instead of reducing power use. Also game devs implement bs graphic details that make nerds horny, and add nothing to gameplay.

 

[DougMcC]

FPS has never been relevant. FPS was invented by folks who thought increasing this number meant you would have a better gaming experience thus pushing you to buy expensive cards. Watch a movie and you realize it's not important. Anything above 90 fps is not detectable, but you can detect a drop in frame rate even if it's 300 fps to 250 fps. Ever wonder what happens to the frames that are generated when you go past the refresh rate of the monitor, they get thrown away. You spend all that power to render the frame and you never see it.

Very few games are getting 90 fps at 4k yet. So until they are, it still holds relevance, even to your standard.

But it actually does matter. Plenty of people have proven that frame latency advantages can give gamers a substantial edge, even among relatively non-pro gamers. At least up to about 240 hz the latency advantage can be quite substantial, similar to an extra network hop (or 3 or 4).

 

[liolio]

Why don't you measure just PC Latency in different DLSS configurations? You can list the fake fps in a footnote or ignore it: that part is not important

 

[Roland Of Gilead]

Some of the over the top reactions here are nuts! Do you think if you complain enough on these forums, nVidia will be listening and about face on their GPU roadmap anytime soon?

Like a lot on here, I too am not necessarily fond of DLSS or FG. I like to play my games at native with high eye candy settings, and FPS above 100, with 1% lows at least at 60fps for some smoothness. I play mostly online shooter types, so I have no need for RT effects, they pass me by unnoticed. Most games run fine and the way I want them to.

With that said, with my current system, DLSS actually helps for some games I play. I find I'm having to use it more, specially with engines like UE5 stomping all over my CPU/GPU. I don't use FG in anything else other than Indiana Jones and the Great circle. It's nice and smooth and the added latency isn't noticeable for that type of game. But, I can't afford to buy a new system every two years so it can keep up with just pure raster performance to max out my monitor. I like many others, will make use of this tech. It gives those with lesser systems and zero disposable cash a longer time in the game (pardon the pun!) without having to drop 300-400 on a kick ass CPU and another 500-600 for a decent mid to high GPU.

For all those saying FG+Multi simply shouldn't be tested are missing the whole point of a GPU review. It's to 'review' the features of the card, not just the features you want reviewed. I think it's a good thing that Jarred/Tom's are reviewing how they bench these cards. There are many more gamers out there who aren't loaded who will want to know this info, how they perform, with AND without FG or DLSS or whatever.

On a side note, wow! 10 hours for one GPU. @JarredWaltonGPU , good luck with that! ☘️

 

[Peksha]

FrameGen would be to compensate for VSync "on" situations

Doesn't turning on FG require Reflex which will disable vsync?

 

[JarredWaltonGPU]

Some of the over the top reactions here are nuts! Do you think if you complain enough on these forums, nVidia will be listening and about face on their GPU roadmap anytime soon?

Like a lot on here, I too am not necessarily fond of DLSS or FG. I like to play my games at native with high eye candy settings, and FPS above 100, with 1% lows at least at 60fps for some smoothness. I play mostly online shooter types, so I have no need for RT effects, they pass me by unnoticed. Most games run fine and the way I want them to.

With that said, with my current system, DLSS actually helps for some games I play. I find I'm having to use it more, specially with engines like UE5 stomping all over my CPU/GPU. I don't use FG in anything else other than Indiana Jones and the Great circle. It's nice and smooth and the added latency isn't noticeable for that type of game. But, I can't afford to buy a new system every two years so it can keep up with just pure raster performance to max out my monitor. I like many others, will make use of this tech. It gives those with lesser systems and zero disposable cash a longer time in the game (pardon the pun!) without having to drop 300-400 on a kick ass CPU and another 500-600 for a decent mid to high GPU.

For all those saying FG+Multi simply shouldn't be tested are missing the whole point of a GPU review. It's to 'review' the features of the card, not just the features you want reviewed. I think it's a good thing that Jarred/Tom's are reviewing how they bench these cards. There are many more gamers out there who aren't loaded who will want to know this info, how they perform, with AND without FG or DLSS or whatever.

On a side note, wow! 10 hours for one GPU. @JarredWaltonGPU , good luck with that! ☘️

Well said. I'm not at all planning to make DLSS on (or FSR2/3/4 on, or XeSS 1/2 on) the default for testing and comparing performance. But it is a feature that can be useful. And when someone says "I have never used DLSS ever" then I wonder if they even have a leg to stand on. Maybe they're only using AMD GPUs and thus couldn't have ever used DLSS?

I do know that from experience, DLSS quality upscaling at 4K gives a solid boost to performance and really doesn't degrade image fidelity much. And with DLSS transformers coming, it might actually look better than non-DLSS in some games. We'll have to wait and see. But discounting features out of hand is a great way to halt progress.

I've poked at framegen (FSR3 and DLSS3) enough now to have a good grasp of what it does and doesn't do. Artifacts tend to be more noticeable with FSR3, DLSS3 tends to provide less of a performance boost. But now we're also getting framegen transformer models and that can reset the experience again. And what will MFG look and feel like? Probably more of the same, and probably you'll want 160 FPS or higher with MFG to make it viable — which means you really should have a 144 Hz or 240 Hz or higher display.

I inwardly cringe every time Jensen brags about rendering one out of four, or one out of eight, or now one out of 32 pixels. Because when he says "render" you should also inwardly say "and sample user input." And suddenly sampling input once for every four "frames" that get sent to your display doesn't sound quite as enticing.

 

[-Fran-]

Doesn't turning on FG require Reflex which will disable vsync?

No idea, hence the questions 😀

Regards.

 

[thestryker]

I used a custom Powenetics setup previously to capture power use. It was totally independent of AMD, Intel, and Nvidia. It also looks janky AF, but it worked. The problem isn't just getting power, but getting it for every test. I had to run a separate workload with Powenetics to capture data, and basically it wasn't feasible to do that for every game. So when Nvidia made the PCAT, I tested some cards with both devices and found the results were basically the same.

The benefit is that FrameView (which uses PresentMon as a base and is free) talks to the PCAT and captures power for every single gaming benchmark I run. It's pretty awesome. People love to be suspicious about stuff like this, but really there's no indication I've seen that the FrameView or PCAT numbers are incorrect. Frankly, with Ada Nvidia pulled so far ahead on efficiency that it feels more like it just wanted to make sure everyone who wanted to work with them could get hard data showing just how efficient the 40-series was.

There's good reason for people to be suspicious of the giant companies who have repeatedly done underhanded things to get ahead in the past. This is not to say I think your experience is inaccurate or that you don't run periodic verifications to ensure things are right. Just that when it's feasible to cut out the vendor provided software (or at least non-open source since technically PresentMon is Intel's) that is the best route to take.

 

[Peksha]

No idea, hence the questions 😀

Regards.

When I tried this in cp2077 on my old 4090 about a year ago, I got exactly that - no Vsync on FG and Reflex. So it didn't help save power.

 

[BryanFRitt]

Dont benchmark it. Measure natively generated frames at correct resolution, compare and thats it. Add a side notes after conclusions about fake frames etc. Excitment about it fuels efforts to work on this crap instead of reducing power use.

Equation idea:
avg*( abs( pre-rendered** - real-time-redered ) ) , lower is better
This could even be made into a video of the differences that could be shared.

example worst cases for a sub pixel of that goes 0-255:
abs( 0 - 255 ) = abs ( -255 ) = 255, expecting black, get white
abs( 255 - 0 ) = abs ( 255 ) = 255, expecting white, get black
example best cases for a pixel of 0-255:
abs( 128 - 128 ) = abs ( 0 ) = 0, the colors exactly match

* for same picture area, like the same part of the image could be shown with 256 pixels or 64, not everything may match pixel to pixel, but can count as a fractions torwards more than one pixel.
** fully-pre-rendered-at-max-possible-quality-with-same-assets or something like this

 

[Notton]

DLSS/FSR/XeSS and FG are important features for entry level laptop GPUs and APUs.

Honestly, the only reason I don't like the 50 series is because Nvidia is pulling Nvidia shenanigans again.
DLSS 3.5/FG locked behind RTX 40 series is kind of understandable as the 20/30 series don't have the hardware for it.
But the DLSS 4.0/MFG locked behind RTX 50 series is a slap in the face when hardly anything changed between 40 and 50 series.

 

[Gaidax]

In my opinion upscaling with DLSS/FSR and XeSS are about to become the standard going forward and all these obnoxious purity checks going on in comments will fade away in a matter of few years.

As a matter of fact, I'd go as far and say reviews should give much more time of the day to the results with these techniques in use, because that would be actually much closer to real world results most gamers will experience with their games.

People turn on DLSS (or well rather, don't turn it off) and that's the reality of it, because in the end of the day more FPS and performance is more important for most people than pixel-perfect rendering.

So yeah, I'd like to see GPU review performance comparisons with most common DLSS/FSR settings used (I imagine that'd be balanced for DLSS) alongside good 'ol image quality checks, like we had back in the good 'ol days 15-20 years ago when every GPU gen came with some new rendering tech.

I think there needs to be some sort of consensus for that among the reviewers.

---

And do take note, I'm not even talking about FrameGen, that is still a tad raw. But upscaling? That's a tech we have for years already, and it is used widely, about time to move it to baseline.

 

[Mama Changa]

I can't see NPUs doing framegen as a viable approach for a variety of reasons. Basically, you have the frames in the GPU memory, and transferring those over to system RAM for the NPU to work on would be a bottleneck. Plus, NPUs are so much slower than GPUs right now on AI stuff. Like, 50 TOPS for INT8 compared to potentially thousands of TOPS. NPUs would need to be 10X~50X faster, and of course Nvidia in particular has no desire to have things run on anything other than GPU.

Just curious why you'd want a current gen NPU which can do maybe 50TOPS to do this rather than the GPU which is 15-50X faster in Blackwell. NPU's make sense in laptops with weak iGPU, not in desktops where th GPU is massively more powerful. I always laughed when Intel hyped Meteor Lake's NPU when is was slower than the iGPU. It's only benefit in a laptop is it is more efficient if performance doesn't matter so would be kinder on battery life. 5080 for example has 1800TOPS!

 

[Gaidax]

Just curious why you'd want a current gen NPU which can do maybe 50TOPS to do this rather than the GPU which is 15-50X faster in Blackwell. NPU's make sense in laptops with weak iGPU, not in desktops where th GPU is massively more powerful. I always laughed when Intel hyped Meteor Lake's NPU when is was slower than the iGPU. It's only benefit in a laptop is it is more efficient if performance doesn't matter so would be kinder on battery life. 5080 for example has 1800TOPS!

Well, one of the use cases could be Nvidia broadcast services, streaming/video image processing or who knows what else will be down the road - like AI NPCs in games and such. This is something you can use simultaneously when gaming and they do use tensor cores.

There are not crazily intensive tasks, but if at least some of them could be offloaded to CPU instead - it would result in increased performance for the main GPU tasks. And yes, probably not decoding, which is done in hardware on the GPU, but maybe some other services Nvidia Broadcast provides don't really need to be done there.

For example, there is a post where Nvidia Broadcast impact was examined on 4090 under various conditions and at times it led up to 10% FPS drops with an average of 4-5%.

Some of the features are camera input filters that by all logic go through CPU first - so they could be just processed there directly instead of weighing on the GPU.

Now it's all a speculation, but just as some things you'd rather have on GPU, there are probably also various other things that would make sense to be offloaded or done on CPU side instead.

 

[hotaru251]

Benchmark should not use upscaling by default.
Turning it on is a personal choice (usually disabled by default in games even).

When a feature (in this case dlss 4 & multi frame gen) have to be enabled by a dev 99% of content does NOT benefit from it thus shouldnt be counted in benchmarks

 

[redgarl]

Jarred, if you follow the guidelines that Nvidia is implying, then no needs to review GPU anymore because you will lose all credibility.

 

[redgarl]

By the way Jarred... why not calling upscaling by their true native resolutions for making apple to apple comparisons?

 

[redgarl]

And thus the other shoe just dropped, nVidia is providing "guidance" for it's affiliated partners, which I suspect they are required to follow under NDA or lose future access to products. This is why I'm going to wait until Steve on GN or Jay over at JTC do fully independent non-affiliated, unguided, and no-NDA testing. Steve was talking about how their team is looking to buy or borrow cards for the testing and report back to people about it.

Steve Walton made a big fuzz about the "EDITORIAL NARRATIVE" from Nvidia guidance a couple of years back... only for nowadays to follow exactly what Nvidia proposed back then for benchmarks suites. Not to mention Tim is now absolutely shilling for Nvidia in indecent manners at any occasion he speaks about DLSS. Jayz is an intel shill and his benchmarks are unreliable. And Steve from GN is always shilling for the high end even if the value is terrible.

View: https://youtu.be/wdAMcQgR92k?si=8RfFlTQEE82adr5z

 

[Notton]

Steve Walton made a big fuzz about the "EDITORIAL NARRATIVE" from Nvidia guidance a couple of years back... only for nowadays to follow exactly what Nvidia proposed back then for benchmarks suites. Not to mention Tim is now absolutely shilling for Nvidia in indecent manners at any occasion he speaks about DLSS. Jayz is an intel shill and his benchmarks are unreliable. And Steve from GN is always shilling for the high end even if the value is terrible.

View: https://youtu.be/wdAMcQgR92k?si=8RfFlTQEE82adr5z

uh, but they (HWU and GN) don't?
IDK about J2C.
DLSS3 is just plain superior to FSR3 and XeSS for image quality, if you have a card that supports it.

but sure, everything you don't agree with is a shill. You might want to get that anger checked.

 

[M0rtis]

Well, one of the use cases could be Nvidia broadcast services, streaming/video image processing or who knows what else will be down the road - like AI NPCs in games and such. This is something you can use simultaneously when gaming and they do use tensor cores.

There are not crazily intensive tasks, but if at least some of them could be offloaded to CPU instead - it would result in increased performance for the main GPU tasks. And yes, probably not decoding, which is done in hardware on the GPU, but maybe some other services Nvidia Broadcast provides don't really need to be done there.

For example, there is a post where Nvidia Broadcast impact was examined on 4090 under various conditions and at times it led up to 10% FPS drops with an average of 4-5%.

Some of the features are camera input filters that by all logic go through CPU first - so they could be just processed there directly instead of weighing on the GPU.

Now it's all a speculation, but just as some things you'd rather have on GPU, there are probably also various other things that would make sense to be offloaded or done on CPU side instead.

I have always said and felt like the next hardware addon to boost gaming should be an AI accelerator, the same way Physx cards were a thing for a bit before nVidia bought them. This was before the current definition of AI so it was along the lines of a cluster of small, low power processors and maybe some memory to boost the intelligence of NPCs in the players vicinity to give them more life like capabilities.
I guess now that NPUs will be standard on CPUs, they can be used for this instead. I dont think that they would be handling any background stuff while gaming and would be sitting idle anyway, so might as well make use of them.
Leave all the graphics based AI stuff to be handled by the graphics card.

 

[Roland Of Gilead]

DLSS/FSR/XeSS and FG are important features for entry level laptop GPUs and APUs.

Honestly, the only reason I don't like the 50 series is because Nvidia is pulling Nvidia shenanigans again.
DLSS 3.5/FG locked behind RTX 40 series is kind of understandable as the 20/30 series don't have the hardware for it.
But the DLSS 4.0/MFG locked behind RTX 50 series is a slap in the face when hardly anything changed between 40 and 50 series.

Yes, but the new transformer model brings a little upgrade for all RTX cards. The 40xx series get all except MFG. So it's not a complete FU to older gen card holders.

 

[JarredWaltonGPU]

I have always said and felt like the next hardware addon to boost gaming should be an AI accelerator, the same way Physx cards were a thing for a bit before nVidia bought them. This was before the current definition of AI so it was along the lines of a cluster of small, low power processors and maybe some memory to boost the intelligence of NPCs in the players vicinity to give them more life like capabilities.
I guess now that NPUs will be standard on CPUs, they can be used for this instead. I dont think that they would be handling any background stuff while gaming and would be sitting idle anyway, so might as well make use of them.
Leave all the graphics based AI stuff to be handled by the graphics card.

The problem is that NPUs are really just specialized hardware to do what a GPU already does. For efficiency purposes, they're better, but if you're intermingling graphics work with the task at hand, keeping everything on a GPU makes sense.

Broadcast type features should definitely be possible on an NPU, though I don't know how powerful it needs to be. Like, Broadcast runs on an RTX 2050 laptop GPU, so that's only about ... well, 41 TFLOPS FP16, but double that for INT8 TOPS. I don't think Nvidia uses INT8 for Broadcast but I could be wrong. Certainly it could be done, and it might need more like 50~100 TOPS to run in real time.

But since Nvidia doesn't make NPUs, it has no need to worry about porting Broadcast or any other features to the NPU architectures. That's on AMD, Intel, Qualcomm, Apple, etc. for now.

 

[tamalero]

Steve Walton made a big fuzz about the "EDITORIAL NARRATIVE" from Nvidia guidance a couple of years back... only for nowadays to follow exactly what Nvidia proposed back then for benchmarks suites. Not to mention Tim is now absolutely shilling for Nvidia in indecent manners at any occasion he speaks about DLSS. Jayz is an intel shill and his benchmarks are unreliable. And Steve from GN is always shilling for the high end even if the value is terrible.

View: https://youtu.be/wdAMcQgR92k?si=8RfFlTQEE82adr5z

GN and HU shilling? roflmao...

 

[M0rtis]

The problem is that NPUs are really just specialized hardware to do what a GPU already does. For efficiency purposes, they're better, but if you're intermingling graphics work with the task at hand, keeping everything on a GPU makes sense.

Broadcast type features should definitely be possible on an NPU, though I don't know how powerful it needs to be. Like, Broadcast runs on an RTX 2050 laptop GPU, so that's only about ... well, 41 TFLOPS FP16, but double that for INT8 TOPS. I don't think Nvidia uses INT8 for Broadcast but I could be wrong. Certainly it could be done, and it might need more like 50~100 TOPS to run in real time.

But since Nvidia doesn't make NPUs, it has no need to worry about porting Broadcast or any other features to the NPU architectures. That's on AMD, Intel, Qualcomm, Apple, etc. for now.

I was thinking more along the lines of using the NPU units on the processor for NPC/Enemy AI advanced behavior, you know the actual place where gaming related "AI" makes sense. I also said that the graphics related "AI" can all be handled by the graphics card itself.