Outside of perhaps rendering input lag figures for specific games, I dont know what other features you think are being neglected so egregiously, or what 'popular usage' aspects of gamers are being ignored.
Digital Foundry Article Technical Discussion [2025]
- Thread starter Shifty Geezer
- Start date
Yes, it will be interessting to see if nVidia use a generated frame as base for one or two frames or they are offsetting these from the previous and current "real" one.
I'm still curious what the definition of a "real" frame actually is. Must the host system be involved? But isnt this something we do not want? CPU limitation is a huge problem... Or must a "real" frame involve new input information? What happens to engine like UE5 which do not couple input lantency to render time? In Black Myth capped FPS to 60 i can have 27ms+ (4K DLSS Performance) or 67ms+ (nativ 4K) input lantency. Even with DLSS Performance UE5 is producing frames with the same input information. Can we call these "useless" frames "fake", too?
How many reviews have you seen cover overlays and filters which are hugely popular in actual usage?
The common definition of a frame includes advancing the simulation so yes the host is involved.
Saying that using filters is 'hugely popular' is probably an overstatement of the century. I'd be surprised if more than 5% of gamers actually use them on a rare basis, let alone on a regular basis.
And what kind of overlays are you talking about?
I dont really see these things as being important to new product releases, though. By this reasoning, there's a thousand things that could be tested given the incredible amount of things computers can do. Reviewers only have so much time and even as is, reviews tend to be lengthy and reasonably thorough, with more niche or detail aspects or simply extraneous 'interesting things to test' tend to be done in follow up videos and whatnot, because they're usually just not as pertinent for what the vast majority need to know or are most interested in.
Basically, you're asking a LOT and maybe not appreciating how much work goes into reviewing these things in the first place. It's not like a monitor, which basically has fixed characteristics that can be tested, regardless of whatever content you're feeding onto it. That makes things a lot more straightforward.
And I think you’re wrong. Guess we’ll never know. All those people using reshade, sweetfx etc and the recent complaints about the nvidia filter performance hit must be irrelevant to you.
It is funny that you think 5% of the population is a minority though. You must also think reviews of the 5090 are a waste of everyone’s time
Yes I’m asking for more than doing the absolute minimum which is what most reviewers do. Testing a monitor is a lot more complicated than running a benchmark at 3 resolutions and sticking a number in a spreadsheet. The tools and tests are way more involved with tons more variables. Not sure if you’re serious.
It's funny that I think 5% of the population is a 'minority'? Maybe we grew up with different dictionaries, I dont know.And I think you’re wrong. Guess we’ll never know. All those people using reshade, sweetfx etc and the recent complaints about the nvidia filter performance hit must be irrelevant to you.
It is funny that you think 5% of the population is a minority though. You must also think reviews of the 5090 are a waste of everyone’s time
Yes I’m asking for more than doing the absolute minimum which is what most reviewers do. Testing a monitor is a lot more complicated than running a benchmark at 3 resolutions and sticking a number in a spreadsheet. The tools and tests are way more involved with tons more variables. Not sure if you’re serious.
And they cant just test filters, because they're all different and can be configured in countless different ways that will all affect the demand in different ways. Makes no sense. Reshade isn't just 'one thing'.
Might as well just ask reviewers to bench every game in existence in every single level with every single different setting combination and anything less is insufficient.
And again, reviewers are not doing the 'bare minimum'. It's an absolute ton of effort to do these GPU reviews, and you are proving my point about being unappreciative of what they actually do.
The point about monitors is that it is straightforward in terms of coming to a conclusion about the characteristics, which are essentially fixed. You test for those fixed characteristics and that's it, you're done, you've essentially got a definitive 'picture' of what the display can do. Processors are nothing like that, as they can react and perform very differently depending on what software you're asking it to run. Display testing also has the advantage that most of what you run on it is also fixed, unlike software, which can be constantly in flux. So with displays, you can reuse data from previous reviews reliably to compare against, whereas with processors you need to be constantly doing NEW bench runs with all the different hardware you're trying to compare when doing new reviews. It's just a ton more work.
Last edited:
Minority is literally less than 50%, the smaller of two portions.
You've probably been influenced by modern discourse to relate 'minority' to very small proportions in terms of oppressed minorities
It doesn't matter what proportion of people are interested in something so long as it's a large enough number to be worth the effort. 5090 reviews will almost certainly be read by far more than just the potential buyers of said card, so reviews of that GPU are still worth doing even if a small minority of gamers ever own one.You must also think reviews of the 5090 are a waste of everyone’s time
What's the effort needed to create a PC of a configuration installed and set up to benchmark a game, and then boot and load and play a game to capture the video, and then record the findings and edit the video? Genuinely I don't know. Can you just keep swapping GPUs and CPU and RAM in and out without messing the OS up, or do you need to do some installation stuff between? I'm not sure it's that much easier than benchmarking displays, and the reason you get more detailed display benchmarks is because it's easier so they can accomplish more investigations in the same timeframe.
Whatever, the point is picking a workload that is economically viable. There's a lot more data we'd like to see, but if that ends up being very niche and doesn't drive visits etc., it's wasted effort. So your argument has to be that putting in more work will net more gains. You need to present the size of the audience that uses filters and why it's in DF's interests to add that workload on top of their existing benchmarking workload in terms of what they'd gain.
But FG is advancing the simulation just with one frame in the past. Digitalfoundry released a video comparing FG and normal rendering in Spider-Man two years ago:
I meant irrelevant minority. His point is that insufficient people use the feature to be worth a reviewer's time.
Exactly.
Most GPU reviewers don't retest all configurations for each video or article. The effort is spread over many weeks/months. Benchmarking GPUs is certainly more tedious, but I wouldn't say easier. The concepts involved in evaluating display quality are a bit more technical than running a benchmark and writing down the number.
Gains as in more clicks and more ad revenue for DF? I don't know anything about that. I'm looking at it from the perspective of how complete and useful is the actual "product review".
If you guys are happy with FPS graphs that's great but that's not the only thing that impacts the experience of actually owning a graphics card. It's not like reviewers need to cover every feature in every SKU review either. It could be done "per generation" or every few years etc. As it stands today there's nary an investigation into GPU user experience outside of maybe upscaling IQ comparisions. It's a low bar.
FG uses two frames in the past.
It's an interesting question though. If we've already accepted that the images we see on our monitors are in the "past" relative to the CPU simulation time then FG is technically interpolating "simulation steps" as well as "image updates". The hangup with FG is all about latency and reaction time.
Let's say the CPU decides that an object moved x meters in 1 time step (e.g 16.6ms at 60fps). FG is essentially guessing how far it moved in each slice of that interval - say x/2 meters every 8.4ms for 2x framegen. So for passive elements of the simulation FG is essentially mimicking more frequent CPU updates. That's a win.
The main difference is that FG doesn't allow the user to influence the outcome of the full 16.6ms action because that's already in the past from the perspective of the CPU. So if you're a super human gamer with cat like reflexes that could be a problem. For most people it's probably fine though.
FG doesn't do any simulation or extrapolation. It just interpolates between two rendered frames (current one, and previous). It delays the presentation of the most recently produced rendered frame until it has inserted the interpolated one.
Indeed, but I wouldn't say it is "mimicking CPU updates". It is literally doing image-space motion smoothing, with some additional data from the engine to reduce visual artifacts from disocclusion. It really is quite comparable to what TVs do, albeit higher quality due to the additional info (and compute). If the time deltas are small and thus the disocclusion or non-linear animation is minimal, this can work pretty well. Obviously it does not scale down though which is why it should always be discussed more as a high end "this is already very good but we can make it even smoother" feature than something that can be used to bring performance from unacceptable to acceptable.
Last edited:
Things need to be reliably measurable and objective, or else they aren't worth much. It's very hard to review the 'general GPU user experience' as a whole because PC's are by nature ultra configurable things that we all use in endless different ways from each other.
What you're asking is straight up just too much of reviewers. You're treating them like they're lazy and doing the bare minimum when in reality they're spending like 16+ hour days trying to get all the most pertinent info to people so they have good information.
I get you want more 'investigations' into more niche use aspects, but that's also not easy to do. Like you say you think 'filters' are some important and neglected thing reviewers aren't testing. What kind of testing are you even envisioning here? How do you 'test' filters specifically? What are you looking for? Do you want them to benchmark how GPU's perform using injected post-processing elements? I could see that being somewhat interesting, but a very weird and unique situation that is also complicated because how do you separate that from the underlying differences in game performance? You could do the testing with multiple games all with all the different sorts of settings, but again, we're piling on the work here.
And really, you're essentially asking them to test mods here. And that's a whole black hole of options and complications and whatnot. It's not that the information wouldn't be worthwhile at all, but the opportunity cost for all this stuff has to be considered for the people doing this sort of content. You guys are always acutely forthcoming about the 'business realities' of major corporate practices from Nvidia and whatnot, but cant make the same sympathies towards ordinary folks like reviewers, who are simply working hard trying to give consumers useful information?
That’s certainly not true when it comes to product reviews. In many cases the info shared is a subjective evaluation.
I disagree. DF is a standout in terms of evaluating frame consistency, stuttering, image quality etc. Most reviewers don’t go that far. You get avg and maybe 1% fps that’s it. They’re currently doing the least amount of work that would qualify as a review.
To bring this back to the original point. If reviews report only “raw” fps (which I agree they should) then they must either find some way to communicate the other capabilities of these cards or choose to ignore them. Those are the options. They have a very poor record of doing the former.
In terms of Digital Foundry, here's what I'd like to see explored in Frame Generation on a product line like the 50 series:
1) How does frame gen scale with tensor performance? With DLSS 3 frame gen we rarely saw actual doubling of performance. It seemed to range somewhere between 1.5 to 2x, but usually not (if ever) 2x. Do the lower end models like 5060 perform as well at frame gen as the top of the stack with 5090? I'm assuming this is a case of pushing the gpu to high utilization before turning frame gen on, which is another reason why frame limiting with frame gen was ideal.
2) How viable is scaling at the top end of the refresh range? Almost any comparison I've seen is kind of max-settings ray-tracing frame gen to get from 60-120. I don't think I've ever seen someone test 180->360 or 120->240 with quantitative and qualitative comparisons. How much do artifacts reduce when you increase base frame rate? Are there any bottlenecks that get hit with frame gen where it stops being viable beyond a particular starting frame rate? Can DLSS 3 FG even push frames at 360Hz? 240 Hz monitors have been out for maybe 6 or 7 years, though incredibly rare, but gpu reviewers still kind of treat them like they don't exist. We're at 480Hz now, and frame gen from 120 to 480 is an interesting use case.
3) input latency is something they're obviously going to explore in detail. Not sure I have any real feedback here, other than it just being an important consideration.
In terms of latency discussions, it's still very poorly understood by a lot, if not the vast majority of gamers. They don't understand that games have latency. Some people think if they're gaming at 60 fps that they have 16ms of latency, so they don't understand two games can vary significantly. They don't understand what Reflex does, in fact you'll see people recommending people turn it off because it limits frame rate. They'll show how it drops their fps from 400 to 350, and say that's making the game run worse. Ultimately gamers are a victim of dunning kruger. They don't really know what's going on, but they have high confidence that they do because they can see numbers like fps, latency and resolution and they latch onto them like vampires. It's one of the reasons a channel like Digital Foundry is so valuable, because they do a lot of education in an easily digestible way. I always bring him up, but battlenonsense was a legend, and I wish his channel was still making videos.
1) How does frame gen scale with tensor performance? With DLSS 3 frame gen we rarely saw actual doubling of performance. It seemed to range somewhere between 1.5 to 2x, but usually not (if ever) 2x. Do the lower end models like 5060 perform as well at frame gen as the top of the stack with 5090? I'm assuming this is a case of pushing the gpu to high utilization before turning frame gen on, which is another reason why frame limiting with frame gen was ideal.
2) How viable is scaling at the top end of the refresh range? Almost any comparison I've seen is kind of max-settings ray-tracing frame gen to get from 60-120. I don't think I've ever seen someone test 180->360 or 120->240 with quantitative and qualitative comparisons. How much do artifacts reduce when you increase base frame rate? Are there any bottlenecks that get hit with frame gen where it stops being viable beyond a particular starting frame rate? Can DLSS 3 FG even push frames at 360Hz? 240 Hz monitors have been out for maybe 6 or 7 years, though incredibly rare, but gpu reviewers still kind of treat them like they don't exist. We're at 480Hz now, and frame gen from 120 to 480 is an interesting use case.
3) input latency is something they're obviously going to explore in detail. Not sure I have any real feedback here, other than it just being an important consideration.
In terms of latency discussions, it's still very poorly understood by a lot, if not the vast majority of gamers. They don't understand that games have latency. Some people think if they're gaming at 60 fps that they have 16ms of latency, so they don't understand two games can vary significantly. They don't understand what Reflex does, in fact you'll see people recommending people turn it off because it limits frame rate. They'll show how it drops their fps from 400 to 350, and say that's making the game run worse. Ultimately gamers are a victim of dunning kruger. They don't really know what's going on, but they have high confidence that they do because they can see numbers like fps, latency and resolution and they latch onto them like vampires. It's one of the reasons a channel like Digital Foundry is so valuable, because they do a lot of education in an easily digestible way. I always bring him up, but battlenonsense was a legend, and I wish his channel was still making videos.
Last edited:
FG also helps solve a bit of the chicken n egg dilemma when it comes to high refresh monitors.
This year samsung will have 500hz QD OLED monitors which will be sublime for motion clarity. Without FG, in anything but esports titles looking like PS2 era quality, you got no chance at that.
At the end of the day, IQ means nothing if it disappears the moment your mouse moves. That's why I'm a big advocate for oled based high refresh monitors. Yes I understand the burn in but I'm happy with the trade off.
This year samsung will have 500hz QD OLED monitors which will be sublime for motion clarity. Without FG, in anything but esports titles looking like PS2 era quality, you got no chance at that.
At the end of the day, IQ means nothing if it disappears the moment your mouse moves. That's why I'm a big advocate for oled based high refresh monitors. Yes I understand the burn in but I'm happy with the trade off.
Last edited:
Good point. It really only works well for linear movement so your time steps have to be pretty small to begin with.
So if we want to define a real frame as “CPU has decided how the world has changed since the last frame” then frame gen is not producing real frames.
More particularly a real frame is a draw of the complete game state - this can happen on any processor. World updates themselves aren't frames as you can have multiple game updates per drawn frame, particularly if iterating a physics model.
A complete, perfect frame will record per pixel the game state at the time slice that frame was begun to be assembled into an image, assuming the game state can't be in flux at the time of drawing as the drawing process is dependent on geometry remaining fixed during the drawing process.
We can start abstracting some interesting concepts on quality and state and 'pixels' from that notion of a complete, perfect frame, and start evaluating interpolated data in terms of deviation from ground truth. Low res buffers are a deviation from ground truth. Interpolated frames are a deviation. Upscaled images are a deviation.
Yeah we should probably treat CPU simulation as an analog process for the most part with rendered frames being discrete snapshots of that continuum.
That would really get into the weeds! Did the CPU update the wind simulation blowing the trees using the same time instant it used to update the ballistic trajectory of the bullets I just fired? We’ll never know.
Similar threads
- Locked
