Awesome engineering discussion with Guillermo. He definitely joins the list of some of the most educational interviews we've hosted. Looking forward to exploring this topic more! Find another engineering interview here, featuring Tom Petersen of Intel: th-cam.com/video/5hAy5V91Hr4/w-d-xo.html Or our interview with two engineers from AMD (Bill & Amit) over here: th-cam.com/video/RTA3Ls-WAcw/w-d-xo.html Find our engineering interview playlist here: th-cam.com/video/hVSSOs9Z-uY/w-d-xo.html
If I may make suggestions... Would you guys mind, since you are going to add latency to your tests, demystifying some long held beliefs that we (in the know) know to be incorrect or correct but still see the mainstream opinion not catching up entirely? A good example, as Guillermo was saying, how DPI is also important for input latency. Of course even if it doubles input latency from 0.8ms to 1.6 it is not like it matters, but it is too often people say I'm "wrong" in playing at a high DPI (from the sensor, with VERY low sensitivity in game or in my OS [pointer speed on windows would be the equivalent]), but we all know it is 'measurably' better, even if undetectable. Then also the good old polling hz. Like I said in some other comment, comparing amd anti-lag, nvidia reflex and just a good old dumb frame limiter to right below where you'd be GPU bound, like Battle(non)sense suggests.. IDK, if I were to give some thought instead of commenting on youtube I could think of more, but these are just examples off the top
thanks for doing this, its always nice to put a face and personality to the tech we use. reflex in particular has been really good for me so far and it was cool to visualize how both the process it sought to improve and it itself works
@@minecraftmike5193 because developers who implement it needs to optimized it for their game. I just gonna go out on a limb that the Rachet and Clank devs didn't care enough to optimized Reflex for the game when Reflex is mainly used on Competitive games not single-player games..
One clarification to the answer to Steve's question at around 22:30. The input sampling is a "uniform random" variable with an average time of 5 ms in that example. This is because as Guillermo explained earlier, the input could have arrived anywhere from 0 to 10 ms before the frame, so the sampling causes an average of 5 ms of latency, though any particular frame has a roughly equal probability of being any value from 0 to 10 ms.
@GamersNexus The best analogy is video camera. As the shutter opens, information starts to accumulate, but at some point shutter closes, and everything that has been collected up to this point in time gets saved as a frame. Then shutter opens again and the cycle repeats. And since a click can occur at any point of that cycle with equal probability, statistical mean value is going to be half of the cycle time.
I would like to point out that in case of the cpu load is low enough, you could get away with something like generate the new frame asap, poll continuously for inputs and if you get new inputs generate the new frame and throw away your old frame sitting in the render buffer. This removes the risk of losing frames you'd get by waiting, but will hit your cpu harder.
@@griffin1366 Battlefield 1 was a diamond in the rough. It runs so good these days. It's optimised to high heaven. Bigger maps, more players, same FPS as CS2 LOL.
I’m a consultant in the advanced technology space and have to give technical “talks” for a living - this was great. It’s really hard to explain technical complexity in a way that is accessible; it’s all abstraction of language we are dealing with so things get confusing quickly. This audience is likely much more technical than most, however thats still very cool something this in depth and a bit esoteric has 200k views. Solid work. Looking forward to Pt 2!
i often end up explaining how pcs work to some of my friends and this video has given me a new perspective on how to explain the cpu and gpu in a very simple fassion, the cpu draws everything and the gpu colours it xDDD
Thank you for looking at this. Latency is definitely the most underrated metric when measuring system performance right now. I purposefully optimize my graphics settings to get the lowest amount of input lag possible, rather than going for more FPS.
You nailed it - very underrated, and we've been a part of the FPS focus (even with frametimes) and want to expand the scope! A lot of times, the two align and are correlated. But there is definitely a lot more to the story!
Same. Limiting FPS for low latency and consistent frame times almost always leads to a significantly better experience than running the GPU at 100% for peak FPS.
Yeah, take Frame Generation for instance - it's being marketed as a performance boost (which I absolutely hate), because the FPS values are higher (bigger number bigger better), but the latency is in some cases tripled. It should be advertised as motion smoothing technology or something
@TK-11 yeah at least in val at 900 fps uncapped it's saying my game to render latency is about 3ms while capping it to 500fps reports roughly 2ms. Interesting seeing im using reflex which is meant to eliminate the latency overhead u just mentioned.
I love listening to people who love and care about what they do talk about their work, especially highly technical things being explained in a more accessible way. Great stuff from the both of you, this video was really cool.
Fantastic video, the guy from Nvidia was an absolute pleasure to listen to and very professional. I now understand why when I run with reflex on my game runs so much smoother. Look forward to part 2. Thanks Steve and team
Fantastic coverage on this underrated metric. It felt very enjoyable to watch and not at all 'lecture-esque' as you thought it would be! Thanks for the great video and awesome content :D
So basically I never want to be gpu bound or what? Even if it's like a 4090 + 14900k in that case i'm still gonna be gpu bound right. Or maybe it doesnt matter at high fps say240+ fps?
I'm really glad that latency is getting so much focus these days honestly. For a long time i feel like it's not really been talked about much, despite being so critical for the user experience
@@GamersNexus Awesome video, primarily playing FPS games lately and adding this information to reviews will allow me to purchase the products that will be more attuned to my gaming goals rather then just 'more fps = better'.
@@Reznap7 I game between 60Hz and 85Hz on a CRT monitor (i could push it beyond 85Hz but a CRT is so fast that higher refresh rates than those are kinda pointless, it already creates natural motion blur at 85Hz). We are on this absurd bubble of Hz and fps for gaming, it's like the loudness wars for music. Finally, since VGA CRTs are analog, not digital, they don't add lag (you don't need anti-aliasing, AI upscaling and frame gen since CRTs don't have pixels or a fixed native spatial and temporal resolution).
Latency is not just a gaming thing! It's of utmost importance for professional audio and video recording too, so not only a graphics thing. Imagine recording like 15 microphones to separate tracks simultaneously, with EQ and many other effects on each, and the whole, and slamming it on a drive, as close to real time to be able to monitor it all! Same goes for recording scenes with many cameras, all at once, or even editing live between them. Please do more on the subject Steve, and also give it more love from a processor/RAM... perspective.
and to circle back, audio latency can also be a huge gaming thing. Certain audio cues are used in place of visual cues in cases which they are faster. For example, it has been studied that the top LG (lightning gun) players in Quake have reaction times that exceed what is possible from visual input. One theory is these players use the audio cue of the hit indicator to calculate when the enemy player changes direction before they actually see it. Human reaction to audio stimuli is also slightly faster than visual stimuli
@@chasmurphy1227 Audio latency should not exist in 2024, that was mostly eradicated when quad cores became mainstream. Even 32 bit 192khz should still run perfectly fine. If your CPU is running at 100% then maybe there's a difference, but firstly you shouldn't be letting your CPU hit 100% as that caused extreme instability on your PC, and secondly most games won't utilize every thread on 16 thread and even some 12 thread CPUs. Audio is extremely important, but I don't see audio latency/lag being an issue still.
My goodness, this was an amazing compilation and summary around everything to do with input latency and render latency. Glad you are expounding on this topic more since FPS numbers tell so little about how the game feels to play. Thanks for doing this!
Really excited to see you guys are going to add latency to your testing. This is awesome news and something that hasn't gotten much attention elsewhere.
I respect your engineering and testing methodology. No one else takes the time to look at it with this attention to detail and consistency. In addition, you take the time to explain it so I know why the numbers matter and how you derived them. Spewing out a ton of data can quickly get boring and become meaningless, but you manage to keep the detail while making it relevant and understandable. That's not easy.
Glad that you're starting to look into this. Though this is MUCH larger than just for graphics. I.e. a very good example is storage (nvme/ssd/hdd); memory access; as well as general interrupt handling delays. The biggest item I run into pretty much daily is data subsystem performance (drive I/O basically). This is made worse as basically OEM's and all reviewers seem to be fascinated by 'streaming throughput' to a subsystem when this workload basically is completely non-existent for any real world workload. (anyone who doubts this, load up diskmon and check for yourself, it doesn't exist). Wish you can add similar level of detail to any reviews you do with these other subsystems.
100% agree but if u care about storage io latency so much those optane drives have been on flash sale recently and in case it wasn't obvious, they demolish any normal ssd in io latency both in averages and especially consistency.
Right, though for *average* latency you can't measure things by the weakest link in the chain (e.g. the SSD), because there are a bunch of layers of cache in between that prevent you from being bottlenecked by it. SSD -> RAM -> L3 -> L2 -> L1 -> CPU would be the rough chain for CPU work on a modern PC (SSDs can also have RAM, but it doesn't quite fit in with the other components in the chain which go from big to small, slow to fast). Things get slow if you miss every cache and have to ask the SSD directly, but in a lot of workloads everything can be loaded in advance or consistently streamed in before it's needed, and you never hit that slowest path at a critical moment. In practice there *are* going to be misses, which is one source of frame time spikes, but if things are working correctly these are low probability events (but of course it depends on how much data the workload churns through).
@EmanuelHoogeveen Yeah true but remember to a cpu each miss is super detrimental when it's running at like 0.2 nano seconds per clock while a good ssd has an access latency of 50 microseconds, ie 250,000 times that, so a lot of wasted cpu cycles whenever it is waiting on data from the ssd. That's why hard drives had been such an enormous bottleneck for so long. To a cpu each miss as technically super rare as it is (proportionally i mean, it happens pretty frequently in absolute terms, check resource monitor and look for "page hard faults") is detrimental when they're that fast. Oh yeah and literally almost any time u load something ur almost definitely gonna get tons of misses. That's basically what loading something is, moving data from storage into ram. Oh and fyi ssds have inherently really bad writes which is why dedicated dram cache on them is literally only used for writes and never reads so for latency it isn't really much help as generally, ur always read bound not write bound. TLDR: No one's assuming ur cpu is pulling data from the ssd 100% of the time, the point is it it affects ur system latency regardless.
@@FrozokenI didn't realize that SSD RAM is mostly for writes but that makes sense, thanks :) I was trying to figure out where it fits in since it's so small relative to the SSD and even your regular RAM, but if it's used as a buffer more than cache that makes more sense.
I have always and continue to appreciate GN's fine line-walking act between presenting the technologies companies implement to make things better... and an ad. This video is great! It doesn't feel like an ad for Nvidia, it feels like education on why Nvidia is a leader in competitive gaming and what technology they have invented to make sure the human component of competitive gaming is the one that matters most! It also helps explain why mysterious settings like Nvidia Reflex might show up in games you play!
This feels like I'm in a lecture and I should be taking notes! Really interesting video; he's clearly very dedicated to what he does. I think more gamers should watch stuff like this, get a better understanding of what's going on closer to the hardware level and gain some appreciation to everyone making this stuff possible.
SUPER useful piece, GN! I've admittedly downplayed response times in my choices of display thinking it wasn't really going to matter to me. This gives me a much clearer picture of what's actually going on and will allow me to make better decisions in the future. Thanks!
This is great content. I love understanding the graphics pipeline. Very clear presentation. Shoot more videos like this, Steve! I enjoyed the TAP VR pipeline presentation years ago. I still go back to it from time to time.
Thats pretty cool. I was still playing consoles when people were playing that game. I the whole csgo generation is a really important chunk out of the overall history of gaming. I always imagine what video games either traditional or VR will be like in a few hundred years if we don't kill the planet with nuclear war and find a way to coexist regardless of borders.
This educational deeper dive into the he why and how is invaluable to consumers. Most of us care about the entire experience of gaming not just fps. Looking forward to part 2!
Latency used to be something often tested. And then it suddenly went away. I remember a common one was the test with a mouse click and led to when the gun muzzle flash. Or whatever was supposed to be triggered. And there was more and more focused on frame. To the point we are now counting fake frames. How is that different to a time when we had games making 30 frames but displaying 60 by duplicating them. And the test was usually done for displays reviews only. Man I loved Guillermo. Not just the all the info but how well he presented and explained it. Also how he kept moving in place. XD I'm always amazed how host like Steve can manage to stay still for so long. It was nice see someone like me that can stay still for 3 seconds.
This was a fantastic and very educational video! Thank you for always putting in the work to not just give us information but to teach us how understand it also. This was an awesome video!
Amazing stuff! I love that you picked Counter-Strike as an example and I'm looking forward to part 2. After playing CS2 the whole evening with Reflex off, it might be the time to finally give it another shot!
Mentioning Battle(non)sense was so nice to see! Love the channel and I actually hope he can monetize his content better. It's a real shame he can't do more of his stuff!
Awesome job, learned so much, thank you! Like I said in the comments of a previous video, "input lag is more important than FPS, frame time, frame spacing, etc." is the hill I'm ready to die on.
God I'm so happy you're doing this. The speedrunning community has been on this train for a while now; we play old games that are often locked as low as 30fps, so we've needed other means to improve the game feel and quickly realized how much more important latency actually is. But it's been so hard to find coverage on latency as a buyer, which it looks like you guys are striving to fix. Looking forward to part 2!
this kind of testing is really important, recently i have been trying out frame generation in a bunch of games thanks to Lossless Scaling's implementation, and every single time i have decided against frame gen due to input latency. Frame gen from 85 FPS to 170 FPS feels so much worse than 90-130 real FPS with equivalent settings. EDIT: Lossless Scaling just updated to not require a frame cap when generating frames, meaning I can now play at an unlocked 90-130 FPS with frame gen. After 5 minutes of getting a feel for this, I think it is AMAZING! Maybe I'll be noticing the input latency after trying this for longer but for now this is definitely how I'm playing my games.
It really depends on the game though, and its implementation. I've had frame generation probably add 10-15ms input latency (you can do a test that shows how sensitive or accurate you are when it comes to telling how much input latency you're getting and FG felt like 10-20 in some games I've tested it with). But with other games, it easily adds 30-50ms. For context, I have a 4080 super
BRILLIANT! No wonder your channel is second to non! What blows my mind, is how freakin fast this process is repeated every second per frame! incredible!
Doesn't matter how fast you can get data from your hand to the computer; it's still going to sit in a cache until the CPU until it's ready to be processed. This is why polling will always be inferior to interrupt, which tells the computer to "drop everything you're doing and process this data". PS/2 purists aren't just being luddites.
@@MyNameIsBucket This was correct but modern chipsets put the PS/2 interface through the same caches as polled devices. Not only that, the PS/2 protocol is so slow it takes around a millisecond to transmit a single scancode, so one or multiple milliseconds per pressed key. A 1000Hz USB keyboard refreshes its entire rollover in the time it takes a PS/2 keyboard to signal a single key.
@@mitsync Unless your CPU is very busy completing other tasks like, oh I don't know, trying to render a CPU-bound game at 120Hz. Then the input from that USB keyboard takes quite a bit longer to register in the game. The PS/2 keyboard will still take exactly the same time.
@@MyNameIsBucket You are prioritising input over simulation speed. The game isn't going to do anything with your input until the next tick anyway, and very very few CPU bound games run over 60 tick. 1000hz polling is already well past the point of diminishing returns, you aren't gaining anything by trying to squeeze your inputs in quicker than that.
I absolutely love this. Seeing numbers on charts can help get to the idea of what's doing what in your system, but having an engineer break down exactly how the pieces work together, and what their tech is doing to optimize it, that I love. One of the best videos I've seen in awhile.
This is really helpful and interesting stuff. This vid helps tremendously in understanding what really happens when we're running into performance limitations. Thanks for this!
Steve working to force all of tech youtube to get better on their game again. First it was pushing the moves away from FPS, now its in coming up with entirely new understandings for how to demonstrate user experience. Great work as always!
Finally!!! I'm really looking forward to this journey. I'd love to also consider adding some Linux tests now and then, since Nvidia Reflex with latency markers will be added soon to Proton.
This was VERY interesting and quite enlightening. It was also explained in a way that i think even people without much theoretical knowledge can understand .
Loved this! This kind of in-depth explanation of complete system latency is Amazing! I thoroughly enjoyed this video... and have enjoyed all of your in depth & complex dives into how the tech is actually functioning! I'm Yearning for more like it! :)
This is literally gonna be the next step in hardware reviews. Hardware has become so good only unoptimized AAA titles and I guess casual games with good graphics need more FPS, but a ton of people who play competitively already get more than enough FPS. What really matters is whether the frame you see is even the one you're supposed to see and if the input you made gets processed fast enough that it's accurate. There's a lot of aspects to it and FPS can only do so much in that regard. I hope other tech reviewers start looking into it as well, it looks hella interesting.
For this to work it'd require measuring latency in complex states/environments. Just firing a gun without anything else is informational but it could be rather misleading. It's difficult to be the "next step" as it'd require to be somewhat standardized/repeatable as well.
These are the type of metrics that probably matter most. It doesn’t matter if you have the fastest processor and GPU if you are bottlenecked somewhere. Thanks for starting to add these!
Yep. We need someone to create a tool like PCPartPicker but with product-specific latency listed that lets you choose all your hardware and estimates your total system latency and breaks it all down, showing you where your bottleneck is and if there's any weird interactions between specific hardware, etc.
Been waiting on a video like this! If anyone is interested in knowing more about polling (specifically on old consoles), there is a video called "You will never see a Bomb Torizo skip in a World Record Super Metroid speedrun (NTSC)" which talks about issues emerging at polling rates around 60Hz. It's very focused on Super Metroid, but I found it very interesting.
This is awesome, I like these deep dives, made me realize that all of us take for granted to a large degree, the engineering behind playing a video game and all that goes into making it a great time.
This is fascinating and exactly the kind of content I crave for being an FPS nerd. I would point out that if you're going to go focus on latency in competitive games, optimizing settings for performance would help quite a bit. For instance, your R6S testing is done on DX11 while the game run significantly better on Vulkan (at least for relatively recent hardware). It would be nice to see these tests done on minimum setting and at 1080p (and, maybe, 1440p), since that's how people who really care about performance tend to play them.
What an amazingly informative video, thanks so much I learned a lot. And thanks to buddy from nvidia for helping us understand latency from a more technical perspective. Great video and why I love your videos!
thanks for talking about this as a fighting game player it's my primary concern. I would argue in this realm it's even more important than in competitive shooters.
Great video! Input lag is something that is commonly addressed in the retro gaming community, but most of the discussions center around the display, and namely the benefits of using a CRT to eliminate the display side of lag. People never seem to care that certain ports/games can have more or less lag inherent to the game software or hardware itself. One interesting solution that is viable for retro games is frame look ahead, where the emulator/game engine is running multiple instances of the game to predict inputs and thereby reduce input latency on the game software side. Obviously this is mainly possible due to the vast difference in hardware power that allows us to simulate many simultaneous versions of these old games, but I do wonder if we'll ever see something like that come to modern games to reduce latency. Maybe in some ideal world our hardware power continues to increase, but game graphics/overhead stays similar allowing us to run multiple instances. Highly unlikely, but a fun thought.
I have a lot of fun playing Fall Guys, but then i noticed that the game is locked to 60fps. Some games simply don't have unlocked framerates. Depending on the game, 60Hz can look a bit flickery on a CRT.
@@saricubra2867According to the PCGamingWiki for Fall Guys, the game is not capped at 60 Hz: "Since Season 5, you can freely adjust the FPS limit in-game when VSync is turned off."
Awesome! I've been wanting a deep dive into the importance of latency. It seems that 4k gamers are missing out on a lot of responsiveness, so the new oled monitors that can switch between 240hz 4k and 480hz 1080p should be a godsend. Now I'm just waiting for a reflex oled monitor at 1080p . . . one can dream.
this is the type of detail test i love to see. because for serious gamer and specially competitive gamer - this is meaningful input latency to human reaction time to keyboard press ( some keyboard has higher height before it execute) all adds up to delay which can ruin your game play.
Gamers do latency analysis all the time, but we do it by the seat of our pants rather than by instrumentation. When you do analysis by the seat of your pants, it all comes down to how it _feels_ to play the game. And yes, it matters a lot.
@@redmoogle I'm (somewhat unscientifically) convinced that most of the issues with jittery game play and latency are being caused by the fifty gajillion processes running on any Windows and the effort to run games at high levels on hardware that's not capable of doing it.
Awesome engineering discussion with Guillermo. He definitely joins the list of some of the most educational interviews we've hosted. Looking forward to exploring this topic more!
Find another engineering interview here, featuring Tom Petersen of Intel: th-cam.com/video/5hAy5V91Hr4/w-d-xo.html
Or our interview with two engineers from AMD (Bill & Amit) over here: th-cam.com/video/RTA3Ls-WAcw/w-d-xo.html
Find our engineering interview playlist here: th-cam.com/video/hVSSOs9Z-uY/w-d-xo.html
Notice me senpai
Very interesting!
With hard wired don't forget the 'on time' latency of the diode/L.E.D. lol Screen sensor Latency + 'bullet' flight time LOL
If I may make suggestions... Would you guys mind, since you are going to add latency to your tests, demystifying some long held beliefs that we (in the know) know to be incorrect or correct but still see the mainstream opinion not catching up entirely?
A good example, as Guillermo was saying, how DPI is also important for input latency. Of course even if it doubles input latency from 0.8ms to 1.6 it is not like it matters, but it is too often people say I'm "wrong" in playing at a high DPI (from the sensor, with VERY low sensitivity in game or in my OS [pointer speed on windows would be the equivalent]), but we all know it is 'measurably' better, even if undetectable.
Then also the good old polling hz.
Like I said in some other comment, comparing amd anti-lag, nvidia reflex and just a good old dumb frame limiter to right below where you'd be GPU bound, like Battle(non)sense suggests..
IDK, if I were to give some thought instead of commenting on youtube I could think of more, but these are just examples off the top
great insite ty. (i know i cant spwll)
Thanks so much for having me on! Great to meet you and the team-that was loads of fun. Looking forward to Pt. 2 with n0thing!
Thank you for the clear explanation, it was really easy to follow!
Love your passion and dedication towards the topic
thanks for doing this, its always nice to put a face and personality to the tech we use. reflex in particular has been really good for me so far and it was cool to visualize how both the process it sought to improve and it itself works
@@minecraftmike5193I’ve not ever had an issue in any of my games. Might be an implementation issue in the game? Not sure just guessing.
Very informative and well made presentation!
@@minecraftmike5193 because developers who implement it needs to optimized it for their game. I just gonna go out on a limb that the Rachet and Clank devs didn't care enough to optimized Reflex for the game when Reflex is mainly used on Competitive games not single-player games..
One clarification to the answer to Steve's question at around 22:30. The input sampling is a "uniform random" variable with an average time of 5 ms in that example. This is because as Guillermo explained earlier, the input could have arrived anywhere from 0 to 10 ms before the frame, so the sampling causes an average of 5 ms of latency, though any particular frame has a roughly equal probability of being any value from 0 to 10 ms.
Correct-thanks for clarifying Josef!
Super helpful! Thanks for the extra info!
@GamersNexus The best analogy is video camera. As the shutter opens, information starts to accumulate, but at some point shutter closes, and everything that has been collected up to this point in time gets saved as a frame. Then shutter opens again and the cycle repeats. And since a click can occur at any point of that cycle with equal probability, statistical mean value is going to be half of the cycle time.
Shannon theorem
I would like to point out that in case of the cpu load is low enough, you could get away with something like generate the new frame asap, poll continuously for inputs and if you get new inputs generate the new frame and throw away your old frame sitting in the render buffer. This removes the risk of losing frames you'd get by waiting, but will hit your cpu harder.
Battlenonsense being cited in this video warmed my heart.
Great video with great insights!
I miss him! Hopefully he comes back around in the future.
Battle(non)sense is one of the most MASSIVELY UNDERRATED channls to ever exist on TH-cam.
@@SKO713 He's a Battlefield player so hopefully when the next game drops, in 2025 or 2026... We have been in limbo since 2013.. Some might say 2016.
Yeah, that excellent channel not getting the success he deserved and not making any more is part of why I hate yt so darn much
@@griffin1366 Battlefield 1 was a diamond in the rough. It runs so good these days. It's optimised to high heaven. Bigger maps, more players, same FPS as CS2 LOL.
I really enjoy the shoutout to battlenonsense and that they learned from that. Really humble, likeable guy who explains it well.
They hahaha!!!!!!!!!!!!! NVIDIA Learning (NASDAQ) US$ 726,58 -12,42 (-1,68%) just give us good stuff for good price then you sell more !
I’m a consultant in the advanced technology space and have to give technical “talks” for a living - this was great. It’s really hard to explain technical complexity in a way that is accessible; it’s all abstraction of language we are dealing with so things get confusing quickly. This audience is likely much more technical than most, however thats still very cool something this in depth and a bit esoteric has 200k views. Solid work. Looking forward to Pt 2!
i often end up explaining how pcs work to some of my friends and this video has given me a new perspective on how to explain the cpu and gpu in a very simple fassion, the cpu draws everything and the gpu colours it xDDD
@@Hydroxi123 cpu writes the recipe. gpu mixes the ingredients as per instructions. display cooks until ready.
Thank you for looking at this. Latency is definitely the most underrated metric when measuring system performance right now. I purposefully optimize my graphics settings to get the lowest amount of input lag possible, rather than going for more FPS.
You nailed it - very underrated, and we've been a part of the FPS focus (even with frametimes) and want to expand the scope! A lot of times, the two align and are correlated. But there is definitely a lot more to the story!
Yeah, exactly. It can be a very tricky balancing act with some games, but it's worth it.
Same. Limiting FPS for low latency and consistent frame times almost always leads to a significantly better experience than running the GPU at 100% for peak FPS.
Yeah, take Frame Generation for instance - it's being marketed as a performance boost (which I absolutely hate), because the FPS values are higher (bigger number bigger better), but the latency is in some cases tripled. It should be advertised as motion smoothing technology or something
@TK-11 yeah at least in val at 900 fps uncapped it's saying my game to render latency is about 3ms while capping it to 500fps reports roughly 2ms. Interesting seeing im using reflex which is meant to eliminate the latency overhead u just mentioned.
I love listening to people who love and care about what they do talk about their work, especially highly technical things being explained in a more accessible way. Great stuff from the both of you, this video was really cool.
I really like that you do these deep background pieces, they give the required context to understand your also excellent hardware reviews.
Fantastic video, the guy from Nvidia was an absolute pleasure to listen to and very professional. I now understand why when I run with reflex on my game runs so much smoother. Look forward to part 2. Thanks Steve and team
Fantastic coverage on this underrated metric. It felt very enjoyable to watch and not at all 'lecture-esque' as you thought it would be! Thanks for the great video and awesome content :D
The animations showing the render queue in GPU bound and CPU bound situations was really great. It clarified some things that I had always wondered.
Am I getting this right.. Being GPU bottleknecked is really bad? And that u always want to be cpu bottleneck?
Anyone please help😭
So basically I never want to be gpu bound or what? Even if it's like a 4090 + 14900k in that case i'm still gonna be gpu bound right. Or maybe it doesnt matter at high fps say240+ fps?
@@jokeren8071 Seems like you should set a fps cap close before the gpu bound happens, e.g. if gpu is 100% usage at 250 fps, do e.g. a 240fps cap
I'm really glad that latency is getting so much focus these days honestly. For a long time i feel like it's not really been talked about much, despite being so critical for the user experience
very educational video, i've been playing fps games for years but never really put lots of thought into how it all worked. really good to know
The game engine section was great!
@@GamersNexus Awesome video, primarily playing FPS games lately and adding this information to reviews will allow me to purchase the products that will be more attuned to my gaming goals rather then just 'more fps = better'.
@@Reznap7 I game between 60Hz and 85Hz on a CRT monitor (i could push it beyond 85Hz but a CRT is so fast that higher refresh rates than those are kinda pointless, it already creates natural motion blur at 85Hz).
We are on this absurd bubble of Hz and fps for gaming, it's like the loudness wars for music.
Finally, since VGA CRTs are analog, not digital, they don't add lag (you don't need anti-aliasing, AI upscaling and frame gen since CRTs don't have pixels or a fixed native spatial and temporal resolution).
Latency is not just a gaming thing! It's of utmost importance for professional audio and video recording too, so not only a graphics thing. Imagine recording like 15 microphones to separate tracks simultaneously, with EQ and many other effects on each, and the whole, and slamming it on a drive, as close to real time to be able to monitor it all! Same goes for recording scenes with many cameras, all at once, or even editing live between them.
Please do more on the subject Steve, and also give it more love from a processor/RAM... perspective.
and to circle back, audio latency can also be a huge gaming thing. Certain audio cues are used in place of visual cues in cases which they are faster. For example, it has been studied that the top LG (lightning gun) players in Quake have reaction times that exceed what is possible from visual input. One theory is these players use the audio cue of the hit indicator to calculate when the enemy player changes direction before they actually see it. Human reaction to audio stimuli is also slightly faster than visual stimuli
@@chasmurphy1227 Audio latency should not exist in 2024, that was mostly eradicated when quad cores became mainstream. Even 32 bit 192khz should still run perfectly fine. If your CPU is running at 100% then maybe there's a difference, but firstly you shouldn't be letting your CPU hit 100% as that caused extreme instability on your PC, and secondly most games won't utilize every thread on 16 thread and even some 12 thread CPUs. Audio is extremely important, but I don't see audio latency/lag being an issue still.
@@mttrashcan-bg1ro Sure it is. Wireless/BT peripherals introduce lag.
Latency is definitely a gaming thing.
@@necro4468I clearly wrote: Not JUST a gaming thing, not "Not a gaming thing" at all.
My goodness, this was an amazing compilation and summary around everything to do with input latency and render latency. Glad you are expounding on this topic more since FPS numbers tell so little about how the game feels to play. Thanks for doing this!
This is what I've been saying for 10 years now. I never understood why no one ever cared about input lag. It is SO important.
The professionalism and quality of this channel are rocketing high sky, congrats guys, and thanks for all the info.
Really excited to see you guys are going to add latency to your testing. This is awesome news and something that hasn't gotten much attention elsewhere.
you're doing us a favour with this video. i hope this blows up
Thank you steve.
Just by you doing this, the industry and gaming is going to get a whole lot better.
I respect your engineering and testing methodology. No one else takes the time to look at it with this attention to detail and consistency. In addition, you take the time to explain it so I know why the numbers matter and how you derived them. Spewing out a ton of data can quickly get boring and become meaningless, but you manage to keep the detail while making it relevant and understandable. That's not easy.
Glad that you're starting to look into this. Though this is MUCH larger than just for graphics. I.e. a very good example is storage (nvme/ssd/hdd); memory access; as well as general interrupt handling delays. The biggest item I run into pretty much daily is data subsystem performance (drive I/O basically). This is made worse as basically OEM's and all reviewers seem to be fascinated by 'streaming throughput' to a subsystem when this workload basically is completely non-existent for any real world workload. (anyone who doubts this, load up diskmon and check for yourself, it doesn't exist). Wish you can add similar level of detail to any reviews you do with these other subsystems.
Definitely agreed on storage mattering too! Talked with Wendell about this recently in a video!
100% agree but if u care about storage io latency so much those optane drives have been on flash sale recently and in case it wasn't obvious, they demolish any normal ssd in io latency both in averages and especially consistency.
Right, though for *average* latency you can't measure things by the weakest link in the chain (e.g. the SSD), because there are a bunch of layers of cache in between that prevent you from being bottlenecked by it. SSD -> RAM -> L3 -> L2 -> L1 -> CPU would be the rough chain for CPU work on a modern PC (SSDs can also have RAM, but it doesn't quite fit in with the other components in the chain which go from big to small, slow to fast).
Things get slow if you miss every cache and have to ask the SSD directly, but in a lot of workloads everything can be loaded in advance or consistently streamed in before it's needed, and you never hit that slowest path at a critical moment. In practice there *are* going to be misses, which is one source of frame time spikes, but if things are working correctly these are low probability events (but of course it depends on how much data the workload churns through).
@EmanuelHoogeveen Yeah true but remember to a cpu each miss is super detrimental when it's running at like 0.2 nano seconds per clock while a good ssd has an access latency of 50 microseconds, ie 250,000 times that, so a lot of wasted cpu cycles whenever it is waiting on data from the ssd. That's why hard drives had been such an enormous bottleneck for so long. To a cpu each miss as technically super rare as it is (proportionally i mean, it happens pretty frequently in absolute terms, check resource monitor and look for "page hard faults") is detrimental when they're that fast. Oh yeah and literally almost any time u load something ur almost definitely gonna get tons of misses. That's basically what loading something is, moving data from storage into ram.
Oh and fyi ssds have inherently really bad writes which is why dedicated dram cache on them is literally only used for writes and never reads so for latency it isn't really much help as generally, ur always read bound not write bound.
TLDR: No one's assuming ur cpu is pulling data from the ssd 100% of the time, the point is it it affects ur system latency regardless.
@@FrozokenI didn't realize that SSD RAM is mostly for writes but that makes sense, thanks :) I was trying to figure out where it fits in since it's so small relative to the SSD and even your regular RAM, but if it's used as a buffer more than cache that makes more sense.
That was genuinely a very good one! glad to see you shedding light at this latency topic, looking forward to future content like this!
I have always and continue to appreciate GN's fine line-walking act between presenting the technologies companies implement to make things better... and an ad. This video is great! It doesn't feel like an ad for Nvidia, it feels like education on why Nvidia is a leader in competitive gaming and what technology they have invented to make sure the human component of competitive gaming is the one that matters most!
It also helps explain why mysterious settings like Nvidia Reflex might show up in games you play!
This feels like I'm in a lecture and I should be taking notes!
Really interesting video; he's clearly very dedicated to what he does. I think more gamers should watch stuff like this, get a better understanding of what's going on closer to the hardware level and gain some appreciation to everyone making this stuff possible.
Wanna copy my notes? I expect quiz at the end of pt2.
SUPER useful piece, GN! I've admittedly downplayed response times in my choices of display thinking it wasn't really going to matter to me. This gives me a much clearer picture of what's actually going on and will allow me to make better decisions in the future. Thanks!
This is hands down my favorite video you've ever posted! Thank you!
This is great content. I love understanding the graphics pipeline. Very clear presentation. Shoot more videos like this, Steve! I enjoyed the TAP VR pipeline presentation years ago. I still go back to it from time to time.
Damn i had no idea David does tech stuff youtube channel was also a CS Pro
hahaha. Didn't see the resemblance until you said it!
I was going to write the exact same thing
Same I thought it was him too
the second I saw him on screen I went to check the comments if anybody else noticed him, they look so incredibly similar lol
Thats pretty cool. I was still playing consoles when people were playing that game. I the whole csgo generation is a really important chunk out of the overall history of gaming. I always imagine what video games either traditional or VR will be like in a few hundred years if we don't kill the planet with nuclear war and find a way to coexist regardless of borders.
This educational deeper dive into the he why and how is invaluable to consumers.
Most of us care about the entire experience of gaming not just fps.
Looking forward to part 2!
amazing video, answered so many of my questions that I've had for years. thanks team GN & NVidia!
Guillermo is quite clearly very knowledgeable and also an excellent teacher. Props for an excellent discussion from Steve too!
I love this. Thanks for making a technical subject more accessable.
I’m so glad you guys did this. The fighting game community appreciates this
Great video! This is the level of knowledge dump I usually have to go to a conference to get.
Great video explaining reflex. As I watched I was thinking why not hold back the CPU and then yes - that’s exactly the solution.
Latency used to be something often tested. And then it suddenly went away. I remember a common one was the test with a mouse click and led to when the gun muzzle flash. Or whatever was supposed to be triggered. And there was more and more focused on frame. To the point we are now counting fake frames. How is that different to a time when we had games making 30 frames but displaying 60 by duplicating them. And the test was usually done for displays reviews only.
Man I loved Guillermo. Not just the all the info but how well he presented and explained it.
Also how he kept moving in place. XD I'm always amazed how host like Steve can manage to stay still for so long. It was nice see someone like me that can stay still for 3 seconds.
LCD display added absurd amounts of lag compared to crt. Maybe it was display companies being shady.
Frame generation tech has to be the low point of modern GPUs and gaming. They can't get real performance, so they're literally faking it. Lame.
8 people dont know how silicone works lol yeah its magic yall@@paradoxicalcat7173
@@Touma134 of corse... and is all about money
This was a fantastic and very educational video! Thank you for always putting in the work to not just give us information but to teach us how understand it also. This was an awesome video!
Amazing stuff! I love that you picked Counter-Strike as an example and I'm looking forward to part 2. After playing CS2 the whole evening with Reflex off, it might be the time to finally give it another shot!
great explanation. thanks for putting the time in to do this
Mentioning Battle(non)sense was so nice to see! Love the channel and I actually hope he can monetize his content better. It's a real shame he can't do more of his stuff!
As long as he is healthy and happy, that is much more important than money and power!
Finally. Glad theres someone looking at latency, frametime, micro stutter, theres so many factors involved that make much more sense than fps.
Great video - as always. Keep up the incredible nerd work. This stuff is far more interesting than it should be.
Wow, this is extremely informative, and still digestible. Kudos, and I can’t wait for part 2!
Omg I have been doing and learning all I can to improve latency in windows and beyond. This topic is so interesting and important
Awesome job, learned so much, thank you! Like I said in the comments of a previous video, "input lag is more important than FPS, frame time, frame spacing, etc." is the hill I'm ready to die on.
God I'm so happy you're doing this. The speedrunning community has been on this train for a while now; we play old games that are often locked as low as 30fps, so we've needed other means to improve the game feel and quickly realized how much more important latency actually is. But it's been so hard to find coverage on latency as a buyer, which it looks like you guys are striving to fix. Looking forward to part 2!
this kind of testing is really important, recently i have been trying out frame generation in a bunch of games thanks to Lossless Scaling's implementation, and every single time i have decided against frame gen due to input latency. Frame gen from 85 FPS to 170 FPS feels so much worse than 90-130 real FPS with equivalent settings.
EDIT: Lossless Scaling just updated to not require a frame cap when generating frames, meaning I can now play at an unlocked 90-130 FPS with frame gen. After 5 minutes of getting a feel for this, I think it is AMAZING! Maybe I'll be noticing the input latency after trying this for longer but for now this is definitely how I'm playing my games.
FSR 3 in beta on Starfield has done this adding a load of input lag compared to DLSS in the previous patch, so you are 100% correct.
My Samsung CRT: "What is that?".
It really depends on the game though, and its implementation. I've had frame generation probably add 10-15ms input latency (you can do a test that shows how sensitive or accurate you are when it comes to telling how much input latency you're getting and FG felt like 10-20 in some games I've tested it with). But with other games, it easily adds 30-50ms. For context, I have a 4080 super
BRILLIANT!
No wonder your channel is second to non!
What blows my mind, is how freakin fast this process is repeated every second per frame! incredible!
This is great content. Thanks Steve and Guillermo :D
Incredibly informative video! Thank you for taking the time and putting effort of creating this video, breaking down the PC latency in video games.
Especially with higher polling rates being more and more common, latency deep dives are super necessary
Doesn't matter how fast you can get data from your hand to the computer; it's still going to sit in a cache until the CPU until it's ready to be processed. This is why polling will always be inferior to interrupt, which tells the computer to "drop everything you're doing and process this data". PS/2 purists aren't just being luddites.
with 1ms or less delay from the polling rate with a decent mouse, the savings seem kinda negligable @@MyNameIsBucket
@@MyNameIsBucket This was correct but modern chipsets put the PS/2 interface through the same caches as polled devices. Not only that, the PS/2 protocol is so slow it takes around a millisecond to transmit a single scancode, so one or multiple milliseconds per pressed key. A 1000Hz USB keyboard refreshes its entire rollover in the time it takes a PS/2 keyboard to signal a single key.
@@mitsync Unless your CPU is very busy completing other tasks like, oh I don't know, trying to render a CPU-bound game at 120Hz. Then the input from that USB keyboard takes quite a bit longer to register in the game. The PS/2 keyboard will still take exactly the same time.
@@MyNameIsBucket You are prioritising input over simulation speed. The game isn't going to do anything with your input until the next tick anyway, and very very few CPU bound games run over 60 tick. 1000hz polling is already well past the point of diminishing returns, you aren't gaining anything by trying to squeeze your inputs in quicker than that.
I felt like there was going to be a quiz at the end. That's how detailed this was. Great job everyone.
I’d have flunked it.
This was really informative. Can't wait for this to become a standard addition in tests.
Really enjoyed this detailed lecture/breakdown of Latency! Would love to see more vids like this keep coming. Thanks Steve!
I absolutely love this. Seeing numbers on charts can help get to the idea of what's doing what in your system, but having an engineer break down exactly how the pieces work together, and what their tech is doing to optimize it, that I love. One of the best videos I've seen in awhile.
It would be great to have a similar piece with AMD to understand how Reflex and Antilag differ. Great job GN! 👏
As long as that one brother and sister work there, you will keep laughing hahahaha
This is really helpful and interesting stuff. This vid helps tremendously in understanding what really happens when we're running into performance limitations. Thanks for this!
Steve working to force all of tech youtube to get better on their game again. First it was pushing the moves away from FPS, now its in coming up with entirely new understandings for how to demonstrate user experience. Great work as always!
This was a great video, and an excellent talk/conversation! Thank you for this
Tech Yes City did a fantastic piece on this as well, his finding was latency peaked with intel 10th gen
I wonder whens part 2, man that was so informative and super interesting, thanks
Thanks you so much for mentioning the most UNDERRATED channel on youtube in Battle(non)sense - please help that brilliant man
These are the type of videos that are the best. Thank you.
This is the content I love and trust from GN. Very well delivered!
Finally!!!
I'm really looking forward to this journey. I'd love to also consider adding some Linux tests now and then, since Nvidia Reflex with latency markers will be added soon to Proton.
This was VERY interesting and quite enlightening. It was also explained in a way that i think even people without much theoretical knowledge can understand .
Loved this! This kind of in-depth explanation of complete system latency is Amazing! I thoroughly enjoyed this video... and have enjoyed all of your in depth & complex dives into how the tech is actually functioning! I'm Yearning for more like it! :)
This is literally gonna be the next step in hardware reviews. Hardware has become so good only unoptimized AAA titles and I guess casual games with good graphics need more FPS, but a ton of people who play competitively already get more than enough FPS. What really matters is whether the frame you see is even the one you're supposed to see and if the input you made gets processed fast enough that it's accurate. There's a lot of aspects to it and FPS can only do so much in that regard. I hope other tech reviewers start looking into it as well, it looks hella interesting.
Nothing beats CRT PC gaming, absolutely nothing.
For this to work it'd require measuring latency in complex states/environments. Just firing a gun without anything else is informational but it could be rather misleading. It's difficult to be the "next step" as it'd require to be somewhat standardized/repeatable as well.
This was a dang good video! Also helped me understand what reflex is actually doing! Hope more game studios incorporate it into their games.
These are the type of metrics that probably matter most. It doesn’t matter if you have the fastest processor and GPU if you are bottlenecked somewhere. Thanks for starting to add these!
Yep. We need someone to create a tool like PCPartPicker but with product-specific latency listed that lets you choose all your hardware and estimates your total system latency and breaks it all down, showing you where your bottleneck is and if there's any weird interactions between specific hardware, etc.
Finally had time to watch this. Great stuff! Looking forward to Pt2
4 months out and we're still waiting for PART 2. When? 😋
Fascinating. You guys are doing great work, and this is super informative! Keep it up.
This is really good, thank you for the quality and in depth analyses
Guillermo did an excellent job here explaining in a way that you could follow along without being super tech-savvy.
I love how battlenonsense get quoted here. Guys a goat
thank you for making this
Been waiting on a video like this! If anyone is interested in knowing more about polling (specifically on old consoles), there is a video called "You will never see a Bomb Torizo skip in a World Record Super Metroid speedrun (NTSC)" which talks about issues emerging at polling rates around 60Hz. It's very focused on Super Metroid, but I found it very interesting.
Wow, this is some video. Really high on the educational value. Kudos!
Not up to frametimes, I think it is much more important!
Thank you!
This is awesome, I like these deep dives, made me realize that all of us take for granted to a large degree, the engineering behind playing a video game and all that goes into making it a great time.
Thank you for this interview and the breakdown of this topic.
Awesome video and so excited for what's to come!
This is fascinating and exactly the kind of content I crave for being an FPS nerd. I would point out that if you're going to go focus on latency in competitive games, optimizing settings for performance would help quite a bit. For instance, your R6S testing is done on DX11 while the game run significantly better on Vulkan (at least for relatively recent hardware). It would be nice to see these tests done on minimum setting and at 1080p (and, maybe, 1440p), since that's how people who really care about performance tend to play them.
What an amazingly informative video, thanks so much I learned a lot. And thanks to buddy from nvidia for helping us understand latency from a more technical perspective. Great video and why I love your videos!
This is awesome. As a sim racer I am was always wondering how to get the lowest latency and how it all works.
thanks for talking about this as a fighting game player it's my primary concern. I would argue in this realm it's even more important than in competitive shooters.
Wow this was eyeopening. Thanks! Awesome.
Absolutely amazing content. There aren't many channels like this one. ❤
Great video!
Input lag is something that is commonly addressed in the retro gaming community, but most of the discussions center around the display, and namely the benefits of using a CRT to eliminate the display side of lag. People never seem to care that certain ports/games can have more or less lag inherent to the game software or hardware itself.
One interesting solution that is viable for retro games is frame look ahead, where the emulator/game engine is running multiple instances of the game to predict inputs and thereby reduce input latency on the game software side. Obviously this is mainly possible due to the vast difference in hardware power that allows us to simulate many simultaneous versions of these old games, but I do wonder if we'll ever see something like that come to modern games to reduce latency. Maybe in some ideal world our hardware power continues to increase, but game graphics/overhead stays similar allowing us to run multiple instances. Highly unlikely, but a fun thought.
Replacing an LCD with a CRT for Counter Strike 2 is the best thing i ever did.
I can watch everything so clearly now, peaking is ridiculously easy.
I have a lot of fun playing Fall Guys, but then i noticed that the game is locked to 60fps.
Some games simply don't have unlocked framerates.
Depending on the game, 60Hz can look a bit flickery on a CRT.
@@saricubra2867According to the PCGamingWiki for Fall Guys, the game is not capped at 60 Hz:
"Since Season 5, you can freely adjust the FPS limit in-game when VSync is turned off."
@@GS-xp5jq It doesn't allow capping to any framerate.
Amazing to see this stuff, and that you collaborated with n0thing. Can't wait to see more.
Cool to see some of the visuals from Battle(non)sense in the presentation!
Really excited to see latency metrics- super important in competitive games, ty for adding this!
Awesome! I've been wanting a deep dive into the importance of latency. It seems that 4k gamers are missing out on a lot of responsiveness, so the new oled monitors that can switch between 240hz 4k and 480hz 1080p should be a godsend. Now I'm just waiting for a reflex oled monitor at 1080p . . . one can dream.
this is the type of detail test i love to see. because for serious gamer and specially competitive gamer - this is meaningful input latency to human reaction time to keyboard press ( some keyboard has higher height before it execute) all adds up to delay which can ruin your game play.
thank you very very much!
Edit! PLEASE CHECK motherboard DPC Latency with LatencyMon!
Amazing video guys, thank you!!
Gamers do latency analysis all the time, but we do it by the seat of our pants rather than by instrumentation. When you do analysis by the seat of your pants, it all comes down to how it _feels_ to play the game. And yes, it matters a lot.
@@redmoogle I'm (somewhat unscientifically) convinced that most of the issues with jittery game play and latency are being caused by the fifty gajillion processes running on any Windows and the effort to run games at high levels on hardware that's not capable of doing it.