From time immemorial, gaming capabilities of computers and individual system components have been measured in frames per second, and long-term benchmarks are the gold standard for testing, which allow you to compare different devices in terms of stable performance. However, in recent years, GPU performance has begun to be looked at from a different angle. In video card reviews, graphs of the rendering duration of individual frames appeared, the question of FPS stability arose in full growth, and it is now customary to accompany average frame rates with minimum values filtered along the border of the 99th percentile of the frame time. Improvements in test methods are aimed at finding delays that dissolve into average frame rates, but are sometimes quite noticeable to the naked eye of the user.
Nevertheless, any software measuring tools that work inside the test system give only an indirect estimate of a hidden variable that is of decisive importance for a comfortable game - the delay time between pressing a keyboard or mouse button and changing the picture on the monitor. We have to follow a simple rule, which says that the higher the FPS in the game and the more stable it is, the less the response time to input will be. Moreover, part of the problem has already been solved by fast monitors with a refresh rate of 120, 144 or 240 Hz, not to mention future 360 Hz screens.
However, gamers, especially players in competitive multiplayer games who are looking for the slightest advantage in hardware over their rivals and are ready to build overclocked computers to order for dozens of extra FPS in CS:GO, have not yet been able to directly assess input lag. After all, such precise and time-consuming methods as shooting a screen with a high-speed camera are available only in laboratory conditions.
But now everything will change - meet LDAT (Latency Display Analysis Tool), a universal hardware tool for measuring gaming latency. Readers familiar with acronyms such as FCAT might guess that this is an NVIDIA product. Indeed, the company offered the device to selected IT publications, including the editors of 3DNews. Let's see if a new measurement technique can shed light on the baffling phenomenon of input lag and help gamers choose their esports components.
LDAT - how it works
The principle of operation of LDAT is very simple. The core of the system is a high-speed photosensitive sensor with a microcontroller, which is mounted at the desired point on the screen. A modified mouse is connected to it, and the control software over the USB interface detects the time separating the keystroke and the local jump in the brightness of the image. Thus, if we place a sensor on top of the barrel of a weapon in a shooter, we get the exact amount of delay that the monitor, computer, and the entire software stack (including device drivers, the game, and the operating system) need to respond to user actions.
The beauty of this approach is that the operation of LDAT is completely independent of what hardware and what programs are installed on the computer. The fact that NVIDIA is busy producing yet another measuring tool, which is also available only to a limited circle of IT journalists, hints that the company is trying to emphasize the advantages of its own products in comparison with competitors (this happened with FCAT several years ago). Indeed, 360Hz monitors with G-SYNC support are about to appear on the market, and game developers will start using NVIDIA Reflex libraries aimed at reducing latency in games under Direct3D 12. Nevertheless, we are sure that LDAT itself does not provide any concessions "green" video cards and does not distort the results of the "red" ones, because the device has no access to the configuration of the experimental hardware when it is connected with a USB cable to another machine running the control software.
Needless to say, LDAT opens up enormous prospects in its field of application. Comparing gaming monitors (and even TVs) with different refresh rates and different types of matrices, checking how latency is affected by G-SYNC and FreeSync adaptive sync technologies, frame scaling by a video card or monitor - all this has become possible. But first, we decided to focus on a narrower task and check how several competitive games designed for high FPS and low reaction time work on video cards of various price categories. And if we formulate the problem more precisely, we are interested in two main questions: is an excessive frame rate a guarantee of low latency and under what conditions it makes no sense to increase it (and therefore buy a more powerful video card). In particular, is it useful to exceed the frame rate corresponding to the refresh rate of the screen if you are the lucky owner of a high-speed 240-Hz monitor.
For testing, we selected four popular multiplayer games - CS:GO, DOTA 2, Overwatch and Valorant, which are undemanding enough that modern GPUs, including budget models, can achieve hundreds of FPS performance. At the same time, these games allow you to easily organize an environment for reliable measurement of reaction time, when constant conditions are most important: the same position of the character, one weapon in each test, etc. For this reason, we had to postpone for the time being benchmarks in games like PlayerUnknown's Battlegrounds and Fortnite. In PUBG, there is simply no way to isolate yourself from other players even on a test site, and Fortnite in the single-player Battle Lab mode is still not protected from loot accidents and, accordingly, makes it impossible to test several video cards with the same weapon in a reasonable time.
In addition, select games have the advantage of running the Direct3D 11 API, which, unlike Direct3D 12, allows the graphics card driver to set limits on the render queue of frames that the CPU can prepare for rendering to the GPU in the software graphics pipeline.
Under standard conditions, especially when the bottleneck of the system is the computing resources of the video card, the frame queue increases up to three by default or, if the application requires, even more. In this way, Direct3D provides continuous GPU loading and a constant rendering pace. But at the same time, there is a side effect in the form of a delay in the response to input, because the API does not allow pre-planned frames to be thrown out of the queue. The corresponding settings in the video card drivers, which were popularized by AMD under the Radeon Anti-Lag brand, are aimed at combating lag, and then NVIDIA has a similar Low Latency Mode option.
However, such measures are not a universal remedy for delays: for example, if the performance of the game is limited by the capabilities of the central processor, and not the graphics processor, a short frame queue (or its complete absence) only makes the CPU bottleneck narrower. In addition to the rest of the test program, we intend to find out if the Radeon Anti-Lag and Low Latency Mode "technologies" have any tangible benefits, in which games and on which hardware.
Test stand, testing methodology
| Test stand | |
|---|---|
| CPU | Intel Core i9-9900K (4,9GHz, 4,8GHz in AVX, fixed frequency) |
| Motherboard | ASUS MAXIMUS XI APEX |
| RAM | G.Skill Trident Z RGB F4-3200C14D-16GTZR, 2×8 GB (3200 MHz, CL14) |
| ROM | Intel SSD 760p, 1024 GB |
| Power supply unit | Corsair AX1200i 1200W |
| CPU cooling system | Corsair Hydro Series H115i |
| Chassis | CoolerMaster Test Bench V1.0 |
| Monitor | NEC EA244UHD |
| Operating system | Windows 10 Pro x64 |
| AMD GPU software | |
| All video cards | AMD Radeon Software Adrenaline 2020 Edition 20.8.3 |
| NVIDIA GPU software | |
| All video cards | NVIDIA GeForce Game Ready Driver 452.06 |
Frame rate and response time measurements across all games were performed at or near maximum graphics quality settings to a) accentuate differences between compared devices, b) get results both at high frame rates above screen refresh rate and vice versa . Especially for this article, we borrowed a fast Samsung Odyssey 9 monitor (C32G75TQSI) with a WQHD resolution and a refresh rate of 240 Hz - the maximum for modern consumer monitors before 360 Hz screens were on sale. Adaptive refresh rate technologies (G-SYNC and FreeSync) have been disabled.
The results of each individual test (a specific video card in a specific game with or without anti-lag driver settings) are obtained on a sample of 50 measurements.
| Game | API | Setting | Full screen anti-aliasing |
|---|---|---|---|
| Counter-Strike: Global Offensive | DirectX 11 | Max. graphics quality (Motion Blur off) | 8x MSAA |
| DOTA 2 | Best Looking Quality | FXAA | |
| Overwatch | Epic quality, 100% render scale | SMAA Medium | |
| Valorant | Max. graphics quality (Vignette off) | MSAAx4 |
Test participants
- ;
- ;
- ;
- ;
- ;
- ;
- ;
- .
Note. In brackets after the names of the video cards, the base and boost frequencies are indicated according to the specifications of each device. Video cards of non-reference design are brought into line with the reference parameters (or close to the latter), provided that this can be done without manually adjusting the clock frequency curve. Otherwise (GeForce 16 series accelerators, as well as GeForce RTX Founders Edition accelerators), the manufacturer's settings are used.
Counter-Strike: Global Offensive
The test results in the very first game, CS:GO, provided a lot of food for thought. It's the lightest project in the entire test program, with graphics cards like the GeForce RTX 2080 Ti hitting frame rates beyond 600 FPS and even the weakest of the eight testers (GeForce GTX 1650 SUPER and Radeon RX 590) holding well above refresh rates. monitor at 240 Hz. Nevertheless, CS:GO perfectly illustrated the thesis that increasing FPS beyond the monitor frequency is not at all useless for reducing delays. If we compare the video cards of the upper group (GeForce RTX 2070 SUPER and higher, as well as the Radeon RX 5700 XT) with the younger models (GeForce GTX 1650 SUPER, GeForce GTX 1060, Radeon RX 5500 XT and Radeon RX 590), we are talking about a one and a half times difference in general the time elapsed from pressing the mouse button until the flash appears on the screen. In absolute terms, the gain reaches 9,2 ms - at first glance it is not much, but, for example, changing the screen refresh rate from 60 to 144 Hz (9,7 ms) gives almost the same amount!
As for how video cards belonging to the same broad price category, but based on chips from different manufacturers, correlate in terms of latency, we did not find significant differences in each group. The same applies to options in the accelerator drivers designed to reduce lag by reducing the frame queue in Direct3D 11. On CS:GO (at least in these test conditions), they usually do not have a useful effect. In the group of weak video cards, there is a slight shift in reaction time, but only in the GeForce GTX 1650 SUPER did it reach statistical significance in the results.
Note. Strong color icons indicate results with standard driver settings. Pale icons - with the Low Latency Mode (Ultra value) or Radeon Anti-Lag option enabled. Pay attention to the vertical scale - it starts above zero.
| Counter-Strike: Global Offensive | ||||||
|---|---|---|---|---|---|---|
| By default | Low Latency Mode (Ultra) / Radeon Anti-Lag | |||||
| Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | |
| GeForce RTX 2080 Ti | 642 | 20,7 | 6,5 | 630 | 21 | 4,6 |
| RTX GeForce 2070 SUPER | 581 | 20,8 | 5 | 585 | 21,7 | 5,6 |
| RTX GeForce 2060 SUPER | 466 | 23,9 | 4,6 | 478 | 22,4 | 5,8 |
| GeForce GTX 1650 SUPER | 300 | 27,6 | 4,3 | 275 | 23,2 | 5,4 |
| Radeon RX 5700 XT | 545 | 20,4 | 5,8 | 554 | 21,5 | 4,4 |
| Radeon RX 5500 XT | 323 | 29,3 | 14 | 316 | 26,5 | 14,5 |
| Radeon RX 590 | 293 | 29,3 | 5,8 | 294 | 27,5 | 4,9 |
| GeForce GTX 1060 (6 GB) | 333 | 29,6 | 7,9 | 325 | 28,2 | 12,9 |
Note. Statistically significant differences in the average reaction time (according to Student's t-test) are highlighted in red.
DOTA 2
Although DOTA 2 is also considered an undemanding game by current standards, it is more difficult for modern video cards to reach several hundred FPS in it. So, all budget solutions participating in the comparison fell below the frame rate of 240 frames per second, corresponding to the screen refresh rate. Powerful accelerators, starting with the Radeon RX 5700 XT and GeForce RTX 2060 SUPER, give out over 360 FPS here, but, unlike CS:GO, DOTA 2 more effectively directs excess GPU performance to combat lag. In the previous game, a Radeon RX 5700 XT-level video card was enough to make it no longer worth increasing performance further for the sake of reaction time. Here, the delay continues to decrease on more powerful video cards up to the GeForce RTX 2080 Ti.
It should be noted that it is the results of the Radeon RX 5700 XT in this game that raise questions. AMD's current flagship far surpasses even the GeForce RTX 2060 in latency and performed no better than the younger models, despite the higher frame rate. But reducing the frame rendering queue in DOTA 2 is really useful. The effect is not large enough to be noticed even by experienced cyberathletes, but it is statistically significant for four out of eight video cards
Note. Strong color icons indicate results with standard driver settings. Pale icons - with the Low Latency Mode (Ultra value) or Radeon Anti-Lag option enabled. Pay attention to the vertical scale - it starts above zero.
| DOTA 2 | ||||||
|---|---|---|---|---|---|---|
| By default | Low Latency Mode (Ultra) / Radeon Anti-Lag | |||||
| Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | |
| GeForce RTX 2080 Ti | 418 | 17,7 | 2 | 416 | 17,4 | 1,4 |
| RTX GeForce 2070 SUPER | 410 | 18,2 | 1,6 | 409 | 17,6 | 1,6 |
| RTX GeForce 2060 SUPER | 387 | 20,8 | 1,5 | 385 | 19,8 | 1,6 |
| GeForce GTX 1650 SUPER | 230 | 27,9 | 2,5 | 228 | 27,9 | 2,3 |
| Radeon RX 5700 XT | 360 | 26,3 | 1,5 | 363 | 25,2 | 1,3 |
| Radeon RX 5500 XT | 216 | 25,4 | 1,2 | 215 | 21,7 | 1,4 |
| Radeon RX 590 | 224 | 25 | 1,4 | 228 | 21,8 | 1,3 |
| GeForce GTX 1060 (6 GB) | 255 | 25,8 | 1,9 | 254 | 25,8 | 1,7 |
Note. Statistically significant differences in the average reaction time (according to Student's t-test) are highlighted in red.
Overwatch
Overwatch is the heaviest of the four test games at maximum graphics quality with full-screen anti-aliasing activated. Not surprisingly, every gigaflop of GPU performance here benefits reaction time. The range of lag values in Overwatch between graphics cards like GeForce RTX 2080 Ti and Radeon RX 5500 XT is 2070x. The numbers also show that more powerful video cards than the GeForce RTX 5700 SUPER only increase FPS, but they can no longer speed up the response even nominally. But replacing the Radeon RX 2060 XT or GeForce RTX 2070 SUPER with the notorious RTX 5500 SUPER in theory makes sense in order to minimize lag while maintaining high graphics quality. In addition, in Overwatch, again, one of the accelerators on the “red” chips proved to be not the best side. This time, the Radeon RX XNUMX XT, which significantly outperforms all other budget solutions in terms of average reaction delay.
Overwatch once again helped to prove that a) the performance of a video card even at high frame rates still affects the lag value, b) a formally more productive GPU does not guarantee lower input response lags. In addition to all this, the game demonstrated the reference work of the anti-lag settings of the graphics driver. When playing on relatively weak graphics cards (GeForce GTX 1650 SUPER, GeForce GTX 1060, Radeon RX 5500 XT and Radeon 590), a reduced frame queue can reduce lag by 9 to 17%. Well, for powerful hardware, it is still completely useless.
Note. Strong color icons indicate results with standard driver settings. Pale icons - with the Low Latency Mode (Ultra value) or Radeon Anti-Lag option enabled. Pay attention to the vertical scale - it starts above zero.
| Overwatch | ||||||
|---|---|---|---|---|---|---|
| By default | Low Latency Mode (Ultra) / Radeon Anti-Lag | |||||
| Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | |
| GeForce RTX 2080 Ti | 282 | 35,6 | 10,4 | 300 | 34,2 | 9,6 |
| RTX GeForce 2070 SUPER | 225 | 35,8 | 5,1 | 228 | 36,7 | 8,6 |
| RTX GeForce 2060 SUPER | 198 | 41,2 | 6,4 | 195 | 38,8 | 9 |
| GeForce GTX 1650 SUPER | 116 | 58,2 | 8 | 115 | 51 | 8,7 |
| Radeon RX 5700 XT | 210 | 39,6 | 7,2 | 208 | 41,4 | 7,2 |
| Radeon RX 5500 XT | 120 | 69,7 | 13,2 | 120 | 63,5 | 15,1 |
| Radeon RX 590 | 111 | 61,2 | 8,6 | 111 | 51,7 | 7,7 |
| GeForce GTX 1060 (6 GB) | 121 | 60,7 | 8,7 | 118 | 50,7 | 6,5 |
Note. Statistically significant differences in the average reaction time (according to Student's t-test) are highlighted in red.
Valorant
Valorant stood out among the test games with excellent - or, conversely, mediocre - graphics optimization. The fact is that, despite the huge difference in the potential performance of test GPUs, according to frame rate estimates, they all concentrated in the range from 231 to 309 FPS. And this despite the fact that we deliberately chose the most resource-intensive scene for latency measurements in order to enhance the expected differences. However, in terms of the distribution of lag values, Valorant is somewhat similar to CS:GO. In this game, owners of the GeForce RTX 2060 SUPER or Radeon RX 5700 XT are on an equal footing with users of more expensive and powerful accelerators. Even the junior video cards of the GeForce GTX 1650 SUPER and Radeon RX 5500 XT classes are not so far behind the older ones here. With such input data, it is not surprising that it is useless to limit the Direct3D frame queue in Valorant: the corresponding settings have a statistically significant effect for selected video cards, but its magnitude is completely miserable.
Note. Strong color icons indicate results with standard driver settings. Pale icons - with the Low Latency Mode (Ultra value) or Radeon Anti-Lag option enabled. Pay attention to the vertical scale - it starts above zero.
| Valorant | ||||||
|---|---|---|---|---|---|---|
| By default | Low Latency Mode (Ultra) / Radeon Anti-Lag | |||||
| Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | Average frame rate, FPS | Average reaction time, ms | Art. reaction time deviation, ms | |
| GeForce RTX 2080 Ti | 309 | 19,3 | 2,6 | 306 | 20,2 | 3 |
| RTX GeForce 2070 SUPER | 293 | 19,2 | 3,1 | 289 | 19,5 | 2,9 |
| RTX GeForce 2060 SUPER | 308 | 20,7 | 2,7 | 310 | 19,6 | 2,9 |
| GeForce GTX 1650 SUPER | 251 | 24,5 | 2,9 | 243 | 23,6 | 2,5 |
| Radeon RX 5700 XT | 256 | 21,9 | 3,3 | 257 | 21,9 | 2,7 |
| Radeon RX 5500 XT | 258 | 23,5 | 2,8 | 262 | 22,8 | 2,6 |
| Radeon RX 590 | 237 | 25,8 | 2,7 | 234 | 24,3 | 2,5 |
| GeForce GTX 1060 (6 GB) | 269 | 23,5 | 2,8 | 268 | 23,4 | 4,4 |
Note. Statistically significant differences in the average reaction time (according to Student's t-test) are highlighted in red.
Conclusions
Measuring latency in games with hardware has yielded rich results that, frankly, call into question the methods accepted in the industry for assessing the performance of video cards, when frame rate has been the only parameter measured for decades. Of course, FPS and lag are closely correlated, but at least in eSports games, when there is a struggle for every millisecond of delay, the frame rate no longer allows us to give an exhaustive description of performance.
In a brief study of popular multiplayer projects, we found several interesting phenomena. First, our data refutes the popular belief that it makes no sense to increase the FPS beyond the values corresponding to the screen refresh rate. Even on a very fast 240Hz monitor, games like Counter-Strike: Global Offensive can reduce lag by half by swapping out a budget graphics card for a top model. We are talking about the same gain in reaction time as, for example, when switching from a 60 Hz screen to 144 Hz.
On the other hand, the frame rate can still be excessive, when a more powerful video card only heats the air in vain and no longer helps to deal with the already extremely low latency. In all of our test games at 1080p, we found no significant difference between devices like the GeForce RTX 2070 SUPER and the GeForce RTX 2080 Ti. The absolute minimum reaction time we recorded is 17,7 ms and was obtained in DOTA 2. This, by the way, is not such a modest value, which, if translated into a refresh rate, corresponds to 57 hertz. So the following conclusion suggests itself: upcoming 360Hz monitors will definitely find use in competitive games - this is a direct way to reduce lag when computer hardware has already exhausted its capabilities and runs into the limitations of a thick software stack of the operating system, graphics API, drivers and the game itself.
We then checked to see if there was any benefit from software anti-latency tools, which so far have been limited to limiting the frame render queue in applications relying on the Direct3D 9 and 11 graphics API - the notorious Radeon Anti-Lag in the AMD driver and NVIDIA's Low Latency Mode. As it turned out, both "technologies" really work, but can bring tangible benefits only in conditions where the bottleneck of the system is the GPU, and not the central processor. In our test system with an overclocked Intel Core i7-9900K processor, such tools helped low-cost mid-range graphics cards (Radeon RX 5500 XT, GeForce GTX 1650 SUPER and similar past-generation accelerators), but are completely pointless when you have a powerful GPU. However, when anti-lag settings work, they can be extremely effective, reducing the lag in any Overwatch by up to 10ms, or 17% of the original.
Finally, there were certain differences between graphics cards from different manufacturers that could not be predicted from frame rates alone. So, AMD video cards sometimes provide the same short latency as formally more productive "green" devices (example: Radeon RX 5700 XT in CS: GO), and in other cases they work suspiciously slowly (the same model in DOTA 2). We won’t be surprised that if hardware lag measurement techniques like LDAT become widespread, avid cyberathletes who are fighting for the slightest advantage over their rivals will begin to select video cards for a particular game, depending on which model provides the fastest response time.
But most importantly, thanks to LDAT, we have the opportunity to conduct deeper studies of the delay time. What we have done in this preview is just the tip of the iceberg. For now, topics such as the impact of adaptive synchronization technologies (G-SYNC and FreeSync) on the lag, FPS limits in the game, dependence on CPU performance and much more have remained outside the scope. In addition, we are going to find out if we can achieve a high frame rate of hundreds of FPS and, accordingly, a fast response to input, not only in competitive games specially optimized for these criteria, but also in AAA-class projects that load the system much more. And therefore, does the average gamer, not the champion, need a cutting-edge monitor with a refresh rate of 240 or even 360 Hz? We will answer these questions in future work using LDAT.
Source: 3dnews.ru