Titanic efforts, part 1
Hello. Life is not enough for any PC-geek without a bad idea to install two video cards. I already wrote about the “Trolleybus of Bread” and the two younger NVIDIAs in SLI (then the GTX 950 was used). It is time to switch to heavier drugs.
Today our guest is two 980 Ti in SLI, one GTX 1070, a lot of dirty curses, 4k2k and the questions “how can this be”. In general, everything as always, it will be interesting, look under the cat.
The NVIDIA Maxwell architecture at the time of release was pleasantly surprised by the ratio of performance, heat transfer and price. At that time, it seemed that the next year or two was waiting for us to stagnate, and even two GTX 970 scandals (first with a “lack of” memory, then with crookedly indicated specs regarding ROPs) did not prevent her from becoming one of the most popular video cards in custom PC assemblies. Excellent 1080p performance, allowing you to play at maximum settings, good at 1440p, giving access to high settings, and 4k2k was quite playable if you were willing to sacrifice quality post-processing such as self-shading and any expensive hair type items from NVIDIA HairWorks or image blur.
')
To overpay for the 980th in the absence of 4k2k was a lesson ... not the most useful, the price difference did not cover the performance difference, but a 256-gb SSD for toys could be bought for the same money. The top segment was almost completely occupied by 970s from various vendors.
Approximately the same thing happened with the flagship assemblies, only instead of the 970th and 980th there were 980 Ti and GTX Titan X. On the second side there was twice as much memory and a slightly more efficient stuffing, but it was problematic to utilize all this power in games: Titan was like a lot of VR-enthusiasts, game developers and just people who have a lot of extra money, and remained: for 980 Ti asked for 650 bucks, and for Titan X - just a thousand. And the difference in real cases was at best 10 percent.
And if you had a lot of money, it was easier to spend $ 1,300 for two 980 Ti, than to take one Titan, which you had to find. Easier, it is easier, but ... There are already starting troubles. If the “mother” is old, and the video card with “high” cooling (above two slots) - SLI can simply not fit. Take the appropriate power supply and put it down. All other expansion slots are likely to be forgotten ... In general, there are a lot of nuances. However, it turned out that I had a suitable motherboard, and the opportunity to get two 980 Ti, and that is the best time when there is a 4k2k monitor on the table, and the GTX 1070 box next to it. the hole? : D
With the release of the first benchmarks of video cards on the Pascal architecture, the blogosphere, journalists and iron editions at first took a deep breath and then did not believe their eyes. Not only did Maxwell live only one generation of accelerators, not two (the “tick-to-do” strategy implemented by Intel at one time was successfully adopted by many manufacturers), Pascal simply broke all last year’s solutions, and with the release budgets (like the same 1060), sank more and B / C-iron market. Who needs the 970th warranty and with a 30-40% discount, when you can dig up a month and take 1060, which is faster in everything, has more memory, supports new technologies and will be new from the factory? Here are attached old video cards for friends, for nothing, or in the second computers at home. Yes, NVIDIA immediately stated (and did not lie) that the GTX 1070 will be faster than the 980 Ti, and 1080 will bypass the previous generation Titan X, despite the fact that the chips themselves belong to the consumer and not the prosyumerskoy line, and for the price are a replacement the usual "970 and 980, respectively.
Of course, journalists from the world of iron made benches, 1070 won 980 Ti, and peace-friendship-chewing gum, but ... here we have two 980 Ti. And the smell of fried!
So, the test stand. The pieces of iron are not the newest ones, but they still cannot refuse them with brute force.
The test bench today is Philips 241P6VPJKEB . This monitor belongs to the "business" line, but formally I would call it a variant for workstations. The design has a healthy minimalism, the chips include light sensors, built-in sound, a USB hub and a swivel mechanism, and the characteristics ... Well, let's talk about this in more detail.
This display is built on the basis of a 24-inch matrix with a resolution of 4k (3840 × 2160 pixels). The matrix itself is eight-bit, but the controller allows you to display up to 10 bits per channel, mixing colors from adjacent frames (FRC, also known as frame rate control, if you are not afraid of English - much more is told here). In fact, most of the cheap monitors do this for displaying 8 bits (and the matrix is actually 6-bit). There is a factory calibration for SRGB (not bad, I must admit), but calibrated colors are incompatible with the operation of the light sensor.
According to the table of basic measurements one could say: look, look, what a terrible Delta E. Normal. First, the monitor does not belong to the line for professional work with graphics. Secondly, this is the factory sRGB calibration. In the absence of the user's interest, there are three things to worry about: the uniformity of the color temperature depending on the brightness, the uniformity of the displayed gamma, and the flickering of the backlight. The first problem is this: in 90% of cases, blue LEDs with a yellow phosphor are used. Depending on the brightness of the backlight and how closed the LCD shutter is, the permeability of the resulting “white type” spectrum is different. The second aspect suffers from the fact that the characteristics of the matrix itself are extremely non-linear, and the more precisely the monitor LUT converts the incoming signal - the more linear and correct the result will be in the measurements.
Here is the color temperature:
The bottom 10% (as well as the top 10%, however, too) can hardly be looked at, measured there as a problem, and the above features of producing “white” light introduce their own errors. In the remaining 80% of the coverage, the main working area - the temperature is slightly overestimated from the “nominal”, but stable.
Gamma in bright shades somewhat "stomps", but there, due to the non-linearity of the magnitude and not very relevant range, this will not be noticeable.
The color triangle combines well with the theoretical sRGB: a small excess in the yellow-green part of the spectrum and a dip in the blue is the merit of the backlight. In general, the factory settings for sRGB are 90% better than what regular monitors have in this price range. And who needs calibration and other accuracy, you know, either spend three times more on the monitor, or order a calibrator and bring this model to full coverage.
By itself, the display is pleased with the quality of the picture, the high density of dots + high-quality matrix make themselves felt, but so far not pleased with the thoughtfulness of the design. Up to the same HP or DELL with a normal arrangement of ports, a well-thought-out menu that can be used without constantly checking with the “legend” does not hold out. The connection ports of the signal cables in one place, the input for the USB cable are in another, and even my Type-B cable sticks out because of the frame ...
Approximately the same disease in the other ports: scattered on the monitor surface, as a result, even the cable management bracket does not help you to put things in order: all the wires stretch in different directions and create an untidy "beard" under the monitor.
As for me, attention to detail is not enough, not at all Apple. On the other hand, Apple does not have 4k2k monitors in such a diagonal, and if there were, the price would not have pleased anyone. The layout is simple: yes, usability for the top three. But the monitor is taken for the most part, not to configure, but to work. And here for this, everything is there. Factory calibration under sRGB provides 99% coverage and acceptable accuracy. If you are not engaged in work with photo / video editing - you have manual adjustment, make the image a little warmer, turn on the light sensor, eyes will not get tired. Matting is very neat, I don't want to cry from the crystal effect and noise. True, here a very high density of points also plays its role: the fonts are chiseled, with natural smoothing of the borders, very neat and visually pleasing.
If you want 4k, and there is no room for 27-32 inch models - you are not likely to be confused by such trifles, as if not the most successful cable layout or faddish menu. Once set up and forgotten, but the picture will still please the eye for a long time.
So, 980 Ti in SLI, fresh DOOM on the ultra-settings, the light is extinguished, the sound is louder ... damn, how beautiful he is. And gives from 40 to 55 FPS. And it is in 4k, yes with anti-aliasing! In statics is not impressive, but in the dynamics - breathtaking.
The new NFS (not so new, but still) on the FrostBite 4 engine produces an amazing, really amazing picture in dynamics. Damn, he is not bad in statics, if you choose a good angle.
Of course, in some places the scenery is simply not designed to be looked at point-blank, and in statics the maximum is “mediocre”, but on the go ... No, this is actually one of the most spectacular auto-simulators at the moment. Stable 60 frames per second (rested in vSync), subsidence is mainly due to an unstable Internet connection or when crossing two active races with 5-10 participants. After all, 20 cars in the frame, and even with a bunch of complex light and reflections ... And the picture is still good!
UPD: I don’t know if EA or Nvidia broke there, but just yesterday the system with two 980Ti couldn’t squeeze out more than 50 FPS, the real ones were about 35-40 + permanent departures of the application. Sad, but true.
Okay, go ahead. Forza Motorsport 6: Apex, available in the Windows Store .... Does not support SLI at all. One video card just kicks the bolt. As a result, drawdowns up to 25 FPS, besides, trendy “user-defined” coloring pages downloaded from the Internet have a habit of “disappearing” depending on the drawing distance. Nizakhot.
What is there more severe? GTA V? From 40 to 50 FPS, strongly depends on the situation, sometimes there are microfreezes or FPS drawdowns of up to 25 in just a fraction of seconds. Apparently, the point is in SLI.
Witcher 3 with all DLC, location Tussent. The abundance of grass causes FPS drawdowns even on serious configs, and here we are all super serious. To be honest, the creators of the witcher just scored on the optimization of top settings. Minimum and medium tweaks to play could be on relatively old machines, and high and ultra ...
In general, it is 980 Ti, which is two - the sense is not very much. FPS floats from 25 to 60, depending on what is happening on the screen, the work of NVIDIA HairWorks (this feature is generally better off) and the position of stars in the sky. It is unlikely that the computing power of the two almost top-end video cards of the current generation was not enough.
Turn off the computer, remove a couple of video cards, install the GTX 1070, update the drivers, check that everything works ...
Specifically, I had a GTX 1070 GameRock option from Palit, 1080 and 1070 outwardly differ a little less than nothing (stickers). Theoretically, the GameRock line is designed for modders and fans of customizable backlights: native software should be able to repaint the LED strip on board. I was not engaged in such garbage, I was more interested in performance.
So, DOOM. The same 45-55 FPS, but without jerks, much smoother and more evenly. If the FPS sags, it is only on the number of opponents, and not spasmodically, with the appearance of special effects or sudden movements / at random. With the Vulkan API enabled, the FPS counter freezes at 60 and does not move. WIN.
NFS gives out 45-60 frames per second, drawdowns are also caused either by other players or by problems with the connection: the game is networked. In general, of course, the FPS is slightly smaller than that of Sparky, but the game is ported from consoles, on which 30 FPS are generally stable in the last titles - the ultimate dream, and it is not very well respected, I saw frank lags during group drift on PS4, and graphics and next to not lying with PKshny.
UPD: With the latest drivers and recent changes, a single card produces a more stable result without crashes than two 980Ti in SLI.
Forza 6: Apex, by itself, goes better by 1070 simply because 1070 is faster than 980 Ti, and the game does not know how to work with SLI. There was nothing below 45 frames per second, VSync did not let work above 60 (120-144 Hz 4k2k were not distributed yet), and we didn’t want to observe frame breaks.
Minus Forza 6 is not even that the game is raw, but that there is nothing to do in it. A short campaign, a couple of races of a “seasonal” character and everything. Little content: no tracks, no cars on which I would like to ride. I hope, in due course will correct.
GTA V gave out from 45 to 57 FPS in real juz-cases and up to 60 in cut scenes, in statics and in rooms - the load in this case is lower.
Witcher is another game where the brute force of the 980 Ti in the SLI is still a dodgy novice, couldn’t give out more than 40 FPS in this mode of the GTX 1070. However, the minimum FPS was also higher, there were no drawdowns and uncomfortable lags, and the main guarantee of a pleasant game with such dynamics is not so much a high FPS, as even. So, one video card won out "morally".
The culprits of this result in the picture: large volumes of vegetation and a lot of animations of the surrounding world.
In benchmarks and on demo stands at journalists, 1070 shows excellent results, but bypassing 980 Ti in SLI? .. However, I have already given the answer to this question. Keyword - demostand. For testing use iron, which allows to minimize the impact of the performance of other components on the test object. If they test processors, they put on flagship video and top memory. They test disks - they connect with identical cables, identical iron, maximum spherical conditions in vacuum. Well, game tests and video card tests, as you guessed it, are produced on top-level fresh pieces of hardware, if this is not an article about building a PC as a whole. For toys of my processor, there is enough of a head, even the very first i7 still on socket 1366 with the architecture of Bloomfield and the ancient DDR3 1333 MHz memory are inferior in game tests from 10 to 25% of the flagship i7-6700k, which have higher core performance per clock and higher clock speed and speed memory. The dog is buried in a thing called an API overhead.
If you explain it in a very rude language and do not go into how the video card draws the frame “layers by layer”, then each time the video card needs to draw something: take a model, pull a texture, find a shader, some people have to tell the video card what to do . This one is a central processor. Since 90 (if not 99%) of PC gaming is tied to Windows, in most cases high-level DX10 / DX11 or OpenGL type libraries are used for games that are relevant for 2015-2016.
They make it easier and faster to work with 3D from the point of view of development, it imposes certain limitations and associated costs: in a simple language, you tell the video card what to do, it does.
On consoles, developers can write a much lower-level code, achieving good results from fairly weak hardware by simply reducing the overhead of drawing. Development becomes more difficult, the implementation of the code is tied to one hardware, but for consoles it doesn’t really matter - you didn’t need to choose any graphics settings or take into account different hardware (literally before the current generation).
These same overhead costs grow with the number of special effects, the resolution of the image, various post-processing tools, and so on. Even “insignificant” in terms of direct load on computing power, the Ambient Occlusion setting can lead FPS from a comfort zone to an uncomfortable one, because the video card will not give you a frame until it draws the shadows that objects cast on themselves. And these shadows are calculated through simplified ray tracing, which requires a lot of calls and the speed of the algorithm depends strongly on the resolution of the image.
So here. First, these calls to draw are not very well parallel. In some cases, in benchmarks, asynchronous requests work even worse than synchronous and single-threaded requests. Secondly, for the majority of users, the old Xeon is not that “secured”, the consumer i5 is worth it. Without overclocking. And even in general, the eight-core from AMD with rotten performance per core.
A minute of ageless humor about the effectiveness of "stones" from AMD
That is, a bunch of two 980 Ti in SLI suffers in games at 4k resolution is trivial due to high overhead: in some toys they do not have time to receive teams in proper quantity, are loaded at 50-80%, and even have to synchronize regularly for collaboration and correct display of the picture. As you understand, it is somewhat easier to assign tasks to one performer than to two, who also have to agree among themselves, who does what. Hence, microfreezes, and such performance results on real hardware, rather than on licked demo stands, where they tried to reduce such “hunger” to a minimum, using flagship consumer i7 with frantic frequencies of one core. However, there is a way out.
Well, firstly, do not take the second 980 Ti: the video card is not small, hot, it requires an additional ~ 300 watts of power (yes, 250 W are indicated in specs, but almost all non-references have an improved radiator and higher frequencies for a reason). Secondly, the era of DirectX 12 and API Vulkan is approaching, which drastically reduce such “tupnyak”, providing just a dramatic increase in performance (the increase can be fivefold):
Of all the games at the moment, if my memory serves me correctly, Vulkan normally works only with DOOM. And the titles with full support for DX12 ... In general, by the time it becomes mainstream, 980 Ti is unlikely to be more relevant. And in a year it will be very difficult to get rid of a pair of video cards - a loss in price, and new generations, you know ...
GTX 1070 in this case, eats less, has a memory margin and is not so hot. So if you still think, to take you a second 970/980/980 Ti in conjunction with the first one, or just sell your video card, report back a bit and buy the current generation - just take one, but fresh one. In the current games, the result will be better (the architecture of Pascal works wonders, we'll return to it), in the promising, due to higher frequencies and auto overclocking, I think it will be parity. API Overhead, of course, greatly affects the performance of current games, but you should not forget about iron improvements at a low level and reducing overhead costs even at the iron level. In general, about the magic of Pascal and the prospects of the game development - next time. Have a productive week.
Today our guest is two 980 Ti in SLI, one GTX 1070, a lot of dirty curses, 4k2k and the questions “how can this be”. In general, everything as always, it will be interesting, look under the cat.
Hardware: Maxwell and the game for all the money
The NVIDIA Maxwell architecture at the time of release was pleasantly surprised by the ratio of performance, heat transfer and price. At that time, it seemed that the next year or two was waiting for us to stagnate, and even two GTX 970 scandals (first with a “lack of” memory, then with crookedly indicated specs regarding ROPs) did not prevent her from becoming one of the most popular video cards in custom PC assemblies. Excellent 1080p performance, allowing you to play at maximum settings, good at 1440p, giving access to high settings, and 4k2k was quite playable if you were willing to sacrifice quality post-processing such as self-shading and any expensive hair type items from NVIDIA HairWorks or image blur.
')
To overpay for the 980th in the absence of 4k2k was a lesson ... not the most useful, the price difference did not cover the performance difference, but a 256-gb SSD for toys could be bought for the same money. The top segment was almost completely occupied by 970s from various vendors.
Approximately the same thing happened with the flagship assemblies, only instead of the 970th and 980th there were 980 Ti and GTX Titan X. On the second side there was twice as much memory and a slightly more efficient stuffing, but it was problematic to utilize all this power in games: Titan was like a lot of VR-enthusiasts, game developers and just people who have a lot of extra money, and remained: for 980 Ti asked for 650 bucks, and for Titan X - just a thousand. And the difference in real cases was at best 10 percent.
And if you had a lot of money, it was easier to spend $ 1,300 for two 980 Ti, than to take one Titan, which you had to find. Easier, it is easier, but ... There are already starting troubles. If the “mother” is old, and the video card with “high” cooling (above two slots) - SLI can simply not fit. Take the appropriate power supply and put it down. All other expansion slots are likely to be forgotten ... In general, there are a lot of nuances. However, it turned out that I had a suitable motherboard, and the opportunity to get two 980 Ti, and that is the best time when there is a 4k2k monitor on the table, and the GTX 1070 box next to it. the hole? : D
Match 2: Pascal and death to all living things
With the release of the first benchmarks of video cards on the Pascal architecture, the blogosphere, journalists and iron editions at first took a deep breath and then did not believe their eyes. Not only did Maxwell live only one generation of accelerators, not two (the “tick-to-do” strategy implemented by Intel at one time was successfully adopted by many manufacturers), Pascal simply broke all last year’s solutions, and with the release budgets (like the same 1060), sank more and B / C-iron market. Who needs the 970th warranty and with a 30-40% discount, when you can dig up a month and take 1060, which is faster in everything, has more memory, supports new technologies and will be new from the factory? Here are attached old video cards for friends, for nothing, or in the second computers at home. Yes, NVIDIA immediately stated (and did not lie) that the GTX 1070 will be faster than the 980 Ti, and 1080 will bypass the previous generation Titan X, despite the fact that the chips themselves belong to the consumer and not the prosyumerskoy line, and for the price are a replacement the usual "970 and 980, respectively.
Of course, journalists from the world of iron made benches, 1070 won 980 Ti, and peace-friendship-chewing gum, but ... here we have two 980 Ti. And the smell of fried!
Iron and video card specifications
So, the test stand. The pieces of iron are not the newest ones, but they still cannot refuse them with brute force.
CPU: Core i7-3930k @ 3.2 GHz
SandyBridge-E is not the newest architecture, but the 3930 has six cores, an unlocked multiplier, a huge cache, and 40 PCI-Express lines. The old horse will not spoil the furrow;
Motherboard: Asus Sabertooth X79
Up to 128 GB of RAM, 2 full-speed PCI-Express cards that allow you to install three-slot video cards, very evil power circuits ... In general, an excellent mother with a 2011 socket;
SSD # 1: 80 GB Intel 320
On this old man there is only a system, he should not influence the test performance;
SSD # 2: 480 GB Kingston OC400
High-quality disk, designed for installation in servers and workstations;
Video card # 1: Palit JetStream GTX 980 Ti x2
2 pieces NVIDIA GTX 980 Ti in SLI, automatic mode;
Video Card # 2: Palit GTX 1070 GameRock Edition
NVIDIA GTX 1070 with a custom cooling system and high frequencies.
A few words about the monitor
The test bench today is Philips 241P6VPJKEB . This monitor belongs to the "business" line, but formally I would call it a variant for workstations. The design has a healthy minimalism, the chips include light sensors, built-in sound, a USB hub and a swivel mechanism, and the characteristics ... Well, let's talk about this in more detail.
This display is built on the basis of a 24-inch matrix with a resolution of 4k (3840 × 2160 pixels). The matrix itself is eight-bit, but the controller allows you to display up to 10 bits per channel, mixing colors from adjacent frames (FRC, also known as frame rate control, if you are not afraid of English - much more is told here). In fact, most of the cheap monitors do this for displaying 8 bits (and the matrix is actually 6-bit). There is a factory calibration for SRGB (not bad, I must admit), but calibrated colors are incompatible with the operation of the light sensor.
According to the table of basic measurements one could say: look, look, what a terrible Delta E. Normal. First, the monitor does not belong to the line for professional work with graphics. Secondly, this is the factory sRGB calibration. In the absence of the user's interest, there are three things to worry about: the uniformity of the color temperature depending on the brightness, the uniformity of the displayed gamma, and the flickering of the backlight. The first problem is this: in 90% of cases, blue LEDs with a yellow phosphor are used. Depending on the brightness of the backlight and how closed the LCD shutter is, the permeability of the resulting “white type” spectrum is different. The second aspect suffers from the fact that the characteristics of the matrix itself are extremely non-linear, and the more precisely the monitor LUT converts the incoming signal - the more linear and correct the result will be in the measurements.
Here is the color temperature:
The bottom 10% (as well as the top 10%, however, too) can hardly be looked at, measured there as a problem, and the above features of producing “white” light introduce their own errors. In the remaining 80% of the coverage, the main working area - the temperature is slightly overestimated from the “nominal”, but stable.
Gamma in bright shades somewhat "stomps", but there, due to the non-linearity of the magnitude and not very relevant range, this will not be noticeable.
The color triangle combines well with the theoretical sRGB: a small excess in the yellow-green part of the spectrum and a dip in the blue is the merit of the backlight. In general, the factory settings for sRGB are 90% better than what regular monitors have in this price range. And who needs calibration and other accuracy, you know, either spend three times more on the monitor, or order a calibrator and bring this model to full coverage.
By itself, the display is pleased with the quality of the picture, the high density of dots + high-quality matrix make themselves felt, but so far not pleased with the thoughtfulness of the design. Up to the same HP or DELL with a normal arrangement of ports, a well-thought-out menu that can be used without constantly checking with the “legend” does not hold out. The connection ports of the signal cables in one place, the input for the USB cable are in another, and even my Type-B cable sticks out because of the frame ...
Approximately the same disease in the other ports: scattered on the monitor surface, as a result, even the cable management bracket does not help you to put things in order: all the wires stretch in different directions and create an untidy "beard" under the monitor.
As for me, attention to detail is not enough, not at all Apple. On the other hand, Apple does not have 4k2k monitors in such a diagonal, and if there were, the price would not have pleased anyone. The layout is simple: yes, usability for the top three. But the monitor is taken for the most part, not to configure, but to work. And here for this, everything is there. Factory calibration under sRGB provides 99% coverage and acceptable accuracy. If you are not engaged in work with photo / video editing - you have manual adjustment, make the image a little warmer, turn on the light sensor, eyes will not get tired. Matting is very neat, I don't want to cry from the crystal effect and noise. True, here a very high density of points also plays its role: the fonts are chiseled, with natural smoothing of the borders, very neat and visually pleasing.
If you want 4k, and there is no room for 27-32 inch models - you are not likely to be confused by such trifles, as if not the most successful cable layout or faddish menu. Once set up and forgotten, but the picture will still please the eye for a long time.
Results: obvious and not so
So, 980 Ti in SLI, fresh DOOM on the ultra-settings, the light is extinguished, the sound is louder ... damn, how beautiful he is. And gives from 40 to 55 FPS. And it is in 4k, yes with anti-aliasing! In statics is not impressive, but in the dynamics - breathtaking.
The new NFS (not so new, but still) on the FrostBite 4 engine produces an amazing, really amazing picture in dynamics. Damn, he is not bad in statics, if you choose a good angle.
Of course, in some places the scenery is simply not designed to be looked at point-blank, and in statics the maximum is “mediocre”, but on the go ... No, this is actually one of the most spectacular auto-simulators at the moment. Stable 60 frames per second (rested in vSync), subsidence is mainly due to an unstable Internet connection or when crossing two active races with 5-10 participants. After all, 20 cars in the frame, and even with a bunch of complex light and reflections ... And the picture is still good!
UPD: I don’t know if EA or Nvidia broke there, but just yesterday the system with two 980Ti couldn’t squeeze out more than 50 FPS, the real ones were about 35-40 + permanent departures of the application. Sad, but true.
Okay, go ahead. Forza Motorsport 6: Apex, available in the Windows Store .... Does not support SLI at all. One video card just kicks the bolt. As a result, drawdowns up to 25 FPS, besides, trendy “user-defined” coloring pages downloaded from the Internet have a habit of “disappearing” depending on the drawing distance. Nizakhot.
What is there more severe? GTA V? From 40 to 50 FPS, strongly depends on the situation, sometimes there are microfreezes or FPS drawdowns of up to 25 in just a fraction of seconds. Apparently, the point is in SLI.
Witcher 3 with all DLC, location Tussent. The abundance of grass causes FPS drawdowns even on serious configs, and here we are all super serious. To be honest, the creators of the witcher just scored on the optimization of top settings. Minimum and medium tweaks to play could be on relatively old machines, and high and ultra ...
In general, it is 980 Ti, which is two - the sense is not very much. FPS floats from 25 to 60, depending on what is happening on the screen, the work of NVIDIA HairWorks (this feature is generally better off) and the position of stars in the sky. It is unlikely that the computing power of the two almost top-end video cards of the current generation was not enough.
Turn off the computer, remove a couple of video cards, install the GTX 1070, update the drivers, check that everything works ...
Specifically, I had a GTX 1070 GameRock option from Palit, 1080 and 1070 outwardly differ a little less than nothing (stickers). Theoretically, the GameRock line is designed for modders and fans of customizable backlights: native software should be able to repaint the LED strip on board. I was not engaged in such garbage, I was more interested in performance.
So, DOOM. The same 45-55 FPS, but without jerks, much smoother and more evenly. If the FPS sags, it is only on the number of opponents, and not spasmodically, with the appearance of special effects or sudden movements / at random. With the Vulkan API enabled, the FPS counter freezes at 60 and does not move. WIN.
NFS gives out 45-60 frames per second, drawdowns are also caused either by other players or by problems with the connection: the game is networked. In general, of course, the FPS is slightly smaller than that of Sparky, but the game is ported from consoles, on which 30 FPS are generally stable in the last titles - the ultimate dream, and it is not very well respected, I saw frank lags during group drift on PS4, and graphics and next to not lying with PKshny.
UPD: With the latest drivers and recent changes, a single card produces a more stable result without crashes than two 980Ti in SLI.
Forza 6: Apex, by itself, goes better by 1070 simply because 1070 is faster than 980 Ti, and the game does not know how to work with SLI. There was nothing below 45 frames per second, VSync did not let work above 60 (120-144 Hz 4k2k were not distributed yet), and we didn’t want to observe frame breaks.
Minus Forza 6 is not even that the game is raw, but that there is nothing to do in it. A short campaign, a couple of races of a “seasonal” character and everything. Little content: no tracks, no cars on which I would like to ride. I hope, in due course will correct.
GTA V gave out from 45 to 57 FPS in real juz-cases and up to 60 in cut scenes, in statics and in rooms - the load in this case is lower.
Witcher is another game where the brute force of the 980 Ti in the SLI is still a dodgy novice, couldn’t give out more than 40 FPS in this mode of the GTX 1070. However, the minimum FPS was also higher, there were no drawdowns and uncomfortable lags, and the main guarantee of a pleasant game with such dynamics is not so much a high FPS, as even. So, one video card won out "morally".
The culprits of this result in the picture: large volumes of vegetation and a lot of animations of the surrounding world.
Why is this happening?
In benchmarks and on demo stands at journalists, 1070 shows excellent results, but bypassing 980 Ti in SLI? .. However, I have already given the answer to this question. Keyword - demostand. For testing use iron, which allows to minimize the impact of the performance of other components on the test object. If they test processors, they put on flagship video and top memory. They test disks - they connect with identical cables, identical iron, maximum spherical conditions in vacuum. Well, game tests and video card tests, as you guessed it, are produced on top-level fresh pieces of hardware, if this is not an article about building a PC as a whole. For toys of my processor, there is enough of a head, even the very first i7 still on socket 1366 with the architecture of Bloomfield and the ancient DDR3 1333 MHz memory are inferior in game tests from 10 to 25% of the flagship i7-6700k, which have higher core performance per clock and higher clock speed and speed memory. The dog is buried in a thing called an API overhead.
CPU dependency and API overhead
If you explain it in a very rude language and do not go into how the video card draws the frame “layers by layer”, then each time the video card needs to draw something: take a model, pull a texture, find a shader, some people have to tell the video card what to do . This one is a central processor. Since 90 (if not 99%) of PC gaming is tied to Windows, in most cases high-level DX10 / DX11 or OpenGL type libraries are used for games that are relevant for 2015-2016.
They make it easier and faster to work with 3D from the point of view of development, it imposes certain limitations and associated costs: in a simple language, you tell the video card what to do, it does.
On consoles, developers can write a much lower-level code, achieving good results from fairly weak hardware by simply reducing the overhead of drawing. Development becomes more difficult, the implementation of the code is tied to one hardware, but for consoles it doesn’t really matter - you didn’t need to choose any graphics settings or take into account different hardware (literally before the current generation).
These same overhead costs grow with the number of special effects, the resolution of the image, various post-processing tools, and so on. Even “insignificant” in terms of direct load on computing power, the Ambient Occlusion setting can lead FPS from a comfort zone to an uncomfortable one, because the video card will not give you a frame until it draws the shadows that objects cast on themselves. And these shadows are calculated through simplified ray tracing, which requires a lot of calls and the speed of the algorithm depends strongly on the resolution of the image.
So here. First, these calls to draw are not very well parallel. In some cases, in benchmarks, asynchronous requests work even worse than synchronous and single-threaded requests. Secondly, for the majority of users, the old Xeon is not that “secured”, the consumer i5 is worth it. Without overclocking. And even in general, the eight-core from AMD with rotten performance per core.
A minute of ageless humor about the effectiveness of "stones" from AMD
That is, a bunch of two 980 Ti in SLI suffers in games at 4k resolution is trivial due to high overhead: in some toys they do not have time to receive teams in proper quantity, are loaded at 50-80%, and even have to synchronize regularly for collaboration and correct display of the picture. As you understand, it is somewhat easier to assign tasks to one performer than to two, who also have to agree among themselves, who does what. Hence, microfreezes, and such performance results on real hardware, rather than on licked demo stands, where they tried to reduce such “hunger” to a minimum, using flagship consumer i7 with frantic frequencies of one core. However, there is a way out.
How to live further
Well, firstly, do not take the second 980 Ti: the video card is not small, hot, it requires an additional ~ 300 watts of power (yes, 250 W are indicated in specs, but almost all non-references have an improved radiator and higher frequencies for a reason). Secondly, the era of DirectX 12 and API Vulkan is approaching, which drastically reduce such “tupnyak”, providing just a dramatic increase in performance (the increase can be fivefold):
Of all the games at the moment, if my memory serves me correctly, Vulkan normally works only with DOOM. And the titles with full support for DX12 ... In general, by the time it becomes mainstream, 980 Ti is unlikely to be more relevant. And in a year it will be very difficult to get rid of a pair of video cards - a loss in price, and new generations, you know ...
GTX 1070 in this case, eats less, has a memory margin and is not so hot. So if you still think, to take you a second 970/980/980 Ti in conjunction with the first one, or just sell your video card, report back a bit and buy the current generation - just take one, but fresh one. In the current games, the result will be better (the architecture of Pascal works wonders, we'll return to it), in the promising, due to higher frequencies and auto overclocking, I think it will be parity. API Overhead, of course, greatly affects the performance of current games, but you should not forget about iron improvements at a low level and reducing overhead costs even at the iron level. In general, about the magic of Pascal and the prospects of the game development - next time. Have a productive week.
Source: https://habr.com/ru/post/398479/
All Articles