The nearest future of video cards

He recalled a specialist in the national economy, who even defended his doctoral dissertation, and used a telephone book and several old folk songs as sources for it.
- M. Larni

In this article, using Wikipedia and a few pictures from the Internet, I will try to assess the near future of graphics accelerators.

Let's look at the current S3 Trio video card 20 years ago.


')
Seriously, let's rate it.

1. The minimum electrical power of the order of watts.

2. From item 1. the absence of a core power system follows, the board is very compact.

3. It also follows from p.1 that there is no need for a powerful cooling system. You do not need a huge radiator, as well as fans. Complete silence.

4. Modular memory. Everyone can buy as much video memory as he needs, and then install it on a video card. You can buy video memory when there will be money or when you need to expand its volume.

5. Video output is carried out through VGA, which was then standard. One type of signal transmission, one output, etc.

6. The video card is connected via a PCI slot that is even now used. This video card can be connected to any computer with such a connector even now and it will work.

7. There are almost no radio components on the board that are damaged or cause problems. There are no powerful chokes that can start whistling, there are only a few capacitors - and those with a minimum capacity.

Let's now look at one of the current top, namely the MSI GTX 1080 Ti Lightning Z.
The length of the video card is 32cm



It takes 3 slots, which in general has become a phenomenon, not surprising at the moment



The video card uses three standard PCI-E power connectors.



and in theory consumes only 250W, which causes some doubts, judging by the power and cooling system. Of course, the cooling system should be mentioned separately.



Only 5 heat pipes do not seem to be something exceptional. Well, you have already seen 3 fans.

If someone has S3 Trio and GTX1080ti lying around, I suggest to conduct a simple experiment - erase the dust from the first one, align the capacitors that were crumpled while the video card was lying in the garbage. Familiarize a retired person who is absolutely far from computers with both video cards. And then ask what card, in his opinion, is newer. If the rust on the S3 Trio does not let you down, the answer will be in its favor.

How did we get to this?

It is obvious that the growth of the video card's electrical power has led to such a critical increase in board size and heat generation. The core requires more power; to ensure it, it is necessary to design an increasingly complex power subsystem, since core voltage is very small. In addition, the heat released has to be dissipated, which in turn requires an increase in the efficiency of the cooling system. Since scattering hundreds or more watts (remembering a soldering iron at 60W), and even in the limited space of a hypothetical case, is not an easy task, you have to increase air flow (increase in the number of fans), heat exchange area and thermal conductivity of the radiator itself (increase in the number and size of heat pipes). When it started? The first generation of common graphics accelerators did even without radiators, as we can see from the example of the S3 Trio. With the further development of video cards, installation of small radiators was needed.


ATI Rage 128 Pro 32MB


By the very beginning of this century, many video cards in the line were equipped not just with a radiator, but already with a radiator with a small fan.


ATI Radeon 8500LE 64Mb

If someone experimented with installing small fans anywhere, then he is well aware of their extremely low efficiency. The air flow due to the size of their blades is negligible, and such a fan creates a lot of noise due to a decent speed of rotation. Since at the same time, the power consumption of video cards continued to grow, even in the reference (exemplary) video cards we had to switch to turbines. After about 2010, turbine-type cooling became standard for exemplary video cards, and although we see it live less and less, video chip developers themselves present their default video cards with it. It is interesting to note that at the beginning of this century, the turbine, as a type of cooling, could be sold separately for those who want to upgrade their video card with an ordinary small radiator. Now they, of course, no longer sell. Actually, why the turbine?


Typical turbine type cooling system with casing removed on GTX670

As you can see, the turbine itself captures air from above, redirects it to the sides, after which the air passes through the radiator and cools it. Ideally, hot air is completely thrown out of the enclosure. It should be noted that the turbine will hardly work with the casing removed. The theoretical limit of power dissipation by the turbine is quite large - it is enough to increase the number of radiator fins, attach a pair of heat pipes to it, and even 450W of heat can be dissipated (HD6990). However, the main disadvantage of turbines is noise, since the turbine itself must rotate at high speeds. Therefore, approximately since 2008, production of air-cooled cooling systems, a relatively compact radiator, and several full-sized fans blowing in the direction of a video card through a radiator has begun. A typical example, one of the first swallows


Thermalright HR-03 at 8800GT

At the same time, this type of cooling has already become quite standard - it’s strange to buy a video card with a turbine, in fact. So the situation repeats. Initially, turbines were sold as a custom cooling option, then became the standard. Then radiators on heat pipes with a fan repeated their fate and became the de facto standard. It remains to hope that it will not come to a compact SVO on each video card by default.

This is only an approximate estimate of the growth of power of video cards, which is called “by eye”. Now I propose to make an assessment on more stringent criteria, namely on the charts. To build this and the following graphs, I used for the most part data from the wiki on AMD and NVIDIA video cards, however, I filled in some data myself based on the data stored on the Internet. Unfortunately, the TDP of the old video cards before 2007 is unknown. The graphics include almost all GeForce and Radeon series in about 10 years, except for the 100th and 300th GTX series and the 8000th HD Radeon series. The points on the graphs belong to video cards, the date of the video card output is plotted on the abscissa axis, and the change of those or other values ​​is plotted on the ordinate axis.
So, let's take a close look at the graph of changes in the heats allocated by video cards at work over the past 10 years.



The following conclusions suggest:

1. The reduction of the technical process did not produce the expected revolution in heat dissipation of graphics accelerators.

2. Despite the large technical process, the coldest graphics cards were produced until 2007. The first major jump in heat generation occurred approx. mid-2006 - early 2007. This clearly indicates that heat release is primarily determined by market demand. When the video cards themselves are already badly pulling the game, manufacturers have to overestimate the frequency and voltage, which leads to a very strong increase in heat dissipation.

3. The maximum heat release on the plateau occurred approx. 2009, plus or minus.

4. The plateau of maximum heat generation of powerful models has been holding steadfastly since approx. 250W. This, apparently, is the maximum that the objective properties of a video card as a device in its modern form allow. If it were possible to further increase the heat generation, then, no doubt, video chip manufacturers would continue to increase the frequency and voltage. However, cooling a video card with heat above 300W is already a difficult task, and it will fry in a bad case, and the noise of the fans or the turbine will be above all standards.

5. Medium and weak models also slowly but surely increase their heat release. If, however, the older models have rested against a maximum (plateau) around 2009, then the average models have rested at about 150-180W even earlier. The weak go the same way.

6. There has been a decrease in the number of models presented since 2012. This is especially noticeable regarding models with low heat dissipation. They are getting smaller.

7. It is impossible to write off the last rather long stagnation of attention to the market of mobile solutions, as it was done before. Moreover, just at the moment of maximum popularity of tablets, there is a rather high frequency of new video cards.

8. It is no less paradoxical that it seems that during the years of the global economic crisis, the release of new models took place much more often than now.

9. The last two conclusions again lead us to the dominant role of market demands in front of the objective capabilities of video cards.

Well, maybe we knowingly tolerate an increase in heat generation, surely this is justified by an increase in computational power? Let's try to check it out. Of course, there is no way to unambiguously find the exact performance of a video card, so I turned to the maximum theoretical performance for single-precision calculations.



It is quite nice to see quite exponential growth here. For me personally, it was a surprise that the theoretical performance of the Radeon video cards was noticeably higher. Well, probably an old joke about the lack of a driver



turned out to be true. It is possible that the matter is in the optimization of the games themselves.

Now let's compare the graph of TDP and performance. It is easy to come to a conclusion confirming clause 2:

10 The market requires an exponential increase in productivity, despite the fact that video chip manufacturers are obviously lagging behind this requirement. This is indicated by an increase in TDP.

Now why don't we relate performance to power.



To say exactly what function here sets the curve is difficult. On the one hand, the GeForce seems to have a power function, and the Radeon curve is S-shaped. If we discard the optimization of the driver and games, we find that AMD is ok. 2009 has achieved a good-performing architecture, which has been used for many years and the latest models of which were released in early 2016. This is a plateau on an S-shaped curve. Being well-versed in the video cards coming out, it is easy to understand about which models we are talking about. At NVIDIA, there has been a jump in efficiency since 2012 - most likely, the adoption of some completely new architecture. The GeForce is followed by an Radeon exponential jump since 2015.

I propose now to turn to an increase in the number of transistors. Let's try to check the law of Moore.



Incidentally, by the way, it will be mentioned that the so-called. Moore's law is only an observation that was made in 1965, i.e. at the dawn of the modern electronics industry. And the observation was made on Intel processors. The law itself says:

The number of transistors placed on an integrated circuit chip doubles every 24 months.

Problems with long-term prediction by Moore's law at the point of reference, namely, how many transistors to take. I took 110 million for a medium-weak video card. As you can see, Moore's law is satisfied, the number of transistors is growing, as it should. But here it would be appropriate to ask the question - is it growing due to the area of ​​the crystal? We look at the size of the crystal (or their sum, if we are talking about dual-processor solutions)



In this case, we see a pleasant constancy. Manufacturers trite unprofitable to increase the area of ​​the crystal. But I would like to remind that with progress it would be desirable to see a decrease in the area of ​​the crystal. Let's try to draw conclusions:

11. The ratio of performance to TDP depends almost entirely on the current architecture (or on some not yet obvious factors).

12. Architecture, in spite of marketing changes of names from generation to generation, in fact, in reality, fundamentally change rarely. Not more often than once every 6-7 years.

13. The use of one or another architecture is very clearly seen from a curve describing the relationship of performance to TDP.

14. Most of the changes in the ratio of performance to TDP cannot be described by power functions. Alas, these changes do not have the form of a linear function and, in general, can stand in place for years almost unchanged.

15. The technical process does not affect the ratio of performance to power, because reducing the size of the transistor is accompanied only by the occupation of its place by new transistors. This is related to paragraph 10, i.e. with inflated market demands.

16. Squares of crystals do not change.

17. The increase in the number of transistors continues to conform to Moore's law.

And now let's look at the most interesting, namely the price. Since In Wikipedia, release price data was scarce, I had to search for them on the Internet, but this is still not a big secret. Since the price is a very blurry thing (the release price may differ from the soon-established market, the release prices themselves are not always available, etc.), its inaccuracy requires the use of a more diffuse marker.



Summing up and here:

18 Price is slowly increasing.

19. The general trend of rising prices can be correlated with US dollar inflation ( thanks to American and British economists ). For the mass of models, this can be explained very simply - chip companies come from Silicon Valley. Inflation of the American national currency, of course, will affect them.

20. However, individual plateaus form the highest models of video cards. In particular, for NVIDIA, we see several S-shaped price rises - prices for tops rose in early 2012 and by the summer of 2016. AMD, having switched to the pre-top model range from the HD7970 era, does not have such jumps.

21. Separate pricing points for dual-chip solutions do not deserve separate consideration by themselves. However, at the end they were joined for the first time by a single-chip solution - NVIDIA TITAN V.

Summarizing the above, we will try to create a forecast for the near future of graphic accelerators. Yes, only a fool can climb into the future, and any more or less self-respecting person never descends to forecasts (for they will surely make a mistake in them, anyway). Well, if it makes them, it is extremely blurry to get, as they say, with a finger to the sky. However, I am willing to take a chance:

1. TDP video cards will continue to increase for weak and medium models. This will be a very slow process, but most likely inevitable.

2. For older and older-medium models, TDP will remain ok. 250W - an objective limit for modern cooling systems. A further increase in heat generation will simply require the installation of an unnecessarily expensive cooling system, and even reduce the lifetime of the devices.


3 In the performance of video cards, market requirements are primary. One way or another, the video cards do not satisfy or satisfy them, but with difficulty. If we assume that the market does not reduce its requirements, then expect a decrease in TDP, a reduction in chip size, etc. can not.


4. Increasing the number of transistors and improving the architecture will not improve the ergonomic properties of video cards, but will completely go so far as to catch up with market demands for increased performance.

5. Most likely, you should prepare to reduce the model range of video cards. In addition to commercial moves, chip makers will develop only a few cores, which, depending on their qualities, will already be sorted. This will be done to reduce development costs.

6. The price of video cards will slowly grow, it is possible that literally one lower model will be left at the lowest price (for example, the younger video card of the previous generation), but the price increase in dollars will continue.

7. Of particular concern will be the prices for strong and medium-strong graphics cards. Since NVIDIA has long been on the plateau, we can expect the next jump in prices for the above mentioned video cards in the new generation, or the next. Be that as it may, the jump in prices for tops is just around the corner.


So, nothing hopeful is expected in the near future. Only a naive person can think that he will get the performance of the next generation GTX1060 for the cost of the GT1030. In short, the forecast can be expressed simply by “holding onto our vintage tops and expecting the best.”

A small addition:

Thanks to Edil0 for the idea. The ratio of theoretical performance for single-precision calculations to price.