OpenCl and the tools for it. Reviews and impressions. - page 20
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
I read old ixbt.com archives ("fan" threads) related to the time of AMD and ATI merger. Here is such post (excerpt from it):
Думаю, что в основе лежит идея АТИ использовать графический процессор для решения физических задач. Если такой комбинированный спецпроцессор оснастить двухканальным контроллером памяти и ХайперТранспортом, его можно ставить в АМД4х4. Получится домашний суперкомпьютер. Благодаря НТ сильно уменьшатся задержки между CPU и графическим адаптером, что ускорит игры. Также всякие расчёты будут идти в несколько раз быстрее, чем на универсальном процессоре.
This post was published 5.5 years ago and nobody could have known about Llano stones (and OpenCL) back then. But we do know: buying ATI turned out to be a strategic decision of AMD.
And now - let's look at the AMD A8-3850 CPU and especially the characteristics of its graphics core(here - briefly):
As a result, they managed to cram as many as 400 (in the A8 family) or 320 (A6) graphics processors under a single heat dissipator cover. The first one is already the level of Radeon HD 5570; only clock frequency is 50 MHz less.
Below in the table of the same article the A8 processor graphics core is listed as Radeon HD 6550D.
Of course, there are not 1280 conveyors like in HD 6930, but 400 is not bad at all comparing to Intel HD 2000 integrated graphics.
So, one wonders if it would be worth it to try running OpenCL code on APU integrated graphics unit without any discrete vises. Although, the memory would be DDR3 rather than GDDR5 and the bus width is... ...we don't know what it is. But still I doubt very much that AMD had not thought of such way of using their stones, when making Llano.
P.S. Frankly speaking I've never been interested in CPU from AMD, but in light of recent Metaquotes turn to GPGPU calculations I've somehow started to think about it.
P.P.S. Good news: Next generation Intel processors will support OpenCL.
P.P.S. Good news: The next generation of Intel processors will support OpenCL.
This means that there is no point in buying graphics cards with OpenCL as the slow data transfer bus from the CPU to GPU and back will in many cases negate the performance of the computation. If the GPU sits on a high-speed bus with the CPU, this drawback will be eliminated.
Moreover, it is likely that the graphics cards will be empty, i.e. they will not be equipped with GPUs at all, as was once the case with the so-called software modems, i.e. processorless devices. Presumably, the graphics cards will be built into the motherboard so that everything goes directly without additional slots.
Anyway, who knows, but I'm in no hurry to buy GPU compatible graphics cards because I don't play games and in MT5 I can't draw anything except Mandelbrot fractal at the moment:
1. the tester does not support any OpenCL at all, i.e. it is impossible to ramp up the optimization speed by upgrading GPU hardware, but only through CPU and RAM clock frequency
2. Developers still question the possibility of accelerating calculations for TA indicators and oscillators, as vector calculations have their own features and not all tasks can be performed using them.
This means that there is no point in purchasing graphics cards with OpenCL as the slow data transfer bus from CPU to GPU and back in many cases negates the performance of calculations. If the GPU sits on a high-speed bus with the CPU, this drawback will be eliminated.
Moreover, it is likely that the graphics cards will be empty, i.e. not bundled with GPUs at all, as was once the case with the so-called software modems, i.e. processorless devices. Presumably, the graphics cards will be built into the motherboard, so that everything goes directly without additional slots.
http://www.thg.ru/cpu/amd_apu_llano_a8_3500m/amd_apu_llano_a8_3500m-01.html is a serious review of AMD's APU plans and current capabilities.
Still, Yura, I wouldn't be so categorical. I don't play games myself, and I'm not going to build myself a desktop on AMD APU yet. So what am I supposed to do? Buy a video card.
Fantasy. In the future AMD will have an APU with a more or less competitive CPU and a really powerful integrated GPU on a level no lower than today's HD 68xx-7xxx, and GDDR5 memory... external. You won't get far on DDR3, even on the fastest one.
And Intel is not going anywhere: the competition has not disappeared, it has just shifted to another area. Now AMD has very little chance to overtake Intel in CPU - given their financial possibilities and technological lag. But they are definitely ahead in integrated GPU. And to make graphics with sane heat dissipation (not with present 300 W on powerful graphics cards but at least 5 times less), they will have to work hard.
Tell me, Kisa, as an artist to an artist, can you draw? (c) Ilf and Petrov
The most important question is, what do you buy a video game for, other than gaming?
1. To draw fractals 100 times faster? I don't need them in hell.
2. to learn OpenCL, which is not supported by strategy tester? I don't need it either.
Answer this question: what is the purpose of the video card in autotrading and everything becomes clear.
2. learn OpenCL, which is not supported by the strategy tester? I don't need that either.
What this means is that there is no point in buying OpenCL graphics cards as the slow data bus from CPU to GPU and back, in many cases, negates computational performance. If the GPU sits on a high speed bus with the CPU, this drawback will be eliminated.
In general, I don't know, but I'm in no hurry to buy GPU graphics cards, because I don't play games and in MT5 it's impossible to draw anything except Mandelbrot fractal so far:
1. the tester does not support any OpenCL at all, i.e. it is impossible to ramp up the optimization speed by upgrading the GPU hardware; only the CPU and the RAM clock frequency can be used to do so
In the tester, support has simply been disabled by one condition, as at that time the MQL5 Cloud Network did not take into account the OpenCL status of agents.
In the next build we will enable it - no problems at all. We have already made the necessary changes in the cloud.
2. The developers are still questioning the possibility of accelerating calculations for TA indicators and oscillators, as vector calculations have their own characteristics and not all tasks can be performed using them.
Yes, this is a big problem - the vast majority of algorithms cannot be accelerated on GPUs in principle.
But most importantly, there is a strong trend of universalization in the development of GPU vendors. If 2 years ago it was not visible, a year ago there were doubts about their goals, today the goals and efforts shine a spotlight in their face.
We must not skip this direction.
And Intel is not going anywhere: the competition has not disappeared, but simply shifted to another area. Now AMD has very little chance to overtake Intel in CPU - given their financial possibilities and technological lag. But they are definitely ahead in integrated GPU. And to make graphics with sane heat dissipation (not with present 300 W on powerful graphics cards but at least 5 times less), they will have to work hard.
It is very good that manufacturers have started to warm up to competition in the new field. Otherwise, technological development and competition in the CPU area has been stagnant for years.
If Intel's CPUs really start supporting OpenCL natively, it would mean a rapid arrival of OpenCL in common software.
By the way, we have made a huge breakthrough in the universalization of tasks for agents and the entire MQL5 Cloud Network. You will see the changes in the next builds.
In fact, MetaTrader 5 has become (wait for 2-3 builds for finalization) a universal calculation platform with unlimited scaling. This change is also from the same area of GPU transition to universal calculations. Many people will say it's unnecessary, but to us the goal is clear and the point is made. Look at the effects in the next 5-10 years.
I'm making an EA now. The calculations can actually be done on every tick (very small), but only due to the fact that I put the hardest part in init().
On rather fast i3-2120 these calculations are done in about 10 seconds. There's a simple, very simple loop with a lot of iterations (about a million). It is hoped that there will be a speed boost on OpenCL. And there is no escaping init() at the stage of constant recompilation. Still, you'd better wait for, say, only 0.2 second instead of 10.
For trading, this optimization is of no importance (I can tolerate 10 seconds once when launching an EA), but at the stage of development it is a decent saving of time and nerves.
just like Elder: "The trading system should fit on the back of a postage stamp", but technical analysis is not what it used to be, it's time to break ticks into quanta )))).