OpenCl and the tools for it. Reviews and impressions. - page 10

New comment
 
tara:

Thank you, I saw it. Extending my hand. I will read it tomorrow.
See also this article: Optimal method of calculating the volume of aggregate position by a given magic number. It will also come in handy.
 
tol64:
Take a look at this article: An optimal method of calculating the volume of an aggregate position by a given magic number. It will be very useful too.


Thanks, I will have a look at it.

There is no such thing as an optimal method. Sorry :)

 
tara:


Thanks, I'll have a look.

There is no such thing as an optimal method. Sorry :)

Don't apologise. It's all relative. :)
 

Important note on hardware, PCI-E bus speed for OpenCL here:

https://www.mql5.com/ru/forum/6042/page6


There's a lot of video material on CUDA here:

http://developer.download.nvidia.com/compute/cuda/docs/GTC09Materials.htm

http://developer.download.nvidia.com/compute/cuda/docs/GTC_2010_Archives.htm

Of all this piling up, the easiest and most important for understanding the changes in programming for CUDA and OpenCL is a little-noticed video lecture by one of the CUDA leaders

Jason Sanders

numbered

GTC 2010 - 2131

(... can't find a direct link to the video, the English PDF outline is easily searchable on the web ... found)

http://us.download.nvidia.com/downloads/GTC_Videos/flvs/2131_GTC2010.mp4

http://us.download.nvidia.com/downloads/GTC_Videos/flvs/2131_GTC2010.flv

http://www.nvidia.com/content/GTC-2010/pdfs/2131_GTC2010.pdf

Here is the Russian abstract

http://sporgalka.blogspot.com/2011/10/cuda.html

 
It's all about the Cuda. But in principle, the programming technology is more or less collapsible. Nothing awful seems to be in sight. Thank you.
 
Yes, you can dive into optimisation and never come out.
 
TheXpert:
Yes, you can dive headlong into optimisation and never come out.


I would love to dive headfirst into all these optimization processes. But it would be cool when something more or less stable is dripping into my account. So, for the time being I'll put all these sweets off until I have a white streak in my life. I would immediately, if I had the chance, get myself a rack with 100 CPU clusters. Something like that:

 

Offtopic:

That's interesting. I've started looking for a better graphics card myself, although I don't really need one yet (options are HD 6790, 6850, 6930).

Must be the psychosis.

Sorry for the reflexivity in such a serious topic.

 
Renat:

Have you tried MQL5 Cloud Network with 2000 agents on a dual-core?

Or did you just run a single pass locally and rest on that?

It is with the MQL5 Cloud Network that we have achieved a simple acceleration of hundreds and thousands of times. Without GPU, on any Expert Advisor. Take my old example with video: Example of calculation acceleration using MQL5 Cloud Network


We have made it possible to instantly raise thousands of agents with one button (the warm-up time is about 20-30 seconds) for fast calculation of tasks. Instead of hundreds of hours, you can really get it done in half an hour.

Moreover, the prices are ridiculous - anyone can check them, and everyone who signs up at MQL5.com gets a bonus of $2.

This is a real revolution. Now any user can put almost any supercomputer to shrugs through the use of the cludes. And with the introduction of the GPU (a build will be available on Friday) it will be possible to beat supercomputers in packs.


The performance is enough (or rather you convince yourself of it) for simple cases. The performance is not nearly enough for the most massive and detailed calculations.

The mistake with comparing MT4 vs MT5 testers is that people turn a blind eye to the vastly different level of detail and ability to accurately multicurrency test.

You're wrong about dependency on hardware:

  • MT5 supports both 32 and 64 bit architectures (the 64 bit version is actually faster and more powerful)
  • MT5 uses all local cores
  • MT5 uses remote agents (32 and 64 bit)
  • MT5 uses MQL5 Cloud Network when a simple laptop receives calculation results from computers several thousand times more powerful than the laptop.
  • MT5 uses hardware-independent OpenCL, which allows you to work with a variety of GPU processors

You are blatantly wishful thinking, clearly aware that you are wrong.


While I am not a GPU advocate (have been outspokenly against it for a long time), your conclusions are wrong.

There is no software for GPU, because developers have not yet believed in this direction, but no repeated reference to the mythical binding to a specific hardware. For the sake of decency, you should at least read what OpenCL is and what it was invented for (hardware and platform independent).

Companies are only forced to use software rendering because of inertia and established processes. Getting a company to change software is a titanic effort, often impossible for the next 3-5 years. Given the leap that GPUs have made in the last 3 years and the clearly visible emphasis by their authors on universalization, it is reasonable not only to expect improved results, but also to actively pursue new features.


We've waited long enough, criticised and disagreed too, but the time has come.

And this is the third time the repetition on 'hardware-dependent software' has taken place.

 
Yura, where in the previous post is your answer? I don't want to go back and look for the differences.
1...34567891011121314151617...29
New comment