User experience feedback on the language OpenCL - page 2

New comment
 
Je voulais dire plutôt pour l'optimisation de mon EA puisqu'un article MQL5 précise que je gagne 42 fois plus de temps avec une carte NVIDIA
 
Lorentzos Roussos #:

i think you are overreacting a bit sir  😇

Not at all, just commenting.
 
David CAUJOLLE #:
Je voulais dire plutôt pour l'optimisation de mon EA puisqu'un article MQL5 précise que je gagne 42 fois plus de temps avec une carte NVIDIA

You could start here , the video is from mql5.com's collection for OpenCL (there is a thread) 

This account has 3 more videos on it , i would watch these in sequence first .


 
Many thanks Lorentzos for the videos. My English is not at the top but it should enlighten me a little more on OpenCL.
A simple question, I have a doubt by reading some comments ? Can you confirm that I will save time on my optimizations by programming my EAs with OpenCL ?
If so I wonder why so few traders do not use OpenCL because the big problem with optimizations is its calculation time. And using the NetWork cloud can quickly become very expensive. So OpenCL programming could be a great alternative.
 
David CAUJOLLE #:
Many thanks Lorentzos for the videos. My English is not at the top but it should enlighten me a little more on OpenCL.
A simple question, I have a doubt by reading some comments ? Can you confirm that I will save time on my optimizations by programming my EAs with OpenCL ?
If so I wonder why so few traders do not use OpenCL because the big problem with optimizations is its calculation time. And using the NetWork cloud can quickly become very expensive. So OpenCL programming could be a great alternative.

I haven't tried to optimize in the backtester with openCL . If your calculations are that heavy per bar or per tick you can check what you can do before even touching open CL to speed things up.

They don't use openCL because its difficult . There will be a lot of "wait a minute" s even if you know openCL where you will realise :

  1. i can see if moving this 'outside' speeds it up anyway
  2. i can see if loading this instead of crunching it speeds things up etc .

It sounds boring to test what can you speed up before openCL but if it sounds boring now with mql5 -to which you know your way around- imagine that process and having to care about how things flow in and out of your GPU on top.

You might even consider other avenues . 

For instance , if you are not running all the possible combinations of your parameters you are using the built in genetic algorithm of the mt5 tester .

A genetic algorithm will approach "the genes" or the "settings" that have the best results .

A neural network will try to "bend" the hyperdimensional space of what you gave it so that it will make it work as accurate as possible . 

I see that pyTorch also has genetic algorithms so if you could export some things to python ,and could emulate trading activity as well as mt5 tester does , you would get a superior tester , faster and with a bigger -and better- support community around it. I mean you are concerned with speed , that is where your nVidia chip will give you its all , a pyTorch tester . 

ignore the strikeout text , i thought i read pytorch , it said python ... low caffeine mistake . sorry .

 
David CAUJOLLE #:
Many thanks Lorentzos for the videos. My English is not at the top but it should enlighten me a little more on OpenCL.
A simple question, I have a doubt by reading some comments ? Can you confirm that I will save time on my optimizations by programming my EAs with OpenCL ?
If so I wonder why so few traders do not use OpenCL because the big problem with optimizations is its calculation time. And using the NetWork cloud can quickly become very expensive. So OpenCL programming could be a great alternative.

Because it's far from obvious to switch from serial/linear calculations to parallel calculations, it's not always appropriate. It requires high coding skills, and time investment.

There is a high risk that the time you could potentially win on backtest, you will have to lost it on development.

 
Alain Verleyen #:

Parce qu'il est loin d'être évident de passer des calculs sériels/linéaires aux calculs parallèles, ce n'est pas toujours approprié. Cela nécessite des compétences de codage élevées et un investissement en temps.

Il y a un risque élevé que le temps que vous pourriez potentiellement gagner sur le backtest, vous le perdiez sur le développement.

If it's only the coding part, I don't think I'll waste time because I'll be using a freelance coder. I still have to find a good one in OpenCL, which is not a priori obvious unless I pay more perhaps? Which I don't mind if I really have to save time. However, once I get my code, does it work like a regular EA? I don't know at all and I don't want to fall into a gas plant and waste my time.

 
David CAUJOLLE #:

If it's only the coding part, I don't think I'll waste time because I'll be using a freelance coder. I still have to find a good one in OpenCL, which is not a priori obvious unless I pay more perhaps? Which I don't mind if I really have to save time. However, once I get my code, does it work like a regular EA? I don't know at all and I don't want to fall into a gas plant and waste my time.

Do you want my opinion ? Forget it.
12
New comment