Here's what you can do with OpenCL directly in MetaTrader 5 without any DLLs - page 17

[Aliaksandr Hryshyn]

Renat Fatkhullin:

Yes, decided not to use old cards without double support at all.

The old cards in general are very bad, including bad unrepaired drivers that often crash, dropping the terminal.

In reality, we should clearly control the driver versions, as the developers of computer games do. Microsoft's default drivers will not get you very far.

double- how come?

The difference between double and float on the GPU is very large, and high accuracy is not needed everywhere.

[[Deleted]]

where to get it - OpenCL not found

-------------

I wanted to admire - the seascape

[Dmitry Nazarov]

Renat, first of all, I would like to thank you for the development.

When using OpenCL on the neural network, I got a 100x performance increase.

But as it turned out, this was just another step in the development of neural network trading.

To save power, earlier I had to limit the number of signals, neurons and use the so-called data window.

By cutting artificially needed signals we get not the best forecasts.

There is a need to use it step by step:

1. 7 cards on one motherboard at the same time.

2. A pool of hundreds of cards combined into a pool.

Who could manage this task?

Funding is negotiable.

[Maxim Kuznetsov]

Dmitry Nazarov:
Renat, first of all I would like to thank you for the development.

When using OpenCL on the neural network I got a 100x performance increase.

But as it turned out, this was just another step in the development of neural network trading.

To save power, earlier I had to limit the number of signals, neurons and use the so-called data window.

By cutting artificially needed signals we get not the best forecasts.

There is a need to use it step by step:

1. 7 cards on one motherboard at the same time.

2. A pool of hundreds of cards combined into a pool.

Who would be able to manage this task?

Financing is negotiable.

What prevents you from buying a bigger card?

A100 or something https://www.nvidia.com/ru-ru/data-center/a100/

[Dmitry Nazarov]

Maxim Kuznetsov:

What's stopping you from buying a bigger card?

some A100 https://www.nvidia.com/ru-ru/data-center/a100/

Basically nothing, except to first test the hypothesis that a better version of the library is possible and there are hundreds of shipping container based farms since mining 🤷🏼♂️

[Renat Fatkhullin]

Aliaksandr Hryshyn:

double- how come?

The difference between double and float on the GPU is very big, and not everywhere you need high precision.

Such a big difference between double / float is not on GPU, but specifically on Nvidia. On AMD everything is much better. Nvidia purposely killed the performance of double calculations on regular cards in this way, forcing them to buy a different class of GPU card for math calculations. This is deliberate behaviour.

While we require cards that support double for the sake of cutting out the outright junk, no one is stopping you from using float.

In financial maths, you can get incredibly screwed when using float because of the wildest loss of accuracy. No matter how anyone says "you don't need accuracy, it's enough", it has nothing to do with the reality in the financial world.

We ourselves do not want to transfer financial maths of profits even from double to OpenCL as we catch discrepancies in the distant digits. In finance, in fact, one cannot apply /fast maths (and in OpenCL/GPU it is always in the fast model), but only /precise models.

[Renat Fatkhullin]

Dmitry Nazarov:
Renat, first of all I would like to thank you for developing it.

When using OpenCL on the neural network, I got a 100x increase in performance.

But as it turned out, this was just another step in the development of neural network trading.

To save power, earlier I had to limit the number of signals, neurons and use the so-called data window.

By cutting artificially needed signals we get not the best forecasts.

There is a need to use it step by step:

1. 7 cards on one motherboard at the same time.

2. A pool of hundreds of cards combined into a pool.

Who could manage this task?

Funding is negotiable if required.

In MQL5's OpenCL interfaces you can use many OpenCL devices.

[Dmitry Nazarov]

Renat Fatkhullin:
In MQL5 OpenCL interfaces, you can use multiple OpenCL devices.

Manually assign a different device to each task in the EA ?

[Renat Fatkhullin]

Dmitry Nazarov:
Manually assign a different device to each task in the advisor?

Use CLContextCreate toscan the available contexts successively from 0, then use CLGetDeviceInfo to find out the device capabilities and decide which devices to scale to.

[Alexej Grebyonkin]

Renat Fatkhullin how do i contact you to discuss the possibility of writing in MQL5+OpenCL. i need to use the processing power of video cards. if i understand correctly, then using an example from practice: the written robot optimises only 11 parameters by 3 machines connected via local network, the period is only 1 year 6 hours. i tried to charge 5 years optimization with a full search of data, it showed me that 2 months have to wait.If I have understood correctly, OpenCL will solve the problem. The speed should increase by hundreds of times, since the calculations will not involve processors but video cards. It means that considering the entire trading system, there will be around 200-300 parameters in the settings.