Features of the mql5 language, subtleties and tricks - page 323
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
PS. We should make sure that there is no optimisation due to caching of the last calculations, because we have a very primitive array [1,2,3....100,1,2,....100,1,2............100]
please insert this line into line 117:
and double-check the result to make sure that caching has nothing to do with it.
on an RTX 4060
PS. We should make sure that there is no optimisation due to caching of the last calculations, because we have a very primitive array [1,2,3....100,1,2,....100,1,2............100]
please insert this line into line 117:
and double-check the result to make sure that caching has nothing to do with it.
on RTX 4060 with new line
13790 CPU
on the RTX 4060 with a new line
13790 processor
thanks! So there is no effect of caching on the result. A video card costing ~250 USD beats a very normal modern processor by 700 times in terms of computing power.
In short - everyone should upgrade their GPUs urgently. Including me :)))
and it's time to rewrite kanvas with OpenCL support
By the way, I even tried writing in Python with OpenCL for testing. Everything flies much cooler than native C++ programmes written without GPU.
I just had to mention: using graphics for calculations and CPU for rendering (and Blend2d is pure CPU) is to some extent a mega solution :-) With all the sympathy for Blend, honestly I thought that there will be a move to Skia, it is more hardware dense.
modern way
https://www.mql5.com/ru/forum/487541/page4#comment_56817317
thank you! So there is no effect of caching on the result. A video card costing ~250 USD beats a very normal modern processor by 700 times in terms of computing power.
In short - everyone should upgrade their GPUs urgently. Including me :)))
and it's time to rewrite kanvas with OpenCL support.
In 2D/3D, where matrix multiplications occur, switching to OpenCL is understandable. But in algotrading I don't see any problems. The main tool there is Tester. I don't understand how OpenCL can help.
ZY
Forum on trading, automated trading systems and testing trading strategies.
Peculiarities of mql5 language, subtleties and techniques of work
fxsaber, 2025.12.12 06:49 pm
Did you save the dialogue?
I'm dabbling in LLM on my RTX 3060 with 12 Gb, not enough power. Let's not talk about the CPU. Model 30B - 15 Gb, yes context 48K takes a lot of gigabytes just in VRAM.
If you are serious about local LLM, you should put 2 RTX 5xxx GPUs with 24Gb VRAM.
And if you are very serious - you should put 6 single-slot GPUs with turbines (now there are such novelties, expensive, bastard) and 3 PSUs. Only on regular miniPCI motherboards miniPCI are included in 2x lines, and this is a bottle neck. Not for inference (there is enough there), but for training - there is a big exchange of the card with CPU. I don't even know if there are motherboards that have enough lines for 8x for 6 cards in miniPCI. Mine has 2 cards.
And if it's serious, I'll have to splurge on a home mini-supercomputer. They say it can beat GPU-servers.
Of course, all this is on condition that there will be a financial result. Which, in my opinion, is not far off. In the meantime, we need to get ready for new technologies.
2. optimisation in the tester:
I'm dabbling in LLM on my RTX 3060 with 12 Gb, not enough power. Let's not talk about the CPU. Model 30B - 15 Gb, yes context 48K takes a lot of gigs just in VRAM.
If you are serious about local LLM, you need to put 2 RTX 5xxx GPUs with 24Gb VRAM.
And if you are very serious - you should put 6 single-slot GPUs with turbines (now there are such novelties, expensive, bastard) and 3 PSUs. Only on regular miniPCI motherboards miniPCI are included in 2x lines, and this is a bottle neck. Not for inference (there is enough there), but for training - there is a big exchange of the card with CPU. I don't even know if there are motherboards that have enough lines for 8x for 6 cards in miniPCI. Mine has 2 cards.
And if it's serious, I'll have to splurge on a home mini-supercomputer. They say it can beat GPU-servers.
Of course, all this is on condition that there will be a financial result. Which, in my opinion, is not far off. In the meantime, we need to get ready for new technologies.
I've already done the preliminary calculations. That a satisfactory GPU home server for research work is about 7-10 k USD. (for LLM - ~70 billion parameters)
And a normal server with LLM with 300-400 billion parameters is already ~70-100 k USD. But here you still have to manage to find a GPU cluster for 30 k
In all these subparts, I don't see a way to speed up. It's a sequential enumeration of ticks.
You will have to write your own tester instead of using the MT one.