I want to build a PC to work with MT5, what do you advise and why? - page 64

New comment
 
lynxntech #:

You buy 2 screws there, one I understand for the system on 500 was, so instead of it you should take SSD.

I agree with that, especially if it will be used for user work.

lynxntech #:

and do not expect that they will restore HDD in the service, they will rip off three skins and restore one porn).

Just, usually, HDD dies with symptoms, slowly, which allows you to react and copy the necessary data.

lynxntech #:

I always do another one on a flash drive and sometimes I plug in an ancient screw

Yes, of course you should make copies, and I do, just not as regularly as I should.

 
Aleksey Vyazmikin #:
when large amounts of data are pumped from RAM memory modules to RAM GPU.

Here I understand that memory bandwidth (memory frequency) is important.

it would be interesting to read how it works there, how much memory GPU needs, it is quite small in relation to the system memory.

 
Igor Zakharev #:

Thanks for the info, didn't know that.

Is there any such information on laptops?

I wanted to upgrade (now I have 2 laptops, one of them is gaming and powerful, but it is almost 3 years old, it has aged here on the service, you can say - I took it for MQL5.com :)). ), maybe by the end of the year, I'm looking at one like the one below (G16). The news about new Intel processors slowing down SSDs and all that is alarming. And the price is still too high, in half a year it should settle down. I'm used to Asus laptops, I rarely use stationary computers.


...

You have to look at the performance gain compared to the previous ones. If it is insignificant, there is no point in overpaying :)

For this money you can get a f*cking mac pro

...
 
lynxntech #:

Here I understand that bandwidth (memory frequency) is important.

It would be interesting to read how it works there, how much memory GPU needs, it is very little in relation to the system memory.

They write that PCI 4.0 bandwidth is about 32 gigabytes per second (if 16 lanes). Only top-of-the-line GPU cards have 16 lanes, as I understand it.

DDR5 4800 is 38.4 GB/s - so it looks like RAM frequency won't affect the video card's data transfer rate here.

In general, the code is written for OpenCL in such a way that as little transactional actions as possible are done.

 
Aleksey Vyazmikin #:
DDR5 4800 - 38.4GB/sec.

DDR5 6800C32 - 100gb per second on Intel, I have 2x32gb

+ Latency with 6800 is about 25% less.

Asked a question to neural network, the answer in C++ OpenCL using GPU the bottleneck can be RAM, video memory on modern cards is 200+gbps


 
lynxntech #:

Asked a question to neural network, the answer in C++ OpenCL using GPU the bottleneck can be RAM, video memory on modern cards is 200+gbps

Well, you can believe neural networks.... you didn't tell it that you will transfer large data arrays to the video card from RAM - maybe then it will figure it out correctly.

 
Aleksey Vyazmikin #:

Well, you can trust neural networks... you didn't tell her that you will transfer large data sets to the video card from the RAM - maybe then she'll get it right.

The neural network has given example codes in MQL5, I'll read them later, I'll try to test them and show you what happened.

from a quick look, the speed between RAM and GPU is limited by PCI, and the preparation before transferring to GPU hangs on the speed and latency of RAM.
 
Maxim Dmitrievsky #:

You have to look at the performance gain compared to the previous ones. If it is insignificant, there is no point in overpaying :)

For this money you can get a fucking mac pro

I have no experience with macs, alas. So I wouldn't risk buying a mac. And in terms of performance there is a 100% increase only on CPU, and on GPU I do not know exactly, depends on what you get. 5090 would certainly be desirable, then OpenCL will work properly.
 
Michael Charles Schefe neural networks and compute massive data sets, ea needs to code and hand that task over to the gfx card. Your gfx card does not automatically receive the large data set it is sent to solve the task. If you haven't encoded ea to pass the data to the graphics card, the main system will do more than 90% of the task; that means you just spent over a thousand dollars on that graphics card, plus the cost of the power supply to power it.

It's a little unclear what you mean. Are you talking about the features 

CLContextCreate( arg);

CLProgramCreate( arg, arg);

CLKernelCreate( arg, arg );

CLBufferCreate( arg, arg, arg);

CLExecute( arg, arg, arg);

and others? https://www.mql5.com/ru/docs/opencl/clprogramcreate

Almost everyone knows how to program, there is no question about it, and there are articles about it. We are discussing what will speed up the work of programmes using OpenCL among other things. I didn't even put a lot of experiments, as my RTX solves complex tasks for a long time, and RTX 5090 is another level. But nobody has shared examples on the forum about how 5090 works on practical tasks (speed of neural networks training on MQL5) in comparison with other accelerators.

New NVidia RTX 50-series look very attractive in terms of performance and features.


 
Igor Zakharev #:

It's a little unclear what you mean. Are you talking about the features

yes i was. so it seems that you have that covered.

I dont know if a workstation gfx card will be better than a gaming gfx card, but have you researched that also?

1...57585960616263646566
New comment