OpenCL: internal implementation tests in MQL5 - page 7

[Vladimir Gomonov]

AlexEros:

Don't buy Gigabyte in any case.

For toys it will work, for bitmining, multi-gpu bitcoing it will work, but for MT5-GPU it will not work.

Here's the thing:

Scientific calculations in general and MT5-GPU in particular need high PCI-E bus speed. You'll be screwed if you buy a Gigabyte motherboard with a multi-slot PCI-E bus ..... 2x-4x slower. Check your own link: it says

"Additional info PCI Express x16 slot mode - 2 slots work in x8 mode, 4 slots work in x1 mode; 1 x PS/2 port - combined".

Look carefully at the mainboard specifications in terms of PCI-E speed limit - it has to be x16 at least in one slot.

Be thankful that at least here Gigabyte has written honestly, and actually it has cheated a lot of people like that before.

There is no decent free software for testing PCI-E data transfer speed for OpenCL, but there is for nvidia CUDA.

It's CUDA-Z.

http://cuda-z.sourceforge.net/

CUDA-Z speed results, along with oddities due to PCI-E speed:

http://forums.guru3d.com/showthread.php?t=328274

There memory speed should be at least 1000 MB/s for ""pageable" and 2000-5000 MB/s for "pinned" memory access types between CPU-GPU over the PCI-E bus. (There are two software-defined types of access to the GPU memory via PCI-E bus in CUDA that are fundamentally different in terms of speed, by as much as 3-4 times. The memory speed inside the GPU is about 10 000-22 000 Mb/s. Get used to "hardware" optimization).

Gigabyte will give you 400MB/sec instead of 1800MB/sec. This motherboard-video-card figure is critical for the speed of most scientific tasks.

Otherwise, it will be the drivers and software that will cause most problems, not the hardware.

The subject of hardware for OpenCL is endless and I am not going to speak on the subject of hardware as there are plenty of other sites for that. Except that the name of a good professional chipset 790-, 890-, 990 FX from AMD for Phenom speaks for itself.

Ugh. Thank you! Crossing it out.

Went to overlocker forum and reviews.

[[Deleted]]

MetaDriver:

Ugh. Thank you! Scratch that.

I'm off to the reviews and overlocker forums.

That's me still softening the description of the situation. In fact even MSI couldn't stand the continuation of such years-long Gigabyte's deceit and recently flipped out, forgetting to be "politically correct" towards their competitor:

http://nvworld.ru/news/gigabyte-use-not-true-pcie-30/

"MSI held a presentation detailing what third-generation PCIe-enabled motherboards should really be. In their report, the company pointed out that their competitor Gigabyte, which claims to have 40 motherboard models with PCIe Gen3 support, is not telling the truth.

.....

At the conclusion of their review MSI noted that only one board from Gigabyte has true PCIe 3.0 support, the G1.Sniper 2."

http://www.techpowerup.com/151718/MSI-Calls-Bluff-on-Gigabyte-s-PCIe-Gen-3-Ready-Claim.html

[Renat Fatkhullin]

It looks like we will also use OpenCL for the needs of basic generation of tens of millions of intermediate ticks while simulating the development of price bars.

This could give a serious acceleration on the passes, but we will get the real results after internal tests.

[Renat Fatkhullin]

AlexEros:

Thank you for the very interesting and concentrated information!

[Ivan Chasov]

AlexEros:

I softened the description of the situation. In fact, even MSI couldn't stand the continuation of such years of Gigabyte's deception and recently flipped out, forgetting to be "politically correct" towards the competitor:

......

Thank you, educate yourself. And as an option.... offer the link on mother with 4 PCI-E slots, satisfying conditions on speed and cost up to 5000 r.

Naturally, there are enough variants with the price over 10k.

[Ivan Chasov]

MetaDriver:

...

2. about the case - yes, I know it needs a big one, but it looks like it's no longer relevant - I need to look for another mother. AlexEros wasrather convincing about Gigabyte. For now I'll continue to dig deeper into the topic and later will make my choice based on testimonials and descriptions.

:) nevertheless, I think, it will be topical about the case, since another mother will be for 3-4 double-slot video cards. And horizontally the video cards themselves are not small...

[[Deleted]]

WChas:

Thank you, enlightened. And as an option.... suggest a link to a mother with 4 PCI-E slots, satisfying the conditions on speed and cost up to 5000 r.

Naturally, there are plenty of options with a price over 10k.

Don't bother with the number of slots. You have 42 PCI-E lanes for all your slots anyway. Therefore 4 PCI-E slots on a decent NON-SERVER motherboard will always work with х16+ х8+х8+х8 = 40 lanes. 3 slots = x16+x16+x8 and so on.

Here is the Ukrainian hotline.ua, choose ASUS, then the parameters of PCI-E, and in the brief for each decent motherboard see the maximum scheme of its work. 8 hryvnia = 1 dollar.

http://hotline.ua/computer/materinskie-platy/412-438

There are also reviews and reviews of the motherboards under their description.

[Ivan Chasov]

AlexEros:

Don't bother with the number of slots. You have 42 PCI-E lanes for all the slots anyway. Therefore, 4 PCI-E slots on a decent NON-SERVER motherboard will always work with х16+ х8+х8+х8 = 40 lanes. 3 slots = x16+x16+x8 and so on.

........

Thanks, I'll look into it further....

So more than 2 graphics cards on one mother makes no sense? Or is there an option?

[[Deleted]]

WChas:

Thanks, I'll look into it further....

It turns out more than 2 graphics cards on one mother makes no sense to put? Or is there an option?

If your program is (already) so well optimized, that data transfers between CPU and GPU are insignificant, then 4 x8 scheme will suit you fine. But when you write and REPLAY a (non-optimized) program, the PCI-E speed can become an unpleasant hindrance when you don't understand why there is no speed gain from CUDA/OpenCL.

[Konstantin Grebenshikov]

Will this technology work in the cloud for those who sell their resources ??? or not .... And what kind of result to expect from XFX Readon HD 5850 Black Ediction.

How will it be implemented graphically in the metatester (only client for the cloud)?