OpenCL: internal implementation tests in MQL5 - page 36

[Sceptic Philozoff]

MetaDriver:

by the way, how much did you pay?

// I bought Radeon 6930 for 5650p.

And don't ask, MD. Look on yandex.market, you'll see.

The 6930 is unattainable in terms of performance/price ratio. Well, among the decent ones, of course.

But GTX 570 consumes less, 200-210 watts maximum.

I wish I had your problems, hardcore gamers :)

[Vladimir Gomonov]

Mathemat:

And don't ask, MD. Look on Yandex.market, you'll see.

OK, that's it, I'll shut up.

[Andrey Dik]

MetaDriver:

... Which you could have read about beforehand on overclockers.ru :)

By the way, how much did you pay?

// I got a Radeon 6930 for 5650p.

You just have no idea what I didn't read before buying. Alexey does. ;)

I got it for 10951p.

http://www.nix.ru/hardware-review/video-graphics-gpu-benchmark-performance.html?good_id=112198#good_id_112198

Does it work for you on CountPass 131072 and above?

[Sceptic Philozoff]

joo: http://www.nix.ru/hardware-review/video-graphics-gpu-benchmark-performance.html?good_id=112198#good_id_112198

Nix, of course, is a reputable company, but the tests there are frankly unrepresentative. There are just few of them and a lot of useless synthetics there.

Better still look at ixbt, 3dnews, overclockers, thd etc.

But I've done the main subversive work and have not paid attention to trifles like video card :)

[Vladimir Gomonov]

joo:

1. You just can't imagine what I haven't read before buying. Alexey knows. ;)

2. Does it work for you on CountPass 131072 and higher?

1. He seems to know too much... Such people usually don't live long...

;)

2. It's still capable of doubling up to and including 1024*512.

2012.03.16 17:10:16 ParallelTester_00-01 x Соunt inticators = 16; Count history bars = 144000; Count pass = 524288
2012.03.16 17:10:16 ParallelTester_00-01 x GPU time = 6490 ms
..................
2012.03.16 17:09:32 ParallelTester_00-01 x Соunt inticators = 16; Count history bars = 144000; Count pass = 262144
2012.03.16 17:09:32 ParallelTester_00-01 x GPU time = 3323 ms
..................
2012.03.16 17:08:35 ParallelTester_00-01 x Соunt inticators = 16; Count history bars = 144000; Count pass = 131072
2012.03.16 17:08:35 ParallelTester_00-01 x GPU time = 1701 ms

At 1024*1024 the driver already crashes/restarts.

[Andrey Dik]

MetaDriver:

1. He seems to know too much... Such people usually don't live long...

;)

2. Can still withstand doubling up to and including 1024*512.

At 1024*1024 the driver already crashes/restarts.

1) Good health to the "witness". :)

2) More cores, apparently, that's why it holds more.

And in general, if we load each fly with more calculations (now each thread in the test is not too heavy), it would be interesting how our cards will behave in this case.

[Vladimir Gomonov]

joo:

1) Good health to the "witness". :)

2) There are more cores, apparently, so it holds more.

And in general, if I load more calculations on each fly (now each thread in the test is not too heavy), it would be interesting how our cards will behave in this case.

1. OK, .... through your prayers... it lives. :)

2. it's just that radeon is better! ;-Ь

I just sent you a more advanced version on skype. Take it.

[Sceptic Philozoff]

MetaDriver: 2. Выдерживает ещё удвоения до 1024*512 включительно. 

At 1024*1024 the driver already crashes/restarts.

Yes, very curious. It turns out that this is the kind of computations the graphics card can only hold when they are short-lived. It's too intense compared to gaming and it warms those bees up too much. It's a cause for reflection.

[Vladimir Gomonov]

Mathemat:
Yes, very curious. It turns out that these kinds of calculations only hold up when they're short-lived. It's too intense compared to gameplay, and it warms these bees up too much. That's something to think about.

No, I'm sure it's just not enough memory. Overheating is as far away as the moon.

i'll see what i can check it out.

--

Ow! Folks! Anybody got any ideas on how to find the weak link?

[Andrey Dik]

MetaDriver:

No, I'm sure it's just not enough memory. The overheating is as far away as the moon.

I'll see what I can find out what's holding it up.

--

Ow! Folks! Anybody got any ideas on how to find the weak link?

Yeah, I don't think it's the overheating either -- the driver just crashes instantly, that's all.

And the problem is most likely not in the memory but in the OpenCL driver task paralleling mechanism itself. You need to experiment with multiples of the number of flies.