When overclocking a GPU one normally can modify three values: core, shader and memory. When altering these values which one will be the most effective to fasten up crunching ?

Btw: the PC, or better, the GPU is only used for crunching, so there's no need for a 'good' picture... ;)

so there's no need for a 'good' picture... ;)

Yes, but a need for accurate calculations ;)

You should require both, core and shader clock, whereas memory clock should be pretty irrelevant.

MrS

---

Scanning for our furry friends since Jan 2002

[quote]

You should require both, core and shader clock, whereas memory clock should be pretty irrelevant.

MrS

Sorry, but that's a very popular myth on the memory. In the F@H GPU community people say that the memory doesn't matter--its just not true. Set the core and shaders then start upping the memory and watch the PPD go up. Lower the memory and watch the PPD go down..It works that way with every card that I have (4). I have absolutely no idea why it would be any different for GPUGrid. Just try it for Yourself.

I remember that in folding@home GPU 1 a 50% drop in memory clock would cause a slow-down of 10-15%. So if I'd increase my 9800 mem clock from 1100 to 1160 I could expect a 1.6% performance increase.

Things could have changed with GPU 2 though and things could be different in GPU-Grid. It depends on how localized the algorithmn is, that is how good the caches can be used and if execution ever has to wait for memory requests. GPU-Grid calculates larger molecules than f@h, so they *may* be more memory performance bound. I'll give it a try, but improvements of a few % will be difficult to see and need averaging over several WUs.

MrS

---

Scanning for our furry friends since Jan 2002

Wow, this was unexpected! First result is in:

Previously I was at 61.0 - 62.5 ms/step and never below 61 ms with app 6.45. The WU with mem clock increased from 1100 to 1160 MHz needed 59.5 ms/step. 5.5% mem clock increase -> 2.5 - 5.0% performance increase, depending on what the normal time would have been for that WU.

So I take everything back and claim the opposite: mem clock does matter in GPU-Grid.

MrS

---

Scanning for our furry friends since Jan 2002

Wow, this was unexpected! First result is in:

Previously I was at 61.0 - 62.5 ms/step and never below 61 ms with app 6.45. The WU with mem clock increased from 1100 to 1160 MHz needed 59.5 ms/step. 5.5% mem clock increase -> 2.5 - 5.0% performance increase, depending on what the normal time would have been for that WU.

So I take everything back and claim the opposite: mem clock does matter in GPU-Grid.

MrS

Good to know.
Thanks to that info I just decided to push the memory clock of my GTX260 a little bit further.
I´ll post some results.
Maybe the incerease will be not the same across different cards, as some may be not so much memory bandwith starved as overs.


Cu KyleFL

When running at stock values (600/1500/1000) I had about 65 ms/step...
Now I'm running at 790/1800/1100 and had 52 ms/step for my last WU... ;)
But I'm still wondering which one has got the highest impact on decreasing the calculation times...

Maybe the incerease will be not the same across different cards, as some may be not so much memory bandwith starved as overs.

Yes, I'd expect as much. But most cards are pretty balanced anyway. On GT200 the larger caches should also help.

MrS

---

Scanning for our furry friends since Jan 2002

The next 2 WUs are in, 62.7 ms with some interactive use and 61.3 ms with only minor use. Damn those long term averages..

MrS

---

Scanning for our furry friends since Jan 2002

Can you remind me what it the tool to overclock from Linux?

gdf

There are CoolBits from Nvidia and NVclock...

If you are lucky and the card and driver support it, somehow...
I tried with nvclock (even compiled the newest one from SVN) and coolbits set in xorg.conf, but both give me a white screen when ever I change a value. Here it does not matter if I lower the clocks or increase them.
I logged in via SSH and nvclock -i reported the core speed somewhere over 900MHz, way too much...

I ended up with overclocking my 8800GTS under windows, modifying and flashing the video BIOS to the new values.

I've got an ASUS 9600GT stock at 650/1625/1800 and am wondering the 'best' way to OC it. Do I keep all the ratio's the same or just the engine/shader ratio?

I've upped it to 700/1800/1900 without seeing a noticeable difference and now have it set to 725/1815/1944.

Live long and BOINC!

ps. If I had of known that an 8800 was 'faster' than a 9600 then I would have bought an 8800! Gotta love marketing! Bet then I was a real nOOb, now I'm just a noob.

As anybody overclocked a GTX280?

gdf

Not a GTX280, but a GTX260 (if that helps)

I figured something out:

Overclocking Core&shaders ~15% did have almost a linear impact on WU times.
Overclocking the memory 10% did get me a performance gain ~1-2%.

Stock Clockings:
Core 576 - Shader 1242 - Memory 999 : Time per step: 41,7
Overclocking Core & Shaders:
Core 650 - Shader 1401 - Memory 999 : Time per step: 37,5
Overclocking Memory (and a little bit more on Core & shaders again)
Core 661 - Shader 1425 - Memory 1101 : Time per step: 37,3

It seems to me, that the last gain of 0.2ms is only because of the minimal raised Core & Shaderspeed and not because of the 10% higher memoryclock


Cu KyleFL

I've got an ASUS 9600GT stock at 650/1625/1800 and am wondering the 'best' way to OC it. Do I keep all the ratio's the same or just the engine/shader ratio?

I've upped it to 700/1800/1900 without seeing a noticeable difference and now have it set to 725/1815/1944.

I must admit I'm underwhelmed by the responses to what I thought was a fairly simple question, anyhoo.....

The first wu completed at the new speeds shows a dramatic decrease in time. Fantastic stuff! I'll see how the next one goes.

Live long and BOINC!

I must admit I'm underwhelmed by the responses to what I thought was a fairly simple question, anyhoo.....

No, it's not simple. Intuition tells me that the ratios don't matter, as long as you don't go extreme. But I can't be 100% sure since I didn't test it specifically.

My suggestion: find the maximum stable clock for engine and shader first, then for the memory and afterwards back off a bit for safety. And don't care about the ratios at all.

MrS

---

Scanning for our furry friends since Jan 2002

My suggestion: find the maximum stable clock for engine and shader first, then for the memory and afterwards back off a bit for safety. And don't care about the ratios at all.

MrS

Agreed, that's pretty much how I approach it....

I have it (GTX280) overclocked to:

clock: 697, shaders: 1500, mem: stock

In my last three units I got these times per step:

29.688 ms
29.096 ms
29.391 ms

This OC is not the maximum possible. Simply I tried to put shaders to 1500 (linked clock) and I found it works fine. I have not tried to go beyond it. I have never made an error WU with these settings as you can see in my account.

My suggestion: find the maximum stable clock for engine and shader first, then for the memory and afterwards back off a bit for safety. And don't care about the ratios at all.
MrS

Agreed, that's pretty much how I approach it....

Excellent! Many thanks for the responses. Looks like the 11% engine/shader oc and 8% memory oc gave a 15% decrease in overall gpu time for the first wu. The wu time is now below 24 hrs!