Thanks for posting. Here's th link more convenient:
https://simtk.org/home/memtest/

They say for testing memory and logic.. do you know how "good" it is at finding logic / gpu core errors?

MrS
____________
Scanning for our furry friends since Jan 2002

That makes a lot of sense: if the memory controller in the GPU (the "logic") is broken, it will surely be detected. But it's not a proper test for the shaders, which the title "Cuda Memory tester" already implies.

Drop us a line if you (or someone else) see the finished shader tester. I think this would be even more important than the memory tester.

MrS
____________
Scanning for our furry friends since Jan 2002

I dont like the idea of having to register to download the program.
No doubt they want to share your details with selected partners.

For now "bugmenot" still works ;)

MrS
____________
Scanning for our furry friends since Jan 2002

The Milkyway's ATI project uses different software, so its difficult to compare to GPUGRID Clients and work units.

For the GPUGRID an overclocked card will get through the tasks faster, so you will get more points over a set period of time. I expect that if you underclock the reverse will be true; in a set period of time you will get through less work and therefore get less points. My opinion is that if you can overclock, keep it stable and not increase the Voltage it is perhaps worth doing. Otherwise, not!

I expect there might be the odd quirky setting found, occasionally, to do with DDR3 timings that might rock the boat a bit, but not by much. Perhaps this is what the ATI App guy found? Perhaps it was a bit more specific; underclocking the GPU but not the RAM or he was just talking about reducing power usage.

For GPUGRID, if you can underclock the card and reduce the voltage it might be an option, but unless you reduce the voltage the power usage won't decrease, so it would be a bit pointless.

Power saving is reduced by lowering the Voltage of the GPU &/or Shaders &/or Memory. Just because Power Saving 2D Clocks reduce all three, does not me you have to.

What we are all looking for is better performance and lower cost. So as Bob said, you could look into reducing the RAM speed (and lowering the voltage), but also raise the core & Shader speeds.

If you are 5% slower because of reduced memory speed and Voltage, but 5% faster from a speed bump in GPU and Shaders at no extra Voltage increase, you will be doing the same work but saving a little bit on the Electric Bill. You might also find that a slight Voltage increase for the GPU & Shaders yields much faster results. So the Electric bill would be the same but the result count would rise.

Better to test these before running GPUGRID, just in case your speeds fail and you lose your running tasks. However you really need to be measuring the power usage at the plug on the wall to know how much you are saving or losing to have an idea if it is worthwhile – the clock speeds and voltages don’t really tell you much by themselves.

I also think you are probably slightly better off with a card that has less memory. A 512MB card will perform just as well as a 2GB card for GPUGRID, but will cost less at the outset and will probably use slightly less electric, on RAM. That said, not all DDR3 video RAM is equal, some require slightly higher or lower Voltages, and well made RAM tends to be slightly more energy efficient; better materials, less energy loss, less energy requirements.

There was a thread on overclocking where we agreed on the following (I tested myself):

- GPU-Grid does need memory clock
- downclocking reduces performance significantly, much more so than you save in power
- OC'ing GPU-RAM does speed things up, not as much as shader & core OC, but clearly measureable

With Milkyway it's different. They're doing completely different calculations: the working set is much smaller and thus much less memory bandwidth is needed. And the GDDR5 used on high end ATI is much more power hungry than the GDDR3 used on current nVidia cards. A 4870 512 MB uses ~60W at idle, almost all of which is simply due to the RAM. Downclocking the RAM reduces power usage linearly and thus does help. For the current MW app a memory speed of ~500 MHz (x4) instead of the stock 900 MHz is sufficient, saving you ~30W.

Power saving can be achieved by both, reducing clock and voltage. Reducing clock speed decreases the "dynamic power draw", i.e. the power which is used upon transistor state switches linearly. It doesn't affect sub-threshold leakage, though (the power which each transistor uses even if it shouldn't do anything).
Reducing voltage (V) classically reduces power draw with the square of V, i.e. half the voltage means one quarter the power draw. In modern chips it's more like V³ due to leakage currents. Voltage can't be reduced infinitely, though, as it's needed to achieve higher frequencies.

MrS
____________
Scanning for our furry friends since Jan 2002

Power saving is reduced by lowering the Voltage of the GPU &/or Shaders &/or Memory. Just because Power Saving 2D Clocks reduce all three, does not me you have to.

I meant, Power usage is reduced by lowering the Voltage of the GPU, Shaders and Memory.

So if you decide you want to contribute to the MW GPU project, there is no point trying to underclock your RAM

I meant, Power usage is reduced by lowering the Voltage of the GPU, Shaders and Memory.

Did you read my post?

So if you decide you want to contribute to the MW GPU project, there is no point trying to underclock your RAM

Sorry, but what do you mean? (and why)

Now I understand, thanks! Well.. I expect it to be this way (lower power consumption for GDDR3 compared to GDDR5), but neither did I test it myself nor do I remember reading a proper test of this. It's just that the problem really only revealed itself as a major one to me on the new ATIs. I wouldn't mind being proven wrong here, though.. which would mean that nVidia guys could also save some power via downclocking, should they choose to only run MW. If the latter would make any sense is a totally different question ;)

MrS
____________
Scanning for our furry friends since Jan 2002

I would like to save a few Watts when running GPUGRID, keep the card cooler, the system quieter and still get through the same amount of work. Perhaps some cards can be tweaked by nibbling away at the voltage here, a DIY green card, or even underclocked to use less Watts Per Point? Unfortunately I don’t have the time to look into this but I would like to give it a go.

I recently ran folding@home for a while, which used the GPU more intensively than GPUGRID, much more; It had my GTS 250 running at 88 degrees C, that’s 12 degrees hotter than GPUGRID. It certainly tested the RAM, but I don’t think a burn-in is supposed to run 24/7 for days and weeks! If I decide to dedicate any more time to that project I will seriously consider throttling the card back, or more likely using a more expendable card.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I would like to save a few Watts when running GPUGRID, keep the card cooler, the system quieter and still get through the same amount of work.

I think I have managed to get my GTX260 stable at about 112% but only time will tell for sure.
I thought it might be working at about 118% as it seemed ok completing Milkyway tasks, which run hotter, but the GPUGRID tasks tended to fail. It is Not good when they fail 5 hours into a 19h task!