Author

Topic: I'm not understanding "Worksize" (Read 8356 times)

legendary
Activity: 1512
Merit: 1000
April 06, 2012, 02:21:44 PM
#8


That's interesting. I'm still a little unsure on the full use of powertune. Like on my 5970's, I run the voltage 1137 and clock at 910 without changing the powertune. Would I even have anything to gain by changing it? I've never bothered to try changing it because I'm already struggling to keep card temps down.

I know this is old, but I came across it searching for something unrelated.  Powertune @ 20% gave me an increase of 40MH/s on my unlocked 6950.
sr. member
Activity: 392
Merit: 250
February 05, 2012, 07:50:50 AM
#7
It makes perfect sense.  Decreasing memory clock on a card wouldn't make make anything go faster at all.  At higher core clocks, it does though.  Why?

AMD added a limiter to the cards to prevent damage (known as Powertune on the 6000 series).  With the 4000 series cards, people were actually damaging PCBs, wiring, and GPUs by using synthetic benchmarks like OCCV and Furmark.  ATI/AMD never anticipated that users would do this (because no game would ever stress the ASIC to this degree).  100 percent utilization on every processor in the chip was actually more current (and heat) than the engineers designed for.

When the 5000 series came out, AMD put a limiter on the cards to prevent component damage/overheating.  They set a limit on each card's current draw.  If the card draws too much power, it skips clock cycles to throttle back the power consumption.  When you clock the memory down, you are dropping power consumption.  Dropping the memory clock down frees up additional power that the GPU can use.  If you decrease your memory clock and your hash rate goes up, it's because it was actually throttling your GPU in the first place.

You can also raise the Powertune limit on some cards.  Whether you can or not (and the range you can work with) varies by card (it's set in the card's BIOS).  Some cards (like the MSI Lightning series) have DIP switches to disable the limiter entirely for extreme overclocking.

That's interesting. I'm still a little unsure on the full use of powertune. Like on my 5970's, I run the voltage 1137 and clock at 910 without changing the powertune. Would I even have anything to gain by changing it? I've never bothered to try changing it because I'm already struggling to keep card temps down.
member
Activity: 63
Merit: 10
February 05, 2012, 06:58:32 AM
#6
Don't feel bad i ran a GTX 580 before i had any ATi cards, it was a MSI lightning too, real power hog for 160MHash/s, I never ran into the issue, in fact i think the powertune feature looks more like this:

GPU Clock: 950 - 880 - 920 - 880 - 920  etc, rather than usage, the usage issue you're having sounds more like CPU affinity, try selecting a different CPU core, or more than one, you might even want to change the flags to something like:

-w 512 -v
or
-aggression=4 -gpugrid=64 -gputhreads=384
or
-gpugrid=480 -gputhreads=480

Depending on which miner you're using one or the other will work. See which one gives you the best results.

The 6950's i have are the one's that have powertune, or whatever where i have to increase the power setting slider up to 20% in CCC to stop if from fluctuating, my 6870 does not do this at all.

-Also shata, thanks for Hi-jacking my thread!  Wink


 
sr. member
Activity: 547
Merit: 253
February 05, 2012, 06:44:18 AM
#5

When the 5000 series came out, AMD put a limiter on the cards to prevent component damage/overheating.  They set a limit on each card's current draw.  If the card draws too much power, it skips clock cycles to throttle back the power consumption.

Do you know if this is also implemented on 560Ti cards? I know they arent exactly perfect on mining but thats curretly what im running I gotta 5830 coming in the mail to start my dedicated miner.

Reason im asking is im having huge problem getting my 560ti's to run 99-100% Closest I get them to is roughly 92-96% with -F1 tags even running cuda miner. Ive mined alot before with 5770s and 6970 and never had this problem.
hero member
Activity: 642
Merit: 500
February 05, 2012, 02:14:32 AM
#4
It makes perfect sense.  Decreasing memory clock on a card wouldn't make make anything go faster at all.  At higher core clocks, it does though.  Why?

AMD added a limiter to the cards to prevent damage (known as Powertune on the 6000 series).  With the 4000 series cards, people were actually damaging PCBs, wiring, and GPUs by using synthetic benchmarks like OCCV and Furmark.  ATI/AMD never anticipated that users would do this (because no game would ever stress the ASIC to this degree).  100 percent utilization on every processor in the chip was actually more current (and heat) than the engineers designed for.

When the 5000 series came out, AMD put a limiter on the cards to prevent component damage/overheating.  They set a limit on each card's current draw.  If the card draws too much power, it skips clock cycles to throttle back the power consumption.  When you clock the memory down, you are dropping power consumption.  Dropping the memory clock down frees up additional power that the GPU can use.  If you decrease your memory clock and your hash rate goes up, it's because it was actually throttling your GPU in the first place.

You can also raise the Powertune limit on some cards.  Whether you can or not (and the range you can work with) varies by card (it's set in the card's BIOS).  Some cards (like the MSI Lightning series) have DIP switches to disable the limiter entirely for extreme overclocking.
member
Activity: 63
Merit: 10
February 04, 2012, 08:58:24 PM
#3
No, shares will always be equal no matter your settings. Just find the fastest setting and go with it, as it can vary with cards/drivers/SDK version.

For example, my 5800-series (ubuntu, 11.5 + sdk 2.4) cards with memory underclocked to 300 run fastest with the worksize 256. With stock clock on memory, 128 is faster, although not as fast as 256 with underclocked memory Smiley

Thanks for that reply  Cheesy

I haven't downclocked memory yet, haven't dedicated the rigs to mining just yet but thats strange how downclocking memory makes it faster?
member
Activity: 121
Merit: 10
February 04, 2012, 08:37:29 PM
#2
No, shares will always be equal no matter your settings. Just find the fastest setting and go with it, as it can vary with cards/drivers/SDK version.

For example, my 5800-series (ubuntu, 11.5 + sdk 2.4) cards with memory underclocked to 300 run fastest with the worksize 256. With stock clock on memory, 128 is faster, although not as fast as 256 with underclocked memory Smiley
member
Activity: 63
Merit: 10
February 04, 2012, 06:42:06 PM
#1
I have some 6870's running on GUI+phoenix and iv'e been playing with the flags, i came out with  

-k phatk VECTORS BFI_INT AGGRESSION=14 worksize=128 FASTLOOP=false

On Stream 2.3, 1000mhz core, iv'e been changing things around just to see if it made any real difference and 300 - 307 M/Hashs is pretty much where it planted, now everything else seems to have had an effect one way or another (very small increments) except "Worksize" iv'e changed it to 64, 128, 256 and 512(lol) and didn't see it make any difference in the M/Hashs after starting it back up.

First off, if somebody would be so kind, explain to me what the worksize is exactly, and when mining does it make any difference, for example, if it does 10 shares@Worksize 64 vs 5 shares @ 128 vs 2.5 @ 256 etc, or is that not right?
Jump to: