Pages:
Author

Topic: Lowering mem clock to idle speeds SPEEDS UP Mh/s - page 5. (Read 39034 times)

member
Activity: 79
Merit: 10
XFX 5870 here

I dropped the memory clock to 300.

Mh/s is holding at 336 now where as before it was fluctuating from 330. My temp appears to have dropped by 5c. I'll do more testing but it's looking good so far.
On both my 5870 and 5850, lowering the memory clock decrease hash rate. Does anyone with these 2 cards see the opposite?

Yes, I get a noticeable increase with both of those cards.

What miners/OS/Catalyst/opencl are you guys running?
newbie
Activity: 35
Merit: 0
XFX 5870 here

I dropped the memory clock to 300.

Mh/s is holding at 336 now where as before it was fluctuating from 330. My temp appears to have dropped by 5c. I'll do more testing but it's looking good so far.
full member
Activity: 154
Merit: 100
On both my 5870 and 5850, lowering the memory clock decrease hash rate. Does anyone with these 2 cards see the opposite?

Yes, I get a noticeable increase with both of those cards.
legendary
Activity: 3080
Merit: 1080
Hmm, so the anandtech article on the 5870 mentions:
Quote
The specific CRC function used in GDDR5 can detect 1-bit and 2-bit errors with 100% accuracy, with that accuracy falling with additional erroneous bits. This is due to the fact that the CRC function used can generate collisions, which means that the CRC of an erroneous data burst could match the proper CRC in an unlikely situation. But as the odds decrease for additional errors, the vast majority of errors should be limited to 1-bit and 2-bit errors.

Should an error be found, the GDDR5 controller will request a retransmission of the faulty data burst, and it will keep doing this until the data burst finally goes through correctly. A retransmission request is also used to re-train the GDDR5 link (once again taking advantage of fast link re-training) to correct any potential link problems brought about by changing environmental conditions. Note that this does not involve changing the clock speed of the GDDR5 (i.e. it does not step down in speed); rather it’s merely reinitializing the link. If the errors are due the bus being outright unable to perfectly handle the requested clock speed, errors will continue to happen and be caught. Keep this in mind as it will be important when we get to overclocking.

http://www.anandtech.com/show/2841/12
Maybe with higher memory clocks some computation is delayed because the memory needs to retransmit again. This may be the reason some people see increases in performance and others don't. (some people's cards can transmit 100% at the stock speed and other people's can't)

On both my 5870 and 5850, lowering the memory clock decrease hash rate. Does anyone with these 2 cards see the opposite?

Well that's odd...why does it not affect hash rates on the 5970?

My 5970 is currently clocked at 850 Mhz and 300 Mhz memory and currently pulling 318.2 Mhash/s per GPU...can anyone else corroborate this?
member
Activity: 79
Merit: 10
Hmm, so the anandtech article on the 5870 mentions:
Quote
The specific CRC function used in GDDR5 can detect 1-bit and 2-bit errors with 100% accuracy, with that accuracy falling with additional erroneous bits. This is due to the fact that the CRC function used can generate collisions, which means that the CRC of an erroneous data burst could match the proper CRC in an unlikely situation. But as the odds decrease for additional errors, the vast majority of errors should be limited to 1-bit and 2-bit errors.

Should an error be found, the GDDR5 controller will request a retransmission of the faulty data burst, and it will keep doing this until the data burst finally goes through correctly. A retransmission request is also used to re-train the GDDR5 link (once again taking advantage of fast link re-training) to correct any potential link problems brought about by changing environmental conditions. Note that this does not involve changing the clock speed of the GDDR5 (i.e. it does not step down in speed); rather it’s merely reinitializing the link. If the errors are due the bus being outright unable to perfectly handle the requested clock speed, errors will continue to happen and be caught. Keep this in mind as it will be important when we get to overclocking.

http://www.anandtech.com/show/2841/12
Maybe with higher memory clocks some computation is delayed because the memory needs to retransmit again. This may be the reason some people see increases in performance and others don't. (some people's cards can transmit 100% at the stock speed and other people's can't)

On both my 5870 and 5850, lowering the memory clock decrease hash rate. Does anyone with these 2 cards see the opposite?
full member
Activity: 154
Merit: 100
Just retested this on a 5870 running in my windows machine. The card is overclocked to 930MHz core with drivers 11.2 + SDK 2.2

1200MHz = 352 Mhash/sec
450MHz = 328 Mhash/sec
300MHz = 305 Mhash/sec

The cards run faster in Linux, but the effect of lower memory clocks is similar.

EDIT: I would test this on Linux but changing the clockspeeds outside of the overdrive limits requires BIOS flashing. It's much easier just to test in Windows where I can set it with MSI Afterburner.

Strange. I wish we knew more about this instead of just hit and miss.
full member
Activity: 219
Merit: 120
My best guess as to why you can see an increase in hashrate is that the card might use tighter memory timings when you lower the clockspeed. Mining doesn't use very much (if any) RAM so there is still more than enough bandwidth even when underclocked. Tighter timings = lower latency = faster.

Oddly, for some cards this works and for others it doesn't. I have a 5830 and a 5870 in one system, both with the RAM underclocked to 450. The 5830 shows no difference in hashrate, but the 5870 takes a 20-30 Mhash/sec hit. These cards are in a dedicated mining system using the same drivers. (Linux + 11.2 + SDK 2.1)

Did you happen to try 300 on the 5870?

Just retested this on a 5870 running in my windows machine. The card is overclocked to 930MHz core with drivers 11.2 + SDK 2.2

1200MHz = 352 Mhash/sec
450MHz = 328 Mhash/sec
300MHz = 305 Mhash/sec

The cards run faster in Linux, but the effect of lower memory clocks is similar.

EDIT: I would test this on Linux but changing the clockspeeds outside of the overdrive limits requires BIOS flashing. It's much easier just to test in Windows where I can set it with MSI Afterburner.
full member
Activity: 154
Merit: 100
My best guess as to why you can see an increase in hashrate is that the card might use tighter memory timings when you lower the clockspeed. Mining doesn't use very much (if any) RAM so there is still more than enough bandwidth even when underclocked. Tighter timings = lower latency = faster.

Oddly, for some cards this works and for others it doesn't. I have a 5830 and a 5870 in one system, both with the RAM underclocked to 450. The 5830 shows no difference in hashrate, but the 5870 takes a 20-30 Mhash/sec hit. These cards are in a dedicated mining system using the same drivers. (Linux + 11.2 + SDK 2.1)

Did you happen to try 300 on the 5870?
full member
Activity: 219
Merit: 120
does anyone know why this works from a technical point of view?

i could understand if it didnt have any effect but having a positive effect i dont get.


I too would like to know. Any GPU guru's around here?



My best guess as to why you can see an increase in hashrate is that the card might use tighter memory timings when you lower the clockspeed. Mining doesn't use very much (if any) RAM so there is still more than enough bandwidth even when underclocked. Tighter timings = lower latency = faster.

Oddly, for some cards this works and for others it doesn't. I have a 5830 and a 5870 in one system, both with the RAM underclocked to 450. The 5830 shows no difference in hashrate, but the 5870 takes a 20-30 Mhash/sec hit. These cards are in a dedicated mining system using the same drivers. (Linux + 11.2 + SDK 2.1)
legendary
Activity: 3080
Merit: 1080
I cant get MSI to lower my mem clocks lower than 1k on 5970..

Check this link out:

https://bitcointalksearch.org/topic/lowering-mem-clock-to-idle-speeds-speeds-up-mhs-4292

Ye shall see the light in the very first post Smiley
hero member
Activity: 696
Merit: 500
I cant get MSI to lower my mem clocks lower than 1k on 5970..
legendary
Activity: 3080
Merit: 1080
does anyone know why this works from a technical point of view?

i could understand if it didnt have any effect but having a positive effect i dont get.


I too would like to know. Any GPU guru's around here?

newbie
Activity: 55
Merit: 0
does anyone know why this works from a technical point of view?

i could understand if it didnt have any effect but having a positive effect i dont get.
full member
Activity: 126
Merit: 100
My gaming OC is 940/1100 or 940/1150 (stock for my XFX BE edition). during mining it is at 1020/300

Damn! 1020 for the GPU core?! Gotta love those better binned cards. What temps are you seeing at that clock rate, and is the clock 100% stable?

No voltage change. These are 6870s! 1020 core is pretty much 100% stable for mining if fan is at least at 45%, but not for gaming. 1040 on a single card was 95% stable for mining, as long as I did not too much GPU related things on my PC at the same time, and fan was at 50%, 1045 same but like 90% stable. After that 1048 was perhaps 85% and 1050 was 50%.

up to 1048 is 100% stable if I put -f 120, 1040 was 98% stable if -f 60 was put
full member
Activity: 155
Merit: 100
This is not new information. I used it with 4850 , and 5830 and it worked, but with my 5870 I get a 50mhash drop.
legendary
Activity: 3080
Merit: 1080
My gaming OC is 940/1100 or 940/1150 (stock for my XFX BE edition). during mining it is at 1020/300

Damn! 1020 for the GPU core?! Gotta love those better binned cards. What temps are you seeing at that clock rate, and is the clock 100% stable?
newbie
Activity: 14
Merit: 0
My gaming OC is 940/1100 or 940/1150 (stock for my XFX BE edition). during mining it is at 1020/300
What voltages do you use--in both situations--for the GPUs and memory?
full member
Activity: 126
Merit: 100
My gaming OC is 940/1100 or 940/1150 (stock for my XFX BE edition). during mining it is at 1020/300
legendary
Activity: 3080
Merit: 1080
bump so that every can benefit from this! at worst your temps and power consumption go WAY down!

Yeah if you're using Windows - reason being if I'm not mistaken in Linux performance also suffers if you lower the mem clock, no?

But yeah, lowering the mem clock works for me in Windows and it does lower power consumption and temps. Not sure about "WAY" down though. I does cut quite a bit of "amps" off the VRMs powering the memory especially for me with my 4 GB of VRAM. I'd say it probably lowers power usage by 30 watts easily.



If you are feeling really frisky, the VTT voltages can be lowered via BIOS on a 5970.

Hehehe..not that frisky. Well, I may do that in a dedicated miner, but I won't mess around with that on the card which I intend to use for gaming as well. The reason being that when I want to game I do intend to up the mem clock back to stock speeds Smiley The lower voltage may not work with 1200 Mhz mem clock because this is an OVERLOCKED setting..stock mem is like what 900 Mhz for the 5970, no?
legendary
Activity: 3080
Merit: 1080
bump so that every can benefit from this! at worst your temps and power consumption go WAY down!

Yeah if you're using Windows - reason being if I'm not mistaken in Linux performance also suffers if you lower the mem clock, no?

But yeah, lowering the mem clock works for me in Windows and it does lower power consumption and temps. Not sure about "WAY" down though. I does cut quite a bit of "amps" off the VRMs powering the memory especially for me with my 4 GB of VRAM. I'd say it probably lowers power usage by 30 watts easily.



Well MSIa doesn't work for linux... but it would be weird if mem clock reduction makes performance suffer in linux.

30W is WAY down in my book Cheesy what card are you using?

Sapphire Radeon 5970 OC edition..4 GB VRAM...runs stock at 850 Mhz with meme 1200 Mhz

It is weird, but that's what some people have been saying on this very thread. Check back a few posts. In windows lowering mem does not affect performance at 300 Mhz...
Pages:
Jump to: