going from +1500 to +3000 is also more complicated than just power limits, you can't just raise the power and expect that to work. I would argue that probably no card will be stable with that configuration anyway. it might be a little faster if you could even do it. but you'll get more performance just using the extra power limit to increase core clocks. all of this just starts making the card very inefficient though.
I think the 30-series cards have some form of memory error detection/recovery so that it doesnt just flat out crash. even with a high memory clock, actual performance can degrade because of all the memory errors. you would need to watch if the tasks are actually running faster and validating.
I have experimented with overclocking memory transfer rates to +1,700 and more but I then notice more "invalids" showing up. Even though it does clock at a slightly faster rate, it is offset with more invalids. My 3080Ti is "stable" at a MTR of +1,600.
My just wondering question......what is the longest you ever got a Nvidia card to run here?
For some reason my 660Ti is still running here since Sept. 2012 mainly in the same box but the last 2 months I have plugged it into a couple of my even older boxes.
I guess it is beyond the warranty now
LOL yes it's past ALL warranties now!!
AND that same card ran for years before that in someone elses systems as well.
no that one was a 560 and it died about 5 years ago.......the 660Ti SC is one I bought the month it came out for just over $300.....BUT the box I just put the 660Ti in IS the same one the 560 used to run in......the slot in that OLD box is at the bottom of the board so I had to lay mylar down under the board so it wasn't just laying on the metal box.......I posted a picture here somewhere when that 560 died but can't find it and I even looked around longer than I should have ......in other news my power company is once again turning the power off all over the area to do some goofy repair and they say from 9am to 5pm so my Einsteins will just be sitting there waiting to run when that starts and I am usually not awake that early so I may have to suspend all of them around 3 or 4am so that also means no heat and all the rest
My just wondering question......what is the longest you ever got a Nvidia card to run here?
For some reason my 660Ti is still running here since Sept. 2012 mainly in the same box but the last 2 months I have plugged it into a couple of my even older boxes.
I guess it is beyond the warranty now
LOL yes it's past ALL warranties now!!
AND that same card ran for years before that in someone elses systems as well.
no that one was a 560 and it died about 5 years ago.......the 660Ti SC is one I bought the month it came out for just over $300.....BUT the box I just put the 660Ti in IS the same one the 560 used to run in......the slot in that OLD box is at the bottom of the board so I had to lay mylar down under the board so it wasn't just laying on the metal box.......I posted a picture here somewhere when that 560 died but can't find it and I even looked around longer than I should have ......in other news my power company is once again turning the power off all over the area to do some goofy repair and they say from 9am to 5pm so my Einsteins will just be sitting there waiting to run when that starts and I am usually not awake that early so I may have to suspend all of them around 3 or 4am so that also means no heat and all the rest
Oh sorry!! I hope you stay warm enough to keep the hummers fed and you and the family warm too!!!
I have experimented with overclocking memory transfer rates to +1,700 and more but I then notice more "invalids" showing up. Even though it does clock at a slightly faster rate, it is offset with more invalids. My 3080Ti is "stable" at a MTR of +1,600.
I went and looked at my Invalids rates. Currently they seem to be running 20-30 per day. Then I looked at the Invalid rates as far back as the Website had. It was showing 1-3 a day.
So I have just dropped my OC on the 2 gpu system to 700 from 1500. The reason I picked that OC is Keith posted a script that used that OC.
Both systems with rtx 3080 ti's are currently set to their highest Power Limits.
I am leaving the single rtx 3080 ti at OC of 1500 with its maximum Power Limit.
This could easily be why actual performance has not been tracking calculated estimates very well.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
The overclock you can use depends on your card generation. I explained that. If you tried to use my overclock example for a Pascal card for your Ampere card, no wonder you are getting massive error correction retransmits and tons of invalids.
The overclock you can use depends on your card generation. I explained that. If you tried to use my overclock example for a Pascal card for your Ampere card, no wonder you are getting massive error correction retransmits and tons of invalids.
Later on I dropped OC completely but am maintaining the maximum PL's.
An examination of "marked invalid's" is down drastically. Looks like 4 instead of 30.
====edit============
Apparently the rtx 3080 ti FE's are perfectly happy to run at their maximum PL (400 watts) with no OCing of any type. 0 marked invalids since midnight. Time for processing is running in the 161 sec range.
I think I will just stand pat for now. Its RAC appears to be on the way up again.
The smaller EVGA rtx 3080 ti has had 4 marked invalids since midnight. But the GPU is running very hot (but not throttled, 85-88C?) due to the way the cpu radiator is pointed (it is sucking in air from the outside while attached to inside of the top of the case. The fans are not rattling at all).
This rtx 3080 ti appears to be processing slower without OC but may still have the possibility of reaching 3.3M RAC without any OCing. I still want to OC the graphics memory a little. But more likely start with +500 and watch the marked invalids rate. If it jumps much at all, lower the OC.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
George mentioned a graphics memory OC that was near the +1500 I was usually using. And generating very high marked invalid rates.
But my GPU brands/models are different so I expect to start with +750 and bump up/down depending on the trends in marked invalid's.
Start on that tomorrow after more than 24 hours of presumed lower m.i.
GWGeorge007 wrote:
I have experimented with overclocking memory transfer rates to +1,700 and more but I then notice more "invalids" showing up. Even though it does clock at a slightly faster rate, it is offset with more invalids. My 3080Ti is "stable" at a MTR of +1,600.
Tom... Tom... TOM!!!
Graphics Clock is different than Memory Transfer Rates. And what is Graphics Memory?
You lowered your memory transfer rates or you lowered your graphics clock rates. One or the other...
I didn't suggest you lower your MTR (memory transfer rates) by a SIGNIFICANT amount, just a little bit.
I know I have trouble at times remembering things, but I will ALWAYS read and re-read what I am about to post before I actually hit the "POST COMMENT".
Apparently the rtx 3080 ti FE's are perfectly happy to run at their maximum PL (400 watts) with no OCing of any type. 0 marked invalids since midnight. Time for processing is running in the 161 sec range.
I think I will just stand pat for now. Its RAC appears to be on the way up again.
The EDIT is no longer available on the previous message.
Just looked again and the Graphics Memory clock is actually set to 700 as of the moment. With any luck it will continue to have no "marked invalids".
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
The main reason for using a boosted clock for the memory transfer rate is to overcome the penalty in the memory clock that Nvidia drivers impose on consumer cards when the drivers detect a compute workload.
But the penalty has differed for every Nvidia card generation. For Maxwell and Pascal the penalty was about 1000Mhz. So because the card will be running in P2 power state under BOINC crunching and 1000Mhz lower memory clocks than the published card spec for running games or a video application.
The penalty for Turing cards was only 400Mhz. So a memory boost of 800Mhz puts the effective memory clock at 14400Mhz or only 400Mhz above spec.
So I normally boosted my 1080 cards by 1400Mhz for the memory clocks. That put the card back at the spec P0 rate plus a little 400Mhz overclock on top. You should be able to use 400-800Mhz boost on the memory over the P0 rate and not produce memory errors if you keep the card relatively cool.
Ian&Steve C. wrote: going
)
I have experimented with overclocking memory transfer rates to +1,700 and more but I then notice more "invalids" showing up. Even though it does clock at a slightly faster rate, it is offset with more invalids. My 3080Ti is "stable" at a MTR of +1,600.
Just a thought...
Proud member of the Old Farts Association
mikey wrote: MAGIC Quantum
)
MAGIC Quantum Mechanic
)
Oh sorry!! I hope you stay warm enough to keep the hummers fed and you and the family warm too!!!
GWGeorge007 wrote: I have
)
I went and looked at my Invalids rates. Currently they seem to be running 20-30 per day. Then I looked at the Invalid rates as far back as the Website had. It was showing 1-3 a day.
So I have just dropped my OC on the 2 gpu system to 700 from 1500. The reason I picked that OC is Keith posted a script that used that OC.
Both systems with rtx 3080 ti's are currently set to their highest Power Limits.
I am leaving the single rtx 3080 ti at OC of 1500 with its maximum Power Limit.
This could easily be why actual performance has not been tracking calculated estimates very well.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
The overclock you can use
)
The overclock you can use depends on your card generation. I explained that. If you tried to use my overclock example for a Pascal card for your Ampere card, no wonder you are getting massive error correction retransmits and tons of invalids.
Keith Myers wrote:The
)
Later on I dropped OC completely but am maintaining the maximum PL's.
An examination of "marked invalid's" is down drastically. Looks like 4 instead of 30.
====edit============
Apparently the rtx 3080 ti FE's are perfectly happy to run at their maximum PL (400 watts) with no OCing of any type. 0 marked invalids since midnight. Time for processing is running in the 161 sec range.
I think I will just stand pat for now. Its RAC appears to be on the way up again.
The smaller EVGA rtx 3080 ti has had 4 marked invalids since midnight. But the GPU is running very hot (but not throttled, 85-88C?) due to the way the cpu radiator is pointed (it is sucking in air from the outside while attached to inside of the top of the case. The fans are not rattling at all).
This rtx 3080 ti appears to be processing slower without OC but may still have the possibility of reaching 3.3M RAC without any OCing. I still want to OC the graphics memory a little. But more likely start with +500 and watch the marked invalids rate. If it jumps much at all, lower the OC.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: George
)
Tom... Tom... TOM!!!
Graphics Clock is different than Memory Transfer Rates. And what is Graphics Memory?
You lowered your memory transfer rates or you lowered your graphics clock rates. One or the other...
I didn't suggest you lower your MTR (memory transfer rates) by a SIGNIFICANT amount, just a little bit.
I know I have trouble at times remembering things, but I will ALWAYS read and re-read what I am about to post before I actually hit the "POST COMMENT".
Proud member of the Old Farts Association
Tom M wrote: Apparently the
)
The EDIT is no longer available on the previous message.
Just looked again and the Graphics Memory clock is actually set to 700 as of the moment. With any luck it will continue to have no "marked invalids".
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
GWGeorge007 wrote: Text
)
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
The main reason for using a
)
The main reason for using a boosted clock for the memory transfer rate is to overcome the penalty in the memory clock that Nvidia drivers impose on consumer cards when the drivers detect a compute workload.
But the penalty has differed for every Nvidia card generation. For Maxwell and Pascal the penalty was about 1000Mhz. So because the card will be running in P2 power state under BOINC crunching and 1000Mhz lower memory clocks than the published card spec for running games or a video application.
The penalty for Turing cards was only 400Mhz. So a memory boost of 800Mhz puts the effective memory clock at 14400Mhz or only 400Mhz above spec.
So I normally boosted my 1080 cards by 1400Mhz for the memory clocks. That put the card back at the spec P0 rate plus a little 400Mhz overclock on top. You should be able to use 400-800Mhz boost on the memory over the P0 rate and not produce memory errors if you keep the card relatively cool.