I run some machines with a GTX 1660 and a GTX 1650 combination, under Windows 7.
I find the disparity between task runtimes on the two card types is much greater at Einstein than other projects.
I've recently lost the fans on one of the 1660 cards (gone back for RMA), so I'm just running the 1650 in that machine.
And I've swapped it from the PCIe x8 motherboard slot to the x16 slot. Much faster. Einstein's bus transfer handling is - different? Need to factor that in too.
And I've swapped it from the PCIe x8 motherboard slot to the x16 slot. Much faster. Einstein's bus transfer handling is - different? Need to factor that in too.
First of all - do it need bus or not! As you can see if memory not enough bus transfers increase a lot - and then bus bandwidth do matter of course. But should bus be so in use first hand. Maybe not, maybe the problem is in low GPU RAM, not in bus per se.
But could you clarify how many tasks per GPU you run before and now?
I made assumption it was 3 per card from stderr output but then you said "bump to 2" that would mean before it was 1 task per GPU, not 3 ?...
Raistmer,
I apologize for the delay. Upto to today I have been running 1 or 2 GR gpu tasks on my Nvidia cards.
I dropped the Gtx 1660 Super offline so I could generate a "pure" Gtx 1060 3GB result a couple of days ago.
Later today I will take a stab at displaying Gpu-z results for one, two and three GR GPU tasks on that Gtx 1060 3GB card. I have been giving the one task an extended baseline. I will give 2 and 3 GPU tasks extended run times too.
I will be minding the store closely for the 3 GPU tasks because I expect to run out of video ram and have the apps crashing.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Worth attempt to run 4 at once just to get GPU-Z picture, hardly it will provide any speedup. [With 3 GB of memory 3 FGRPs perhaps fully in GPU mem but 4 will definitely not. So, will you see crash or will you see BUS load increase (memory swapping) ?]
What about performance 3 vs 2 ? Looks like power consumption little increased...
EDIT: if you will attempt to run 4 FGRPs at once better to suspend CPU work at least firstly.
Your GPU-Z log shows great system memory usage (most probably it comes from many CPU cores doing BOINC CPU tasks). If GPU memory swap will occur system will need even more system memory so you could see crashes not because GPU driver can't handle 4 tasks but because of system memory overflow.
Better to distinguish between these reasons for better understanding of results.
Raistmer* wrote:Not so
)
I was under the impression that running that many GR tasks in a 3gb would cause memory based app crashes.
Let me try bumping it up to 2 and see what happens.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
I run some machines with a
)
I run some machines with a GTX 1660 and a GTX 1650 combination, under Windows 7.
I find the disparity between task runtimes on the two card types is much greater at Einstein than other projects.
I've recently lost the fans on one of the 1660 cards (gone back for RMA), so I'm just running the 1650 in that machine.
And I've swapped it from the PCIe x8 motherboard slot to the x16 slot. Much faster. Einstein's bus transfer handling is - different? Need to factor that in too.
Richard Haselgrove wrote: I
)
Sounds like for the purpose of getting a consistent stream of data I need to run one gtx 1060 3gb or one gtx 1660 Super but not both.
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:Richard
)
I am now running only my Gtx 1060 3Gb card.
I will see if I can find the image site I used to use for S@H so I can show the Gpu-Z results.
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:I will see if I
)
If not, Imgur should work for you.
Richard Haselgrove
)
First of all - do it need bus or not! As you can see if memory not enough bus transfers increase a lot - and then bus bandwidth do matter of course. But should bus be so in use first hand. Maybe not, maybe the problem is in low GPU RAM, not in bus per se.
Tom M wrote: I am now
)
My version has embedded image upload service. So no probs at all, it provides BB code itself.
But could you clarify how many tasks per GPU you run before and now?
I made assumption it was 3 per card from stderr output but then you said "bump to 2" that would mean before it was 1 task per GPU, not 3 ?...
Raistmer* wrote:But could you
)
Raistmer,
I apologize for the delay. Upto to today I have been running 1 or 2 GR gpu tasks on my Nvidia cards.
I dropped the Gtx 1660 Super offline so I could generate a "pure" Gtx 1060 3GB result a couple of days ago.
Later today I will take a stab at displaying Gpu-z results for one, two and three GR GPU tasks on that Gtx 1060 3GB card. I have been giving the one task an extended baseline. I will give 2 and 3 GPU tasks extended run times too.
I will be minding the store closely for the 3 GPU tasks because I expect to run out of video ram and have the apps crashing.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Sorry for the delay. Things
)
Sorry for the delay. Things got very busy in my life. And when I wasn't busy I was sleeping :)
Three tasks are now running. And not crashing (so far).
1 task
Two tasks
Three tasks are now running. And not crashing (so far).
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Worth attempt to run 4 at
)
Worth attempt to run 4 at once just to get GPU-Z picture, hardly it will provide any speedup. [With 3 GB of memory 3 FGRPs perhaps fully in GPU mem but 4 will definitely not. So, will you see crash or will you see BUS load increase (memory swapping) ?]
What about performance 3 vs 2 ? Looks like power consumption little increased...
EDIT: if you will attempt to run 4 FGRPs at once better to suspend CPU work at least firstly.
Your GPU-Z log shows great system memory usage (most probably it comes from many CPU cores doing BOINC CPU tasks). If GPU memory swap will occur system will need even more system memory so you could see crashes not because GPU driver can't handle 4 tasks but because of system memory overflow.
Better to distinguish between these reasons for better understanding of results.