Is the card running in P0 or P2 power state when running the Einstein tasks?
I just started up nvidia inspector to examine this point and believe it indicated I was running in P2 state.
There may be an opportunity there. It has been about 18 months since I worked through this matter on my 970, and I've forgotten a lot, and the nvidia inspector user interface looks a bit different. I'll put off this particular joy for a little while yet.
Please do update on on what settings exactly you are running and when in terms of task concurrency. Currently, I am running 2 GTX 970's BRP6 v55 and i'm wringing out damn near maximum throughput from them. 125,000 credits/day/card...last I checked it was the highest placed dual 970 system on einstein.
I'm curious to know the processing times at 3x so I can make a comparison of averages and see if Pascal is really worth the upgrade. Thank you for sharing your results.
Please do update on on what settings exactly you are running and when in terms of task concurrency. Currently, I am running 2 GTX 970's BRP6 v55 and i'm wringing out damn near maximum throughput from them. 125,000 credits/day/card...last I checked it was the highest placed dual 970 system on einstein.
I made some hasty moves probing the environment at first--not suitable for careful measurement.
However for several hours now I have been been running 2X multiplicity. I think these are averaging elapsed times slightly under 1 hour 20 minutes.
The only "overclocking" of my own employed so far has been to adjust the fan curve to a more aggressive one, which at my conditions has moved the reported core clock speed up perhaps 60 MHz.
All I'll say for now is that this card will certainly get me well over 150,000 credit/day, and that I doubt even careful overclocking and tweaking will get me past 200,000.
I actually already have a spreadsheet with numbers for 1X, 2X, 3X, and 4X. But I don't believe the comparisons are accurate, with too small a sample size to address some of the comparison issues. Broadly speaking, with the current BRP6/CUDA55 application I see much smaller benefit in going from 1X to 2X than I recall in the longer ago past, and to first order all of 2X, 3X, and 4X seem quite close to one another.
I definitely plan to share better based numbers in the coming days. Probably I shall spend one day each at 2X, 3X, and 4X without any overclocking, then revert to 2X and attempt memory overclocking (which I suspect is probably the biggest tweaking opportunity at hand).
In between I am losing some time to minor box adjustments (I moved up the speed of the bottom and side case fans this morning, which may have bought me about 1 degree C lower GPU temperature) and to running some testing on SETI. So the daily totals and RAC progress of the box will not reflect the capability of the card for a little while yet.
Here is a report on a first operating point for my MSI GTX 1070 Founders Edition card running Einstein BRP6/CUDA55.
No core clock or memory clock intervention was in place.
The fan curve was appreciably elevated by use of MSI Afterburner (what would have been 53% fan speed averaged to 68%, bringing the reported GPU temperature down more than 10C.
[pre]2 Number of GRP6 GPU tasks at once
1 Number of 1.04 G Wave F tasks running at once
1:18:40 Average elapsed time for GPU tasks
6:47:47 Average elapsed time for CPU tasks
161,085 Daily credit rate, GPU tasks
3,531 Daily credit rate, CPU tasks
164,616 System daily credit rate
183.1 Watts System power draw at the wall (not including monitor)
1862 average core clock rate
1901.2 memory clock rate (did not vary)
64.1 degrees C average GPU temperature
68 average fan speed percentage
90 average GPU load percentage
84 average memory controller load percentage
70.8 card power consumption average percentage of TDP
118 approximate incremental watts above system idle attributable to Einstein work
1394 credit/day per incremental watt
[/pre]
As the card self-adjusts core clock rate in response to temperature, I should mention that the room averaged about 79F during this test, that the case has 8 fans, though most run rather slowly, and that my fan curve adjustment almost certainly slightly increased throughput.
Although this system is intended to become my daily driver, during this testing span it had negligible interactive use.
I plan to transition to 3X GPU unit running now, and expect to report numbers for that test point tomorrow. I probably won't attempt to tinker with memory clock rate until late this week. I think it likely that will be substantially helpful (much more than any slight increase I might get from running 3X or 4X or tampering with the support application priority and affinity).
I have not taken a careful idle power measurement in this configuration. When I do the last two numbers which refer to "incremental watts" will change slightly.
Regarding the credit/day per incremental watt, which I regard as a very important overall performance metric, this beats my 750Ti SC card handily, as the same calculation and measurement methods give about 1130 for it. As I regard the 750/750Ti as previous generation power efficiency kings, this is actually a very good result.
Thanks for the numbers!
I'd be interested in results from running just 1 WU, so I can compare (though not ideally) this with other GPUs, especially AMD which some of them can't run >1 WU.
For comparison, my current numbers for running 1 BRP6 WU:
GPU Power Time Factor
(W) (low-high) 10^6/(Pwr*Time)
----------------------------------------------------
Fury X 150 2150-2400 3.10-2.78
Radeon HD 7950 110 3100-3400 2.93-2.67
Tesla K20c 105 3200-3400 2.98-2.80
GTX 750 Ti 30 6400-6800 5.21-4.90
GTX 1070 FE 100? 2800?? 3.57??
Time reported was depending on CPU load (no/low CPU load - full CPU load).
I have also added your GTX 1070 based on some numbers you reported earlier, but not sure if that's correct.
Nice results, looks like it's a bit more powerful and slightly more efficient than a 7970. Your one run at 4x had you bumping into the 170k daily RAC range, so I think you can squeeze a bit more out of it, and that's without overclocking. Does that put this in the range of a 980?
Your prediction that it would be a better buy than the 1080 (190k) looks to be true. The question is now, this or a 980ti (225k overclocked) after the $125 price drop? I don't know how much a 980ti would draw at the wall.
I still haven't seen any listings to pre-order my RX 480.
Please do update on on what settings exactly you are running and when in terms of task concurrency. Currently, I am running 2 GTX 970's BRP6 v55 and i'm wringing out damn near maximum throughput from them. 125,000 credits/day/card...last I checked it was the highest placed dual 970 system on einstein.
I made some hasty moves probing the environment at first--not suitable for careful measurement.
However for several hours now I have been been running 2X multiplicity. I think these are averaging elapsed times slightly under 1 hour 20 minutes.
The only "overclocking" of my own employed so far has been to adjust the fan curve to a more aggressive one, which at my conditions has moved the reported core clock speed up perhaps 60 MHz.
All I'll say for now is that this card will certainly get me well over 150,000 credit/day, and that I doubt even careful overclocking and tweaking will get me past 200,000.
I actually already have a spreadsheet with numbers for 1X, 2X, 3X, and 4X. But I don't believe the comparisons are accurate, with too small a sample size to address some of the comparison issues. Broadly speaking, with the current BRP6/CUDA55 application I see much smaller benefit in going from 1X to 2X than I recall in the longer ago past, and to first order all of 2X, 3X, and 4X seem quite close to one another.
I definitely plan to share better based numbers in the coming days. Probably I shall spend one day each at 2X, 3X, and 4X without any overclocking, then revert to 2X and attempt memory overclocking (which I suspect is probably the biggest tweaking opportunity at hand).
In between I am losing some time to minor box adjustments (I moved up the speed of the bottom and side case fans this morning, which may have bought me about 1 degree C lower GPU temperature) and to running some testing on SETI. So the daily totals and RAC progress of the box will not reflect the capability of the card for a little while yet.
Thank you for this response. When I made that comment yesterday I cited figures of 125,000/card. I see now my cards are producing 262,xxx meaning figures of 131,xxx per card per day. It will be interesting to see exactly where the 1070 will run at 3x. It will be more interesting to understand what factor of multiplicity will be most favorable for the 1070, and if it will pass to being 3x or 4x per card.
1080 figures will be interesting as well, although I doubt that under current application we will be able to take proper advantage of the memory architecture present in that card.
Here's my Titan Black for reference: https://einsteinathome.org/host/12034687/tasks
No overclock on CPU (1070mhz), mem set to 3500mhz.
Power consumption near 100%, so close to 250watts.
GPU temps running at 48C with an EVGA 980Ti hybrid cooler.
Running 3x BRP6 Beta. Current average of around 6300s, 2111s for a single unit.
There is a 960 in there also running 3X @ 1400mhz gpu & 3500mhz mem.
This 1070 looks promising, and may look into a hybrid model if EVGA makes one or try to adapt the 980Ti kit to fit it.
Here's my Titan Black for reference: https://einsteinathome.org/host/12034687/tasks
No overclock on CPU (1070mhz), mem set to 3500mhz.
Power consumption near 100%, so close to 250watts.
GPU temps running at 48C with an EVGA 980Ti hybrid cooler.
Running 3x BRP6 Beta. Current average of around 6300s, 2111s for a single unit.
There is a 960 in there also running 3X @ 1400mhz gpu & 3500mhz mem.
This 1070 looks promising, and may look into a hybrid model if EVGA makes one or try to adapt the 980Ti kit to fit it.
2110s for 1 WU, that's comparable to Fury X, however that one consumes ~150-160 W.
Which card is "Faster" on
)
Which card is "Faster" on Einstein.
980 TI or, the GT 1080 ?
or, are there still alot of driver and tweaking
improvements to be made on the newer card ?
Bill
RE: RE: Is the card
)
Please do update on on what settings exactly you are running and when in terms of task concurrency. Currently, I am running 2 GTX 970's BRP6 v55 and i'm wringing out damn near maximum throughput from them. 125,000 credits/day/card...last I checked it was the highest placed dual 970 system on einstein.
I'm curious to know the processing times at 3x so I can make a comparison of averages and see if Pascal is really worth the upgrade. Thank you for sharing your results.
RE: Please do update on on
)
I made some hasty moves probing the environment at first--not suitable for careful measurement.
However for several hours now I have been been running 2X multiplicity. I think these are averaging elapsed times slightly under 1 hour 20 minutes.
The only "overclocking" of my own employed so far has been to adjust the fan curve to a more aggressive one, which at my conditions has moved the reported core clock speed up perhaps 60 MHz.
All I'll say for now is that this card will certainly get me well over 150,000 credit/day, and that I doubt even careful overclocking and tweaking will get me past 200,000.
I actually already have a spreadsheet with numbers for 1X, 2X, 3X, and 4X. But I don't believe the comparisons are accurate, with too small a sample size to address some of the comparison issues. Broadly speaking, with the current BRP6/CUDA55 application I see much smaller benefit in going from 1X to 2X than I recall in the longer ago past, and to first order all of 2X, 3X, and 4X seem quite close to one another.
I definitely plan to share better based numbers in the coming days. Probably I shall spend one day each at 2X, 3X, and 4X without any overclocking, then revert to 2X and attempt memory overclocking (which I suspect is probably the biggest tweaking opportunity at hand).
In between I am losing some time to minor box adjustments (I moved up the speed of the bottom and side case fans this morning, which may have bought me about 1 degree C lower GPU temperature) and to running some testing on SETI. So the daily totals and RAC progress of the box will not reflect the capability of the card for a little while yet.
Here is a report on a first
)
Here is a report on a first operating point for my MSI GTX 1070 Founders Edition card running Einstein BRP6/CUDA55.
No core clock or memory clock intervention was in place.
The fan curve was appreciably elevated by use of MSI Afterburner (what would have been 53% fan speed averaged to 68%, bringing the reported GPU temperature down more than 10C.
[pre]2 Number of GRP6 GPU tasks at once
1 Number of 1.04 G Wave F tasks running at once
1:18:40 Average elapsed time for GPU tasks
6:47:47 Average elapsed time for CPU tasks
161,085 Daily credit rate, GPU tasks
3,531 Daily credit rate, CPU tasks
164,616 System daily credit rate
183.1 Watts System power draw at the wall (not including monitor)
1862 average core clock rate
1901.2 memory clock rate (did not vary)
64.1 degrees C average GPU temperature
68 average fan speed percentage
90 average GPU load percentage
84 average memory controller load percentage
70.8 card power consumption average percentage of TDP
118 approximate incremental watts above system idle attributable to Einstein work
1394 credit/day per incremental watt
[/pre]
As the card self-adjusts core clock rate in response to temperature, I should mention that the room averaged about 79F during this test, that the case has 8 fans, though most run rather slowly, and that my fan curve adjustment almost certainly slightly increased throughput.
Although this system is intended to become my daily driver, during this testing span it had negligible interactive use.
I plan to transition to 3X GPU unit running now, and expect to report numbers for that test point tomorrow. I probably won't attempt to tinker with memory clock rate until late this week. I think it likely that will be substantially helpful (much more than any slight increase I might get from running 3X or 4X or tampering with the support application priority and affinity).
I have not taken a careful idle power measurement in this configuration. When I do the last two numbers which refer to "incremental watts" will change slightly.
Regarding the credit/day per incremental watt, which I regard as a very important overall performance metric, this beats my 750Ti SC card handily, as the same calculation and measurement methods give about 1130 for it. As I regard the 750/750Ti as previous generation power efficiency kings, this is actually a very good result.
Thanks for the numbers! I'd
)
Thanks for the numbers!
I'd be interested in results from running just 1 WU, so I can compare (though not ideally) this with other GPUs, especially AMD which some of them can't run >1 WU.
For comparison, my current numbers for running 1 BRP6 WU:
GPU Power Time Factor (W) (low-high) 10^6/(Pwr*Time) ---------------------------------------------------- Fury X 150 2150-2400 3.10-2.78 Radeon HD 7950 110 3100-3400 2.93-2.67Tesla K20c 105 3200-3400 2.98-2.80
GTX 750 Ti 30 6400-6800 5.21-4.90
GTX 1070 FE 100? 2800?? 3.57??
Time reported was depending on CPU load (no/low CPU load - full CPU load).
I have also added your GTX 1070 based on some numbers you reported earlier, but not sure if that's correct.
-----
Nice results, looks like it's
)
Nice results, looks like it's a bit more powerful and slightly more efficient than a 7970. Your one run at 4x had you bumping into the 170k daily RAC range, so I think you can squeeze a bit more out of it, and that's without overclocking. Does that put this in the range of a 980?
Your prediction that it would be a better buy than the 1080 (190k) looks to be true. The question is now, this or a 980ti (225k overclocked) after the $125 price drop? I don't know how much a 980ti would draw at the wall.
I still haven't seen any listings to pre-order my RX 480.
RE: RE: Please do update
)
Thank you for this response. When I made that comment yesterday I cited figures of 125,000/card. I see now my cards are producing 262,xxx meaning figures of 131,xxx per card per day. It will be interesting to see exactly where the 1070 will run at 3x. It will be more interesting to understand what factor of multiplicity will be most favorable for the 1070, and if it will pass to being 3x or 4x per card.
1080 figures will be interesting as well, although I doubt that under current application we will be able to take proper advantage of the memory architecture present in that card.
I will be following your posts closely archae!
Something I had overlooked
)
Something I had overlooked until now:
http://wccftech.com/amd-rx-480-pre-order-in-stock/
RX 480, 470 AND 460 on June 29. I will still be buying and testing the 480.
I'm liking the potential of
)
I'm liking the potential of that 480.
Here's my Titan Black for reference:
https://einsteinathome.org/host/12034687/tasks
No overclock on CPU (1070mhz), mem set to 3500mhz.
Power consumption near 100%, so close to 250watts.
GPU temps running at 48C with an EVGA 980Ti hybrid cooler.
Running 3x BRP6 Beta. Current average of around 6300s, 2111s for a single unit.
There is a 960 in there also running 3X @ 1400mhz gpu & 3500mhz mem.
This 1070 looks promising, and may look into a hybrid model if EVGA makes one or try to adapt the 980Ti kit to fit it.
RE: I'm liking the
)
2110s for 1 WU, that's comparable to Fury X, however that one consumes ~150-160 W.
-----