Initial multiplicity work with core and memory clock very near observed maximum gave me a clear result that 3X beat the 2X condition appreciably. Typical elapsed time for a GRP6/CUDA55 WU was about 1:33:32, yielding a calculated system daily productivity of 206,448 of which 203,222 coming from the GTX 1070. I have enough results from the 4X condition to suggest it is somewhat higher yielding than 3X.
I should offer a few reservations here: as this it to be my daily driver PC, I expect not to reach these calculated output numbers by reason of competition for resources from my normal use of the machine. Also I expect to back off another increment in both core and memory clock rate for long-term use, and may find I need to go down farther than that. I don't consider either spontaneous reboots or WU computation failures acceptable even at low rates if avoidable by clock rate.
While all my early results had default condtions across the system, for some time now I've been using higher fan speeds. I believe the GTX 1070 would have lowered clock rate at my current (and many previous) operating points without this. Much more recently I set up conditions governing task CPU affinity, and priority in an effort to stabilize WU elapsed times, especially at higher multiplicity. It is possible, even likely, that these measures have slightly improved overall output, and that they may help the higher multiplicity test points in comparison to the lower. In previous adventures with Process Lasso, I have found it very hard to predict which specific interventions will help or hurt. As my primary current goal is to investigate the 1070 running conditions, I don't plan to tweak the affinity/priority settings in the near term, as my main goal of significantly reducing task to task variation seems to have been achieved.
Lastly, there are reports that a GTX 1060 card has been sighted in Hong Kong, and that Nvidia may conduct announcement and release sooner than previous guessed--possibly in July. If this card has useful Einstein performance and is suitably priced, it may well be more widely useful to Einstein users than either the 1070 or 1080. We can hope that Nvidia will feel pressure from the new AMD RX 480 family offerings to set a good price for the 1060.
I imagine Gamboleer is far ahead on this shopping game, but if others here are interested in buying RX 480, GTX 1080, or GTX 1070 while they are scarce, they may find the facilities at Nowinstock.com to be useful.
For example, at the moment their RX 480 page shows a few places which may be among those accepting orders on rollout day tomorrow, though perhaps for just a few minutes. While this finding tool (and use of Distill) has helped diligent users find open offerings, it is part of a massive arms race in which people eager to buy swamp offering sites within minutes.
Some RX 480 pages went live on Amazon a few minutes ago, while I was putting food away in the kitchen, so I got to the party a little late. I have:
- A "who knows when it will ship" Sapphire order.
- A "you'll get in in a month" MSI order, and
- An "it will arrive Thursday" XFX order.
I'm only intending to buy one card for now, but wanted to put in some extra orders for the others in case stock levels change. I'm not familiar with XFX brand. Is it decent?
My 4X run on the heavily overclocked GTX 1070 gave results indicating over 1% higher throughput than my 3X run, with average elapsed time for BRP6/CUDA55 work of 2:03:20 and computed daily system credit of 208,375, of which 205,270 coming from the GPU work.
I tried to move on to 5X, but the system continued to run just 4 GPU WUs. Is this normal (I know Gary Roberts has gone higher, but perhaps it is relevant that he runs in Linux while I'm on Windows?
Details:
Windows 10 Pro 64-bit
Boinc version 7.6.22
host CPU is a 4 core Haswell, without HT.
Einstein preference for "GPU utilization factor of BRP apps" was set to 0.2 (having been 0.25 for the successful 4X run)
Computing preference for "use at most" nn% "of the CPUs" had been set to 40 (I was intending to get and got one Einstein GW task), and I bumped it up to 48 without effect.
Yes, my system has downloaded fresh BRP6 work since I made the change.
Any wisdom out there on "beyond 4X".
The truth is, I'm not very enthusiastic about going above 4X, and am only asking because I had stated I would look up until I saw turnover.
My short-term intention is to turn both core clock and memory clock down one increment, return multiplicity to 3X, and run for a day to raise stability confidence and measure productivity. Then I plan a brief trial of SETI.
If I can get an RX 480 at a reasonable price reasonably soon, I think I should swap that into this box, which would allow a pretty direct comparison controlled for host characteristics (and for owner proclivities). Otherwise I plan to put a GTX 750Ti in the box and see what fraction of their individual productivities results when they run together.
My last 3 AMD based cards have all been XFX. I'm very pleased with them.
Thank you. It appears this is the one I'll be testing, as it has shipped. I'll have it in about 30 hours, and will be able to substitute it into an i5-3570 system that's been running 3x BRP6 with a slightly underclocked 7970.
I'll be able to get a few units done overnight and will post here before shutting all the computers down for the USA holiday weekend.
Not a place i deem to recall going, but try 0.19 instead of 0.20 and ensure you have enough cpu resources free (i think you have - i'd up it 100% availability and suspend all cpu tasks as well to see it it that starts the 5x process)
No wisdom but some experiences I appreciate to share with you and other users :-)
Going from e.g. 4x to 6x I'm doing the following: E@H Setup
> allowing ONLY tasks for GPU(s) (e.g. BRP6)
> setting GPU utilization factor of BRP apps to 0.16 Computing Preferences
> work for 10 days + 10 additional days (less would be also ok)
The E@H Server will then start automatically with downloading new tasks (BRP6)
Check if new downloaded tasks are (0.2 CPUs + 0.16 Nvidia GPUs (device x))
If so wait - latest if one BRP task 0.25 has finished and uploaded there should run now 6 task on each GPU (that's what E@H is alsways doing for me)
Do not forget
- to set back the cache (days of work) and
- allowing other applications to run too again
(if you like).
Cheers,
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
Thanks for the various suggestions. I've decided to kill two birds with one stone by revising my already started run at 3X GPU + 1 CPU job at two increments below maximum core and memory clock to run with zero CPU jobs, while setting the fraction of allowed CPUs to 100%, and the "GPU utilization factor of BRP apps" to 0.19, but putting my requested queue at 3 days.
As a just-completed GW CPU jobs has BOINC imagining that my 212 GPU tasks represent over 16 days of work, this won't give me downloads (and thus won't change my operating multiplicity) for well over a day, but is a great chance to check out the zero CPU jobs 3X GPU performance. On setting the condition I noticed an immediate rise in GPU utilization and GPU fraction of TDP, so there is at least a chance that casting aside the single CPU task may raise total system production in my current configuration.
As GW tasks are currently scarce, I should not persist in this condition for long, either working to burn off the tasks, or returning them so someone else can have them.
Initial multiplicity work
)
Initial multiplicity work with core and memory clock very near observed maximum gave me a clear result that 3X beat the 2X condition appreciably. Typical elapsed time for a GRP6/CUDA55 WU was about 1:33:32, yielding a calculated system daily productivity of 206,448 of which 203,222 coming from the GTX 1070. I have enough results from the 4X condition to suggest it is somewhat higher yielding than 3X.
I should offer a few reservations here: as this it to be my daily driver PC, I expect not to reach these calculated output numbers by reason of competition for resources from my normal use of the machine. Also I expect to back off another increment in both core and memory clock rate for long-term use, and may find I need to go down farther than that. I don't consider either spontaneous reboots or WU computation failures acceptable even at low rates if avoidable by clock rate.
While all my early results had default condtions across the system, for some time now I've been using higher fan speeds. I believe the GTX 1070 would have lowered clock rate at my current (and many previous) operating points without this. Much more recently I set up conditions governing task CPU affinity, and priority in an effort to stabilize WU elapsed times, especially at higher multiplicity. It is possible, even likely, that these measures have slightly improved overall output, and that they may help the higher multiplicity test points in comparison to the lower. In previous adventures with Process Lasso, I have found it very hard to predict which specific interventions will help or hurt. As my primary current goal is to investigate the 1070 running conditions, I don't plan to tweak the affinity/priority settings in the near term, as my main goal of significantly reducing task to task variation seems to have been achieved.
Lastly, there are reports that a GTX 1060 card has been sighted in Hong Kong, and that Nvidia may conduct announcement and release sooner than previous guessed--possibly in July. If this card has useful Einstein performance and is suitably priced, it may well be more widely useful to Einstein users than either the 1070 or 1080. We can hope that Nvidia will feel pressure from the new AMD RX 480 family offerings to set a good price for the 1060.
I imagine Gamboleer is far
)
I imagine Gamboleer is far ahead on this shopping game, but if others here are interested in buying RX 480, GTX 1080, or GTX 1070 while they are scarce, they may find the facilities at Nowinstock.com to be useful.
For example, at the moment their RX 480 page shows a few places which may be among those accepting orders on rollout day tomorrow, though perhaps for just a few minutes. While this finding tool (and use of Distill) has helped diligent users find open offerings, it is part of a massive arms race in which people eager to buy swamp offering sites within minutes.
Some RX 480 pages went live
)
Some RX 480 pages went live on Amazon a few minutes ago, while I was putting food away in the kitchen, so I got to the party a little late. I have:
- A "who knows when it will ship" Sapphire order.
- A "you'll get in in a month" MSI order, and
- An "it will arrive Thursday" XFX order.
I'm only intending to buy one card for now, but wanted to put in some extra orders for the others in case stock levels change. I'm not familiar with XFX brand. Is it decent?
My last 3 AMD based cards
)
My last 3 AMD based cards have all been XFX. I'm very pleased with them.
My 4X run on the heavily
)
My 4X run on the heavily overclocked GTX 1070 gave results indicating over 1% higher throughput than my 3X run, with average elapsed time for BRP6/CUDA55 work of 2:03:20 and computed daily system credit of 208,375, of which 205,270 coming from the GPU work.
I tried to move on to 5X, but the system continued to run just 4 GPU WUs. Is this normal (I know Gary Roberts has gone higher, but perhaps it is relevant that he runs in Linux while I'm on Windows?
Details:
Windows 10 Pro 64-bit
Boinc version 7.6.22
host CPU is a 4 core Haswell, without HT.
Einstein preference for "GPU utilization factor of BRP apps" was set to 0.2 (having been 0.25 for the successful 4X run)
Computing preference for "use at most" nn% "of the CPUs" had been set to 40 (I was intending to get and got one Einstein GW task), and I bumped it up to 48 without effect.
Yes, my system has downloaded fresh BRP6 work since I made the change.
Any wisdom out there on "beyond 4X".
The truth is, I'm not very enthusiastic about going above 4X, and am only asking because I had stated I would look up until I saw turnover.
My short-term intention is to turn both core clock and memory clock down one increment, return multiplicity to 3X, and run for a day to raise stability confidence and measure productivity. Then I plan a brief trial of SETI.
If I can get an RX 480 at a reasonable price reasonably soon, I think I should swap that into this box, which would allow a pretty direct comparison controlled for host characteristics (and for owner proclivities). Otherwise I plan to put a GTX 750Ti in the box and see what fraction of their individual productivities results when they run together.
RE: My last 3 AMD based
)
Thank you. It appears this is the one I'll be testing, as it has shipped. I'll have it in about 30 hours, and will be able to substitute it into an i5-3570 system that's been running 3x BRP6 with a slightly underclocked 7970.
I'll be able to get a few units done overnight and will post here before shutting all the computers down for the USA holiday weekend.
FYI: Radeon Software
)
FYI:
Radeon Software Crimson Edition 16.6.2 has been released for Windows®
Included Highlights:
• Support for Radeon™ RX 480
• Support for Oculus Rift™ SDK v1.3 and HTC Vive™
• Hitman™ Crossfire Profiles update
RE: Any wisdom out there
)
Not a place i deem to recall going, but try 0.19 instead of 0.20 and ensure you have enough cpu resources free (i think you have - i'd up it 100% availability and suspend all cpu tasks as well to see it it that starts the 5x process)
Overclockers UK RX-480
In stock and ... XFX ordered.
RE: Any wisdom out there on
)
No wisdom but some experiences I appreciate to share with you and other users :-)
Going from e.g. 4x to 6x I'm doing the following:
E@H Setup
> allowing ONLY tasks for GPU(s) (e.g. BRP6)
> setting GPU utilization factor of BRP apps to 0.16
Computing Preferences
> work for 10 days + 10 additional days (less would be also ok)
The E@H Server will then start automatically with downloading new tasks (BRP6)
Check if new downloaded tasks are (0.2 CPUs + 0.16 Nvidia GPUs (device x))
If so wait - latest if one BRP task 0.25 has finished and uploaded there should run now 6 task on each GPU (that's what E@H is alsways doing for me)
Do not forget
- to set back the cache (days of work) and
- allowing other applications to run too again
(if you like).
Cheers,
Arthur
I know I am a part of a story that starts long before I can remember and continues long beyond when anyone will remember me [Danny Hillis, Long Now]
RE: Any wisdom out there on
)
Thanks for the various suggestions. I've decided to kill two birds with one stone by revising my already started run at 3X GPU + 1 CPU job at two increments below maximum core and memory clock to run with zero CPU jobs, while setting the fraction of allowed CPUs to 100%, and the "GPU utilization factor of BRP apps" to 0.19, but putting my requested queue at 3 days.
As a just-completed GW CPU jobs has BOINC imagining that my 212 GPU tasks represent over 16 days of work, this won't give me downloads (and thus won't change my operating multiplicity) for well over a day, but is a great chance to check out the zero CPU jobs 3X GPU performance. On setting the condition I noticed an immediate rise in GPU utilization and GPU fraction of TDP, so there is at least a chance that casting aside the single CPU task may raise total system production in my current configuration.
As GW tasks are currently scarce, I should not persist in this condition for long, either working to burn off the tasks, or returning them so someone else can have them.