Does a memory overclock improve production? I have a 1070 but haven't really overclocked the memory. I think like 50mhz only.
Yes it does. In fact on the Pascal cards I've worked with so far, the memory overclock I've reached has contributed far more production improvement than has the core clock.
Be careful, though, to check results. On one of my cards even a substantial memory overclock completed and reported normally, and passed the sanity test at the beginning of quorum validation. The first hint of trouble was an inconclusive comparison out of multiple validation attempts, followed--typically days later, by invalid conclusions reached by comparison with the tie-breaker when it came back. I thought I was moving pretty slowly and carefully, but wound up wasting the computation of about 15 BRP6 WUs this way.
Another of my cards had a nasty habit of doing un-requested reboots if the memory overclock was too high. I'm putting together some observations, thoughts and conclusions on Pascal overclocking, and mean to start a thread on that topic "real soon now".
Is your 203k RAC at one task per GPU or several? I've been bumping up my GPU clock gradually and checking for all valid results and I've gone from 123k RAC to 132k RAC. Thats an average of 3124 seconds down to 2898 seconds averaged over like 10 units then calculated RAC based on the average time. Memory changed from +50 to +500.
Win7 driver 368.81. My core clock has been stable at 2050 so just a single strap higher.
GPU util is only had 61-63% so I was running several BRP6-Beta-cuda55 tasks per GPU but dropped it to 1 for consistent times.
I'm currently running all three of my Pascal cards at 3X on BRP6/CUDA55 work only. All three are also overclocked. The 1070 is getting 20% productivity increase from overclocking, the 1060 6GB 14%, and the 1060 3GB 13%. The big overclocking advantage of my 1070 over the other two is in amount of memory overclock. I don't know whether that is just luck of the silicon lottery, or something about 1070s vs. 1060s, or even something about the specific models.
Hmm ok. I had been running at 4x on my 1070 before this testing. I think it was only at maybe 80-85% GPU util. Not sure an extra 20% GPU util gets me 70k RAC.
1080: +1000MHz Mem Clock in Precision X totaling 5514MHz Mem Clcok, Driver 372.70
1070: +774MHz Mem Clock in Precsion X totaling 4576MHz Mem Clock, Driver 372.70
I have both GPUs running BRP6-Beta-CUDA55 tasks at 3x. With this setup, both cards are completing tasks in nearly identical times: around 1 hour 35 minutes. Both cards are running in P2, but the 1070 boosts to around 1974MHz while I can never get the 1080 to boost above its base speed of 1708MHz unless I have a webpage with video open (YouTube, Facebook, etc.).
I am running both a 1070 and a 1080. <snip> I have both GPUs running BRP6-Beta-CUDA55 tasks at 3x. With this setup, both cards are completing tasks in nearly identical times: around 1 hour 35 minutes.
Your 1070 report looks quite similar to my 1070 experience. In light of that, your 1080 report is quite disappointing given the much higher cost, and (I assume) power consumption of the card. I gave up on my plan to buy a 1080 when the earliest Einstein report seemed not to warrant the price, and I've not seen much new information since. Thanks for your report.
I am running both a 1070 and a 1080. <snip> I have both GPUs running BRP6-Beta-CUDA55 tasks at 3x. With this setup, both cards are completing tasks in nearly identical times: around 1 hour 35 minutes.
Your 1070 report looks quite similar to my 1070 experience. In light of that, your 1080 report is quite disappointing given the much higher cost, and (I assume) power consumption of the card. I gave up on my plan to buy a 1080 when the earliest Einstein report seemed not to warrant the price, and I've not seen much new information since. Thanks for your report.
Yes, I've been wondering if it's due to some combination of the setups, as the OS, CPU, storage drive, and RAM type and speed are different between the two machines. A lot of variables. As far as power, the 1080 is running around 49% and the 1070 at around 78%. Of course the 1080 has a higher power limit, but not quite that high. This is also due to the fact that the 1070 is boosting its core clock whilst the 1080 is not.
I am running both a 1070 and a 1080. <snip> I have both GPUs running BRP6-Beta-CUDA55 tasks at 3x. With this setup, both cards are completing tasks in nearly identical times: around 1 hour 35 minutes.
Your 1070 report looks quite similar to my 1070 experience. In light of that, your 1080 report is quite disappointing given the much higher cost, and (I assume) power consumption of the card. I gave up on my plan to buy a 1080 when the earliest Einstein report seemed not to warrant the price, and I've not seen much new information since. Thanks for your report.
Yes, I've been wondering if it's due to some combination of the setups, as the OS, CPU, storage drive, and RAM type and speed are different between the two machines. A lot of variables. As far as power, the 1080 is running around 49% and the 1070 at around 78%. Of course the 1080 has a higher power limit, but not quite that high. This is also due to the fact that the 1070 is boosting its core clock whilst the 1080 is not.
Yes, I've been wondering if it's due to some combination of the setups, as the OS, CPU, storage drive, and RAM type and speed are different between the two machines. A lot of variables. As far as power, the 1080 is running around 49% and the 1070 at around 78%.
I think the lack of core clock boost on the 1080 may somehow be a symptom of underutilzation, but not the major cause of underperformance. It seems somehow the host is not keeping the 1080 very busy. Could you observe and report the % GPU Load as reported by GPU-Z on the two systems? On my main system, running 3X on BRP6/CUDA55, my 1070 reports an average of 95%, and my 1060 of 94%.
While I suspect your host system of somehow failing to support the 1080 adequately to keep it busy, I doubt the cause would be any of the hardware or software differences you list. Are you interested enough in that answer to try swapping the cards?
I think the lack of core clock boost on the 1080 may somehow be a symptom of underutilzation, but not the major cause of underperformance. It seems somehow the host is not keeping the 1080 very busy. Could you observe and report the % GPU Load as reported by GPU-Z on the two systems?
I was also wondering about under utilization as well. It's not always a matter of GPU load, but how hard that work is pushing the card. Anyway, the 1080 is running around 90% GPU Usage and around 75% MCU Usage at 3x on these tasks. The 1070 is at around 94% GPU Load and 85% MCU Usage.
I may try swapping the cards around in the near future to see if that does anything.
One more interesting note, I recently updated the driver on the 1080 system to the latest, and decreased my task times at 3x by 10 minutes each. Before this, the 1070 was actually completing work FASTER than the 1080.
archae86 wrote:mmonnin
)
Is your 203k RAC at one task per GPU or several? I've been bumping up my GPU clock gradually and checking for all valid results and I've gone from 123k RAC to 132k RAC. Thats an average of 3124 seconds down to 2898 seconds averaged over like 10 units then calculated RAC based on the average time. Memory changed from +50 to +500.
Win7 driver 368.81. My core clock has been stable at 2050 so just a single strap higher.
GPU util is only had 61-63% so I was running several BRP6-Beta-cuda55 tasks per GPU but dropped it to 1 for consistent times.
mmonnin wrote:Is your 203k
)
I'm currently running all three of my Pascal cards at 3X on BRP6/CUDA55 work only. All three are also overclocked. The 1070 is getting 20% productivity increase from overclocking, the 1060 6GB 14%, and the 1060 3GB 13%. The big overclocking advantage of my 1070 over the other two is in amount of memory overclock. I don't know whether that is just luck of the silicon lottery, or something about 1070s vs. 1060s, or even something about the specific models.
Hmm ok. I had been running at
)
Hmm ok. I had been running at 4x on my 1070 before this testing. I think it was only at maybe 80-85% GPU util. Not sure an extra 20% GPU util gets me 70k RAC.
Some observations from my
)
Some observations from my setup:
I am running both a 1070 and a 1080.
The 1070 specs are as follows:
OS: Windows 10 x64
CPU: Intel Xeon E3-1240 v5
RAM: 16GB DDR4 @ 2133MHz
Running BOINC from SSD
The 1080 system is as follows:
OS: Windows 7 x64
CPU: Intel i7-4790k
RAM: 16GB DDR3 @ 1866MHz
Running BOINC from HDD
The two GPUs are configured as follows:
1080: +1000MHz Mem Clock in Precision X totaling 5514MHz Mem Clcok, Driver 372.70
1070: +774MHz Mem Clock in Precsion X totaling 4576MHz Mem Clock, Driver 372.70
I have both GPUs running BRP6-Beta-CUDA55 tasks at 3x. With this setup, both cards are completing tasks in nearly identical times: around 1 hour 35 minutes. Both cards are running in P2, but the 1070 boosts to around 1974MHz while I can never get the 1080 to boost above its base speed of 1708MHz unless I have a webpage with video open (YouTube, Facebook, etc.).
Matt_145 wrote:I am running
)
Your 1070 report looks quite similar to my 1070 experience. In light of that, your 1080 report is quite disappointing given the much higher cost, and (I assume) power consumption of the card. I gave up on my plan to buy a 1080 when the earliest Einstein report seemed not to warrant the price, and I've not seen much new information since. Thanks for your report.
archae86 wrote:Matt_145
)
Yes, I've been wondering if it's due to some combination of the setups, as the OS, CPU, storage drive, and RAM type and speed are different between the two machines. A lot of variables. As far as power, the 1080 is running around 49% and the 1070 at around 78%. Of course the 1080 has a higher power limit, but not quite that high. This is also due to the fact that the 1070 is boosting its core clock whilst the 1080 is not.
Matt_145 wrote:archae86
)
Matt what model of 1080 and 1070 do you owe?
I have EVGA Superclocked for
)
I have EVGA Superclocked for both.
Matt_145 wrote:Yes, I've been
)
I think the lack of core clock boost on the 1080 may somehow be a symptom of underutilzation, but not the major cause of underperformance. It seems somehow the host is not keeping the 1080 very busy. Could you observe and report the % GPU Load as reported by GPU-Z on the two systems? On my main system, running 3X on BRP6/CUDA55, my 1070 reports an average of 95%, and my 1060 of 94%.
While I suspect your host system of somehow failing to support the 1080 adequately to keep it busy, I doubt the cause would be any of the hardware or software differences you list. Are you interested enough in that answer to try swapping the cards?
archae86 wrote:I think the
)
I was also wondering about under utilization as well. It's not always a matter of GPU load, but how hard that work is pushing the card. Anyway, the 1080 is running around 90% GPU Usage and around 75% MCU Usage at 3x on these tasks. The 1070 is at around 94% GPU Load and 85% MCU Usage.
I may try swapping the cards around in the near future to see if that does anything.
One more interesting note, I recently updated the driver on the 1080 system to the latest, and decreased my task times at 3x by 10 minutes each. Before this, the 1070 was actually completing work FASTER than the 1080.