I've just upgraded one of my machines from a GTX 680 and a GTX 670 (both factory OCed) to a pair of GTX 780s (thank you, Craigslist!), also slightly factory OCed. I run EVGA Precision for fan control, and it also boosts the clock of the GPUs considerably (the 780s, for example are 886MHz/941MHz boost, but Precision runs them at 1019 and 1034, respectively. The removed cards were running in the 1050-1080MHz region).

The new cards, like the old ones, are running 3 WUs at once, and the GPU usage is in the upper 90% range, so running a 4th WU wouldn't seem to be helpful.

My question is: how much improvement in RAC should I expect? The machine is running in the low 40K range, about 90% or more due to the old GPUs (CPU is a 4770K Haswell with HT off due to heat problems running 3 of 4 cores with SETI).

Based on the 97% GPU utilization I see, plus the fact that they are running 5-10% slower than the 680/670, why would I see any improvement at all? And the new cards seem to be using more power, according to the UPS app I run to keep track of that.

Thanks!

---

That's a loaded question considering we just switched over to v7...lol. Prior to v7, I would have said you would have seen at least a doubling of your RAC. That is if you could keep them feed with AP's and enough CPU core time. Now with v7?..I would probably say mid 50s..maybe. credit from v7 is about 1/3 to 1/2 what it used to be. So given the new system, I would way, it will be an improvement but not as much as it would have been with the old credit system.


Happy Crunching....


Zalster

That's a loaded question considering we just switched over to v7...lol. Prior to v7, I would have said you would have seen at least a doubling of your RAC. That is if you could keep them feed with AP's and enough CPU core time. Now with v7?..I would probably say mid 50s..maybe. credit from v7 is about 1/3 to 1/2 what it used to be. So given the new system, I would way, it will be an improvement but not as much as it would have been with the old credit system.


Happy Crunching....


Zalster

Whoops - I don't want to get into THAT debate, just want to know what RAC increase I can expect for a given workload. Can I look at the APR for my machine for a given app (like MB7 for CPU and GPU)? The CPU should remain the same, and the GPU ought to go up consonant with RAC increase, yes?

In which case, assuming a new APR for the new GPU, how long must I wait for it to settle at its new level? Anyone? Buehler?

---

...just want to know what RAC increase I can expect for a given workload. Can I look at the APR for my machine for a given app (like MB7 for CPU and GPU)? The CPU should remain the same, and the GPU ought to go up consonant with RAC increase, yes?

In which case, assuming a new APR for the new GPU, how long must I wait for it to settle at its new level? Anyone? Buehler?

RAC is an overall measurement and takes on the order of 5 weeks to adjust. Given the intoduction of AP v7 there will be disturbances too.

I suggest you watch the APR for SETI@home v7 (anonymous platform, NVIDIA GPU). That will adapt fairly quickly, about 320 tasks validated from the 780s will be enough to see almost all the change. But it's also volatile enough that the ratio of shorties to midrange can cause a significant hump or valley, and pendings done by the 670 and 680 will have some effect for awhile. I guess after a week or so you should have a reasonable idea of the change.

A quick comparison of run times for midrange tasks suggests to me the 780s are about 30% faster than the 670 was, and about 25% faster than the 680.

                                                                  Joe

Thank you, Joe.

I was planning on keeping a (daily) eye on the APR, and see what happens,

---

So far, the APR for GPU processing of MB7 has jumped from 126 to 148, a gain of about 17%. Looking good, so far.

---

During the extra long "planned" outage today, my newly upgraded machine ran completely out of MB7 for the GPUs, while the other machine (dual 680s) still had a bunch left (like maybe 30 or so). So, assuming the average WU on the machines was the same in computation needed, I can safely say the 780s are roughly 20% faster than the 680s.

And I also note the following: the dual 680 machine always had more tasks in "validation pending" state - like 800+ to 700+, though it varied considerably (as we all know). Now, the dual 780s have caught up and passed the dual 680s; of course, that may be an artifact of the big surge when the outage finally cleared and I got both machines to report - depends on when my wingpersons report. I will keep an eye on that.

---

One last post:

The (now) dual 780 machine, (previously 680 + 670), which used to lag behind the dual 680s in RAC (roughly 40-42K - 44-47K on a weekly basis) has now reversed the situation, leading by roughly the same margin this past week, even allowing for the 18+ hour outage this past weekend. (I have yet to begin processing AP7 on either machine).

Also, the Average Processing Rate has gone up from about 125 to around 155-160, in line with the predicted (above)increase. (Thanks, Joe!).

---