Apparently the Nvidia NDA-imposed silence on initial 1070 reviews has expired, so there are lots of reviews up. The great majority of the energy in them seems to be devoted to running games and posting frame rates for them in comparison to other cards. This part tells me little for my interests.
One thing I did learn from more than one of the reviews is that the cooling hardware on the Founders Edition 1070 differs from that on the 1080 by omitting use of a vapor chamber. This steals away some of the lower sound advantage you might have hoped for from a lower-power consuming card, and, at least in gaming applications, means it is still not so hard for the 1070 to reach die temperatures that cause performance limiting clock rate reductions.
I still intend to buy a 1070 shortly after they are available, and as the higher fraction of heat dumped out the rear grille is attractive to me in at least one potential usage site, it may be a Founders Edition, but if I like the card and buy two more, at this point I seriously doubt I'll use any more FE.
It remains possible that the use of 8 GHz GDDR5 on the 1070 as opposed to 10 GHz GDDR5X RAM on the 1080 may impose a severe loss in Einstein BRP6 performance, at least at base clock rates. For those willing to overclock (and with good case ventilation) the utility of the card may greatly depend on how much their particular sample of the card will successfully overclock when running the Einstein application. That is doubly something one cannot get from reviews.
I've cancelled my (backordered) FE 1080 order. Thanks to Dave we have early data on 1080 performance here at Einstein. My current plan is to watch how the non-Founder 1080 cards roll out to help choose which 1070 card to order.
It is possible that my hope/guess that the "X" in the memory used difference between the 1080 and 1070 might give the 1080 a special advantage here on Einstein BRP6 code, in which case the 1070 may yet disappoint my adjusted expectations, but finding that out would itself have some value.
The 1080 seems to be dissapointing for einstein, and the 1070 should be also.
The 1080TI with is 384 bits memory bus should be much better here at einstein.
Wondering if the new GPU architecture from ATI (Polaris) will keep the large bus as in the R9 380 (384bits) ans R9 390 (512 bits)
The 1080 seems to be dissapointing for einstein, and the 1070 should be also.
I haven't looked into it, but won't the 1070 have the same number of memory conductors as the 1080? Other things being equal, that should give it relatively good performance for the price, at least eventually.
My Fury X running only 1 task does ~170k / day and draws ~150W.
The GTX 1080 does ~190k / day but running 4 tasks. Estimated power ~130W.
I expected a better score and power consumption.
Running four WUs my system draws 180W in total.
Will do some more testing. Which tasks were CUDA only and which one OpenCL? Am I running the OpenCL right now? Sorry, but haven't been running much Einstein lately. Only GPUGRID, but getting very few WUs there at the moment.
Running four WUs my system draws 180W in total.
Will do some more testing. Which tasks were CUDA only and which one OpenCL? Am I running the OpenCL right now? Sorry, but haven't been running much Einstein lately. Only GPUGRID, but getting very few WUs there at the moment.
For clarification: your 180W is measured at the wall socket connection to the entire PC? If you have a power meter and it is convenient, the idle number would be good to have. Of course the running number excess over idle is made up of added GPU power, plus added CPU power, plus added memory power, plus added overhead hardware power, plus added power supply losses, but it all goes into the cost of GPU computation, which is the real matter of cost interest here. My own recent build box is about 55W at the wall at idle--I'll guess yours may not be so very different from that--and I'm not counting the monitor, by the way.
I believe all tasks your PC has seen so far have been just one flavor--the CUDA55 "beta test" BRP6 Parkes PMPS task. This is the dominant GPU work available here at Einstein at the moment, and you are running the more desirable of the two ways of running it (the other you would get if you turned off your acceptance of test applications in your Einstein preferences, which would be CUDA32, but there is little reason to do that save perhaps intellectual curiosity).
My GTX 750 Tis on singles BRP6 work units are averaging 7800 seconds at 60% TDP as measured by GPU-Z, or 36 watts (Win7 64-bits). I see no overwhelming reason to upgrade until Volta.
I'd be interested to know how
)
I'd be interested to know how much takes just one WU on the 1080.
BTW, HWiNFO might also show you the real power usage in W (instead of % of TDP).
FYI - one BRP6 takes ~3100 secs on HD7950, ~2150 s on Fury X.
BRP6 CUDA tasks are much better optimized than OpenCL.
-----
Apparently the Nvidia
)
Apparently the Nvidia NDA-imposed silence on initial 1070 reviews has expired, so there are lots of reviews up. The great majority of the energy in them seems to be devoted to running games and posting frame rates for them in comparison to other cards. This part tells me little for my interests.
One thing I did learn from more than one of the reviews is that the cooling hardware on the Founders Edition 1070 differs from that on the 1080 by omitting use of a vapor chamber. This steals away some of the lower sound advantage you might have hoped for from a lower-power consuming card, and, at least in gaming applications, means it is still not so hard for the 1070 to reach die temperatures that cause performance limiting clock rate reductions.
I still intend to buy a 1070 shortly after they are available, and as the higher fraction of heat dumped out the rear grille is attractive to me in at least one potential usage site, it may be a Founders Edition, but if I like the card and buy two more, at this point I seriously doubt I'll use any more FE.
It remains possible that the use of 8 GHz GDDR5 on the 1070 as opposed to 10 GHz GDDR5X RAM on the 1080 may impose a severe loss in Einstein BRP6 performance, at least at base clock rates. For those willing to overclock (and with good case ventilation) the utility of the card may greatly depend on how much their particular sample of the card will successfully overclock when running the Einstein application. That is doubly something one cannot get from reviews.
RE: I've cancelled my
)
The 1080 seems to be dissapointing for einstein, and the 1070 should be also.
The 1080TI with is 384 bits memory bus should be much better here at einstein.
Wondering if the new GPU architecture from ATI (Polaris) will keep the large bus as in the R9 380 (384bits) ans R9 390 (512 bits)
RE: The 1080 seems to be
)
I haven't looked into it, but won't the 1070 have the same number of memory conductors as the 1080? Other things being equal, that should give it relatively good performance for the price, at least eventually.
My Fury X running only 1 task
)
My Fury X running only 1 task does ~170k / day and draws ~150W.
The GTX 1080 does ~190k / day but running 4 tasks. Estimated power ~130W.
I expected a better score and power consumption.
-----
Running four WUs my system
)
Running four WUs my system draws 180W in total.
Will do some more testing. Which tasks were CUDA only and which one OpenCL? Am I running the OpenCL right now? Sorry, but haven't been running much Einstein lately. Only GPUGRID, but getting very few WUs there at the moment.
RE: Running four WUs my
)
For clarification: your 180W is measured at the wall socket connection to the entire PC? If you have a power meter and it is convenient, the idle number would be good to have. Of course the running number excess over idle is made up of added GPU power, plus added CPU power, plus added memory power, plus added overhead hardware power, plus added power supply losses, but it all goes into the cost of GPU computation, which is the real matter of cost interest here. My own recent build box is about 55W at the wall at idle--I'll guess yours may not be so very different from that--and I'm not counting the monitor, by the way.
I believe all tasks your PC has seen so far have been just one flavor--the CUDA55 "beta test" BRP6 Parkes PMPS task. This is the dominant GPU work available here at Einstein at the moment, and you are running the more desirable of the two ways of running it (the other you would get if you turned off your acceptance of test applications in your Einstein preferences, which would be CUDA32, but there is little reason to do that save perhaps intellectual curiosity).
Poor overclocking results on
)
Poor overclocking results on ASUS custom board:
http://videocardz.com/60631/asus-rog-strix-geforce-gtx-1080-offers-poor-overclocking
Meanwhile, Jeroen appears to be getting ~215k RAC each on what I assume are overclocked 980tis:
https://einsteinathome.org/host/7181095
RE: I'd be interested to
)
This night I was running only single BRP6 WUs sequentially.
It took 3095s on average (13 WUs).
Power draw was measured at the wall withouth the LCD. At idle it's 40W.
My GTX 750 Tis on singles
)
My GTX 750 Tis on singles BRP6 work units are averaging 7800 seconds at 60% TDP as measured by GPU-Z, or 36 watts (Win7 64-bits). I see no overwhelming reason to upgrade until Volta.