Nvidia Pascal and AMD Polaris, starting with GTX 1080/1070, and the AMD 480
Language
Copyright © 2024 Einstein@Home. All rights reserved.
Copyright © 2024 Einstein@Home. All rights reserved.
Keith Myers wrote: Looks to
)
That depends on the user situation. If you own several PCs which already run 24/7 and have vacant PCIe slots waiting for a graphics card or two, then the cheapest purchase price per unit Einstein productivity drives you to the 1060 3G.
But for many people at least part of the host PC price needs to be considered. Some people build entire PCs with no use save Einstein production, in which case the entire cost of new components purchased for the build needs to be considered. This pushes people in that situation in the direction of more performance per box, even if the graphics card price per unit performance is somewhat above the lowest available.
Otherwise, before Pascal, we Nvidia users would all have been building nothing but GTX 750 cards into our boxes. Now it was a fine little card, and still is remarkably competitive for purchase price efficiency, and more than competitive for power efficiency, but many of us want to get more out of a box than that can give.
For people who don't have off-site rack space (think Gary Roberts), it also may not be attractive to fill the house with more boxes.
I have a little trouble seeing the use of a single 1080, which may struggle to beat a pair of 1060 6GB cards that cost less), but a dual configuration of that may make sense for some users, and all the other three currently announced cards appear potentially attractive to some builds, in my opinion. The price slope seems to mean that in most cases a dual-card version using lower cards will beat single-card options. We don't yet see a Pascal generation equivalent of the GTX 960, which in Einstein application gained little praise (they cut back too drastically on the memory resources on that one). Time will tell if there is a "new 750Ti" coming our way.
archae86 wrote:Time will tell
)
One of the big advantages of the GTX 750 Ti (I bought six of them) is their relatively high memory bandwidth as compared to the number of shaders. However, all the Pascal cards now seem a bit memory bandwidth limited as I understand the numbers, even if not as much as the GTX 960, which is a good card for other uses. I expect that is because both Nvidia and AMD were originally planning for HBM2 memory by now, which did not happen, so they had to fall back to Plan B.
I think when we see the next generation of cards with HBM2, the story will be different.
archae86 wrote:With all those
)
Thanks, that's very close to my number ofr the 1060 6 GB. What about the 3 GB version? Is it more power efficeint than the 6 GB model? That's what I should have asked in the 1st place.
And you're right, these numbers capture none of the factors "surrounding" the card. But I don't think they should, as those would differ from PC to PC anyway. If one wants to know it exactly, one must include the specific system. Which is not "the power consumption of the card itself".
MrS
Scanning for our furry friends since Jan 2002
ExtraTerrestrial Apes
)
I only took a very short average on the 3GB, a couple of days ago, and it was reported as so very little less than the 6GB as to mean if true it is actually slightly less power efficient. One reason I don't trust the numbers at that level of detail is the surprisingly high idle power it showed, whereas multiple reviewers have commented favorably on low Pascal card idle power. Responsible reviewers don't rely on the self-reported numbers.
As fortune would have it, I may get an opportunity for 1060 6GB vs. 3GB comparison by swap sometime soon. It seems sensible to move the dual-fan 1060 6GB into my more poorly ventilated case, and the single-fan 3GB into the case which has fans on five sides. If I do, I'll try to take careful comparison measurement of wall-socket power for the fully idle state in the two configurations, and also take comparison measurements differing only in enabling 3X BRP6 on the 1060.
I had expected the 3GB to burn enough less power than the 6GB to have a power efficiency advantage. Of course there are two different brands involved here, plus my doubts on the reporting accuracy, but taken at face value it could mean that Nvidia does not implement an effective method of saving the power consumption of the unused part of the GPU chip in this particular case.
I cannot see in einstein
)
I cannot see in einstein preferences how to tell einstein to only snd gpu tasks isntead of cpu one (on gtx 1070).
Edit: Einstein runs well on
)
Edit: Einstein runs well on GTX 1060/1070
@Xixou: Account / preferences
)
@Xixou: Account / preferences / project / use CPU (2nd point) set to "no".
@Peter: the cards should have the same default power limit, or at least that's what the same TDP implies. If the 6 GB is power limited (should not be the case at 95 W), the 3 GB with less shaders should be able to boost higher and reach the same power consumption with less efficiency. However, if the cards run at similar frequency & voltage, the 3 GB model may have an advantage.
MrS
Scanning for our furry friends since Jan 2002
None of my three Pascal cards
)
None of my three Pascal cards has ever reported being anywhere near power limit. Plenty of gamers get there, but I've not seen it in running BRP6/CUDA55 at Einstein. (oddly enough one of my four 750Ti cards, which arrived factory preset to higher voltage than the others, does get in the neighborhood of power limit at reachable overclock--so even on the same card model, your results may vary).
As GPU temperature rises, you can watch the core clock speed drop (slightly) and what GPU-Z calls VDDC drop (slightly). I believe one or both of these effects becomes much more drastic fairly abruptly at about 83C, but I've not explored that operating region.
Memory clock rate, on the other hand, I've not seen adjust to GPU temperature. So long as it is still running BRP6, it stays to exactly the same memory frequency as reported to GPU-Z hour after hour.
So employing extra case ventilation, or setting a higher GPU fan curve, directly gives performance improvement with these cards. However the effect is very small. The core clock does not move by very much in the plausible range of 50 to 80C, and the sensitivity of Einstein BRP6/CUDA55 productivity to Pascal core clock rate is not particularly strong.
The second point does not
)
The second point does not exists.
xixou wrote:The second point
)
You mean you don't find it? "Account" is the 2nd menu item at the top of the page, just below the "einstein home" label. If you click it a navigation row pops up below the 1st menu. Preferences is the rightmost item.
MrS
Scanning for our furry friends since Jan 2002