Nvidia Pascal and AMD Polaris, starting with GTX 1080/1070, and the AMD 480

Einstein got a huge increase in productivity on all the Nvidia cards whose owners opted to allow Beta test applications simply by using the CUDA55 code generation path (as opposed to CUDA32). That has zero awareness of even Maxwell opportunities, and of course is only provided to users aware and interested enough to select the non default option. Clearly getting more out of the Nvidia cards is not a very high priority here. I think the opportunities in the Einstein code itself and in CUDA version selection probably very greatly outweigh any driver improvement opportunities.

I'm amused at the insistence on comparing single-instance processing time. If the interest is in Einstein productivity, that makes sense when considering users running pure defaults (which many do) but for the propeller-heads debating here, the sensible comparison is at best operating point.

Just because some of the AMD cards are not functional using current drivers and code at multiplicity greater than one is no reason to force the comparison to that point save for the purpose of falsely favoring AMD.

The 750 and 750Ti cards are remarkably good power efficiency for their generation. If the 1070 or 1080 can equal them that would be a considerable improvement over the Maxwell2 cards (not to mention any AMD card). We'll see. In comparing power it is important to select equivalent measurement. Even if power reporting from the card itself as displayed by a monitoring application were accurate (which in several cases I have observed not to be true), it does not actually incorporate the adders to power dissipation elsewhere in the user system associated with running Einstein work on the GPU, including CPU, motherboard RAM, motherboard I/O, and power supply imperfect efficiency losses. For that reason I think total box power measurements are greatly to be preferred.

That is impressive. I mentioned the power efficiency of the 750 cards previously. I have one dual-750Ti PC, which runs just under 140,000 credit/day at 173.1 wall watts operation, up from about 74 watts idle. These are overclocked beyond the EVGA "superclocked" point but have been stable for months at the current clock rates.

The cards in question (the EVGA "superclocked" variant) are currently available new for $110, or "used" from Amazon Warehouse Deals for about $97 each.

So compared to your 7970 reference, this dual 750 Ti system has somewhat higher purchase price, but considerably better power efficiency than the specific 7970 system you mention. On the other hand, two 750s uses up two slots, so with beefy power supplies and very generous case cooling one could stuff more compute power in a case using the 7970 option, which is an advantage in amortizing the system overhead over higher case level production.

Personally, I continue to think the 750 cards are great entry-level Einstein participant cards. Modest cost, small size means an easy case fit, surprisingly decent productivity (especially with some time invested in overclocking) and really almost wonderful power efficiency and very few would need to invest in a better power supply to use one. When Nvidia addresses that price point in the Pascal generation, they may greatly exceed the power efficiency, but probably not sooner. It is curious that Maxwell parts were rolled out bottom-up, while Pascal is coming out upper-middle first, then extending both down and up later, according to rumor.