Founders Edition vs. Reference, fan considerations
Initial press comments on the "Founders Edition" cards sprayed around a bit, but comments seem to have converged more recently toward these points:
1. these are not built by Nvidia themselves by their partners.
2. but they are built closely to an Nvidia design, and quite likely some components (the shroud) are only from a specific source, and should be extremely similar whether assembled for Asus, EVGA, Gigabyte, or ...
3. Despite the name change and attempted pricing premium, most commenters think these cards are very analogous to the reference cards of previous generations.
4. Like those reference cards, these cards have a single fan, in this case located far from the I/O panel end of the card, with ducting assuring that the very high majority of the heated air exits the case through the grillwork of the card at the I/O panel. This is generally held to be a substantial difference from the great majority of card designs, which generally do push a good bit of their heated air out the back, but also share a good bit with the rest of the interior of the case.
5. Many reviewers and commenters hold that in the past the higher-spec cards private labelled by partners have offered superior performance to reference cards, derived from superior cooling, better power regulation/delivery, and cherry picking of the faster GPU parts from the distribution.
Some details of my circumstances leave me thinking that I'll opt for an early Founders Edition 1080. I'll watch for indications of how quickly other variants become available, their characteristics, and price, before deciding on a comparison 1070 card.
It continues to be my guess that for people with a primary Einstein computing interest in these cards, the 1070 may well be the superior option in price/performance and possibly (depending on host cost allocations) in power/performance.
A lot depends on how well the Einstein code likes these cards and their drivers, which can only be found by trial.
Founders Edition vs. Reference, fan considerations
It continues to be my guess that for people with a primary Einstein computing interest in these cards, the 1070 may well be the superior option in price/performance and possibly (depending on host cost allocations) in power/performance.
A lot depends on how well the Einstein code likes these cards and their drivers, which can only be found by trial.
Yup, its IMHO the drivers that are the problem:-/ Along with the OS.
Before a recent problem with Win10[now removed due to boot problems] it was crunching faster than Win7Ult.. with early drivers.
Now I'm back to Win7Ult and the same drivers crunching is slower.. Still I've put Win10 in a new ssd and am waiting to swap crunching over to that OS in the near future.. When I can be sure I don't end up with the same PC being recorded twice due to the OS change.
Dunno if NVidia will release new drivers specifically for the 10 series, if not I'll stick with my present drivers as and when funds permit a new 1080:-)
any of you have any feeling for the 1060 scheduled for a fall release. Thumbs up/down for this one? From my earlier post looked like a low power user. But not much info on performance that I could find.
Apparently the release from NDA (Non-disclosure agreement) restriction for the the reviewers given early samples and information was about an hour and a half ago.
Many reviews are up now, as a quick search using a string such as "GTX 1080 Review" will show.
At the moment the web site VideoCardz appears to be maintaining a frequently updated list of links to 1080 review here.
Of the ones I've looked at so far, the most substantial review to my personal taste was this one Tom's Hardware GTX 1080 review
Nvidia has not yet, I think published any appreciable detail on even the 1070 besides price, while 1060 guesses at this point are just guesses.
I still intend to make a 1080 purchase as soon as I spot availability, and a 1070 very soon after they are available. I continue to guess the 1070 is likely to be a much better buy in performance per purchase dollar.
If you are interested I suggest you review this page at anandtech.
Highlights as I see them:
The 1070 uses the same GPU die as does the 1080.
1920 of 2560 "CUDA cores" are enabled
All 64 ROPs are enabled
120 of 160 texture units are enabled
Nominal core and boost clock rates are modestly lower
The memory type used is 8 Gbps GDDR5 instead of 10 GBps GDDR5X
TDP is 150 instead of 180W
From a marketing product positioning perspective, Nvidia is pushing the 1070 as a nice upgrade for 970 users or earlier generation mid-range products.
If the 1080 turns out to be "worth the price" over the 1070 for Einstein, I imagine most likely the reason would be memory performance. It all depends on whether the "X" in the memory type name and the 10 vs. 8 in the memory speed translate into a really great improvement for the applications here. Also crucially important is how successful the software and drivers are at employing the additional parallel resources.
One interesting question is what the power consumption will turn out to be when running Einstein work (for both 1070 and 1080). My own 1070 running Einstein has never gotten to the built-in power dissipation limit, even under rather substantial overclocking. On the other hand lots of reviewers report getting to the 1080 power limit in actual game play.
My personal bets--subject to revision on exposure to data--
The 1070 would be a nice upgrade for almost any Nvidia user here not running 980 or 980Ti.
The 1070 will give much more performance per purchase dollar than the 1080 at the card level
System level performance per purchase dollar may be better for the 1080 when considering the cost of the "everything else" in addition to the card.
Rumors continue to fly regarding likely availability sequencing and the proper description of the Founders Edition.
My updated understanding:
Founders Edition cards will be sold under many labels, not just the Nvidia label.
The card houses have been working on their non FE cards for some time, and are in an advanced state of development on at least some of them.
Cards other than FE are likely to be available rather soon after the May 27/June initial shipment dates for FE 1080/1070.
It appears that the previous standard card control utilities are incompletely able to control 1080 overclocking, or perhaps not at all. Reviewers report having to use a buggy preliminary release of PrecisionX which only got to them shortly before NDA release date for initial reviews.
I'm seeing 6.5 TFlops for the 1070, and 8.9 TFlops for the 1080. Assuming that translates directly into increased E@H performance, the 1080 is 50% more expensive ($600 vs $400) for 37% increase in performance, but TDP is only 20% higher, so potentially could approach the 1070 in cost per unit over the life of the card.
Looking forward to seeing some work unit completions!
I just noticed something apropos to Windows boxes.
I have several 7970s. They are almost exactly 4x more powerful at crunching BRP6 tasks than my GTX 650 (~36k for GTX 650, 150k-160k for 7970).
According to this table, a 980 is about 2x the speed of a GTX 650. This means it's about 0.5 the speed of a 7970 for E@H (so, 75k-80k units per day on BRP6?).
If I have been reading correctly, the 1080 is supposed to have a little less than 2x the computing performance of a 980. If the performance issue with Maxwell-generation cards persists into Pascal, and my previous assumptions are correct, then a 1080 will perform about the speed of a stock 7970, or 150-160k units a day doing BRP6. It will in theory do so with better TDP (180w for 1080 vs 250w for 7970), but my 7970 only draws 160 extra watts doing 3 BRP6.
I can't compare on Linux, as there isn't enough data for AMD cards, but I see the 980 is rated 7x (!) as good as a 650.
As I get about 130,000 cobblestones/day out of a 970 running BRP6, I think it rather likely the 1080 will do very much better than the number you propose. Admittedly my 970 has significant memory overclock and a modest core clock overclock. But I think your suggestion is likely rather a long way below the truth.
I hope to report validated units within ten days. We'll see.
As I get about 130,000 cobblestones/day out of a 970 running BRP6, I think it rather likely the 1080 will do very much better than the number you propose. Admittedly my 970 has significant memory overclock and a modest core clock overclock. But I think your suggestion is likely rather a long way below the truth.
I hope to report validated units within ten days. We'll see.
That would seem to indicate a problem with the tables. Any idea how they are populated? The ratios posted are fairly consistent with my results for my 650, 750ti, 950 and 960, but my only high-end cards are a 7950 and two 7970s.
On another note, NewEgg finally has pages for the 1080 Founder's:
That would seem to indicate a problem with the tables.....
I'm not at all a betting man but I'd be prepared to wager my entire credit stash on one simple proposititon :-).
I believe the tables don't take into account the concurrency with which the GPU tasks have been crunched.
Here's a scenario to think about. I run Linux and I have lots of HD 7850 (Pitcairn) GPUs. The table rates them at 0.208 compared to a Tahiti at 0.538. In other words, the ratio between them is 2.59 to 1.
My Pitcairns all run 4x and are run on a variety of both AMD and Intel boards from quite recent to eight years old. The results are surprisingly consistent across all architectures (this wasn't always the case) and really indicates what a great job Bikeman (Heinz-Bernd) did in making the app run at top efficiency on the different PCIe generations. Over all these different hosts, the crunch time per task averages out to be around 77 mins with very little deviation, maybe +/- a couple of minutes.
I recently bought a Tahiti series GPU on ebay. It's running 6x and currently averaging 46 mins per task. I'm still testing but I feel it may be at its 'best' efficiency for my purposes. So I would contend that the true ratio between a Tahiti and a Pitcairn is more like 1.67 to 1. So why the big difference from what the table gives?
I would contend that all my Pitcairn results are being taken as 308 mins (the actual running time of a task) without any regard for the fact that the concurrency of 4x makes it really 77 mins per task. So nasty people like me are skewing the results to make them look worse than they really are :-). Of course I don't know this for a fact so I am only guessing :-).
BTW, thanks very much for the link. I'm getting so old and decrepit that I'd completely forgotten that such information existed. I had seen it quite a while ago but had completely forgotten about it.
EDIT: I should also make it very clear that this isn't the only factor - there are probably quite a few. An obvious one is that there are big differences in the performance of a given app type over time. CUDA32 compared to CUDA55 is one that's currently in play. Another is that the very high-end AMD GPUs (Hawaii??) return invalid results above a concurrency of 1x. This would skew the relative worth of affected GPUs to look better than they really are because some of the GPUs below them are going to be running 2x, 3x, etc, and adding longer run times to the stats.
The 'take-home' message should be to treat the numbers as a rough guide only and hope that early adopters like Archae86 give us the 'real deal'. It's great that there are people willing to do this and, particularly, willing to document it to the level of detail that gives great confidence that the numbers are accurate.
Founders Edition vs.
)
Founders Edition vs. Reference, fan considerations
Initial press comments on the "Founders Edition" cards sprayed around a bit, but comments seem to have converged more recently toward these points:
1. these are not built by Nvidia themselves by their partners.
2. but they are built closely to an Nvidia design, and quite likely some components (the shroud) are only from a specific source, and should be extremely similar whether assembled for Asus, EVGA, Gigabyte, or ...
3. Despite the name change and attempted pricing premium, most commenters think these cards are very analogous to the reference cards of previous generations.
4. Like those reference cards, these cards have a single fan, in this case located far from the I/O panel end of the card, with ducting assuring that the very high majority of the heated air exits the case through the grillwork of the card at the I/O panel. This is generally held to be a substantial difference from the great majority of card designs, which generally do push a good bit of their heated air out the back, but also share a good bit with the rest of the interior of the case.
5. Many reviewers and commenters hold that in the past the higher-spec cards private labelled by partners have offered superior performance to reference cards, derived from superior cooling, better power regulation/delivery, and cherry picking of the faster GPU parts from the distribution.
Some details of my circumstances leave me thinking that I'll opt for an early Founders Edition 1080. I'll watch for indications of how quickly other variants become available, their characteristics, and price, before deciding on a comparison 1070 card.
It continues to be my guess that for people with a primary Einstein computing interest in these cards, the 1070 may well be the superior option in price/performance and possibly (depending on host cost allocations) in power/performance.
A lot depends on how well the Einstein code likes these cards and their drivers, which can only be found by trial.
RE: Founders Edition vs.
)
Yup, its IMHO the drivers that are the problem:-/ Along with the OS.
Before a recent problem with Win10[now removed due to boot problems] it was crunching faster than Win7Ult.. with early drivers.
Now I'm back to Win7Ult and the same drivers crunching is slower.. Still I've put Win10 in a new ssd and am waiting to swap crunching over to that OS in the near future.. When I can be sure I don't end up with the same PC being recorded twice due to the OS change.
Dunno if NVidia will release new drivers specifically for the 10 series, if not I'll stick with my present drivers as and when funds permit a new 1080:-)
Regards,
Cliff,
Been there, Done that, Still no damm T Shirt.
any of you have any feeling
)
any of you have any feeling for the 1060 scheduled for a fall release. Thumbs up/down for this one? From my earlier post looked like a low power user. But not much info on performance that I could find.
Apparently the release from
)
Apparently the release from NDA (Non-disclosure agreement) restriction for the the reviewers given early samples and information was about an hour and a half ago.
Many reviews are up now, as a quick search using a string such as "GTX 1080 Review" will show.
At the moment the web site VideoCardz appears to be maintaining a frequently updated list of links to 1080 review here.
Of the ones I've looked at so far, the most substantial review to my personal taste was this one Tom's Hardware GTX 1080 review
Nvidia has not yet, I think published any appreciable detail on even the 1070 besides price, while 1060 guesses at this point are just guesses.
I still intend to make a 1080 purchase as soon as I spot availability, and a 1070 very soon after they are available. I continue to guess the 1070 is likely to be a much better buy in performance per purchase dollar.
Finally some more details
)
Finally some more details regarding the GTX 1070.
If you are interested I suggest you review this page at anandtech.
Highlights as I see them:
The 1070 uses the same GPU die as does the 1080.
1920 of 2560 "CUDA cores" are enabled
All 64 ROPs are enabled
120 of 160 texture units are enabled
Nominal core and boost clock rates are modestly lower
The memory type used is 8 Gbps GDDR5 instead of 10 GBps GDDR5X
TDP is 150 instead of 180W
From a marketing product positioning perspective, Nvidia is pushing the 1070 as a nice upgrade for 970 users or earlier generation mid-range products.
If the 1080 turns out to be "worth the price" over the 1070 for Einstein, I imagine most likely the reason would be memory performance. It all depends on whether the "X" in the memory type name and the 10 vs. 8 in the memory speed translate into a really great improvement for the applications here. Also crucially important is how successful the software and drivers are at employing the additional parallel resources.
One interesting question is what the power consumption will turn out to be when running Einstein work (for both 1070 and 1080). My own 1070 running Einstein has never gotten to the built-in power dissipation limit, even under rather substantial overclocking. On the other hand lots of reviewers report getting to the 1080 power limit in actual game play.
My personal bets--subject to revision on exposure to data--
The 1070 would be a nice upgrade for almost any Nvidia user here not running 980 or 980Ti.
The 1070 will give much more performance per purchase dollar than the 1080 at the card level
System level performance per purchase dollar may be better for the 1080 when considering the cost of the "everything else" in addition to the card.
Rumors continue to fly regarding likely availability sequencing and the proper description of the Founders Edition.
My updated understanding:
Founders Edition cards will be sold under many labels, not just the Nvidia label.
The card houses have been working on their non FE cards for some time, and are in an advanced state of development on at least some of them.
Cards other than FE are likely to be available rather soon after the May 27/June initial shipment dates for FE 1080/1070.
It appears that the previous standard card control utilities are incompletely able to control 1080 overclocking, or perhaps not at all. Reviewers report having to use a buggy preliminary release of PrecisionX which only got to them shortly before NDA release date for initial reviews.
I'm seeing 6.5 TFlops for the
)
I'm seeing 6.5 TFlops for the 1070, and 8.9 TFlops for the 1080. Assuming that translates directly into increased E@H performance, the 1080 is 50% more expensive ($600 vs $400) for 37% increase in performance, but TDP is only 20% higher, so potentially could approach the 1070 in cost per unit over the life of the card.
Looking forward to seeing some work unit completions!
I just noticed something
)
I just noticed something apropos to Windows boxes.
I have several 7970s. They are almost exactly 4x more powerful at crunching BRP6 tasks than my GTX 650 (~36k for GTX 650, 150k-160k for 7970).
According to this table, a 980 is about 2x the speed of a GTX 650. This means it's about 0.5 the speed of a 7970 for E@H (so, 75k-80k units per day on BRP6?).
https://einstein.phys.uwm.edu/gpu_list.php
If I have been reading correctly, the 1080 is supposed to have a little less than 2x the computing performance of a 980. If the performance issue with Maxwell-generation cards persists into Pascal, and my previous assumptions are correct, then a 1080 will perform about the speed of a stock 7970, or 150-160k units a day doing BRP6. It will in theory do so with better TDP (180w for 1080 vs 250w for 7970), but my 7970 only draws 160 extra watts doing 3 BRP6.
I can't compare on Linux, as there isn't enough data for AMD cards, but I see the 980 is rated 7x (!) as good as a 650.
RE: 150-160k units a day
)
As I get about 130,000 cobblestones/day out of a 970 running BRP6, I think it rather likely the 1080 will do very much better than the number you propose. Admittedly my 970 has significant memory overclock and a modest core clock overclock. But I think your suggestion is likely rather a long way below the truth.
I hope to report validated units within ten days. We'll see.
RE: RE: 150-160k units a
)
That would seem to indicate a problem with the tables. Any idea how they are populated? The ratios posted are fairly consistent with my results for my 650, 750ti, 950 and 960, but my only high-end cards are a 7950 and two 7970s.
On another note, NewEgg finally has pages for the 1080 Founder's:
http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&DEPA=0&N=8000&Order=BESTMATCH&Description=PPSSVYJXPPLKEI
If they all look the same and have the same specs, sales figures would be revealing for brand loyalty.
RE: That would seem to
)
I'm not at all a betting man but I'd be prepared to wager my entire credit stash on one simple proposititon :-).
I believe the tables don't take into account the concurrency with which the GPU tasks have been crunched.
Here's a scenario to think about. I run Linux and I have lots of HD 7850 (Pitcairn) GPUs. The table rates them at 0.208 compared to a Tahiti at 0.538. In other words, the ratio between them is 2.59 to 1.
My Pitcairns all run 4x and are run on a variety of both AMD and Intel boards from quite recent to eight years old. The results are surprisingly consistent across all architectures (this wasn't always the case) and really indicates what a great job Bikeman (Heinz-Bernd) did in making the app run at top efficiency on the different PCIe generations. Over all these different hosts, the crunch time per task averages out to be around 77 mins with very little deviation, maybe +/- a couple of minutes.
I recently bought a Tahiti series GPU on ebay. It's running 6x and currently averaging 46 mins per task. I'm still testing but I feel it may be at its 'best' efficiency for my purposes. So I would contend that the true ratio between a Tahiti and a Pitcairn is more like 1.67 to 1. So why the big difference from what the table gives?
I would contend that all my Pitcairn results are being taken as 308 mins (the actual running time of a task) without any regard for the fact that the concurrency of 4x makes it really 77 mins per task. So nasty people like me are skewing the results to make them look worse than they really are :-). Of course I don't know this for a fact so I am only guessing :-).
BTW, thanks very much for the link. I'm getting so old and decrepit that I'd completely forgotten that such information existed. I had seen it quite a while ago but had completely forgotten about it.
EDIT: I should also make it very clear that this isn't the only factor - there are probably quite a few. An obvious one is that there are big differences in the performance of a given app type over time. CUDA32 compared to CUDA55 is one that's currently in play. Another is that the very high-end AMD GPUs (Hawaii??) return invalid results above a concurrency of 1x. This would skew the relative worth of affected GPUs to look better than they really are because some of the GPUs below them are going to be running 2x, 3x, etc, and adding longer run times to the stats.
The 'take-home' message should be to treat the numbers as a rough guide only and hope that early adopters like Archae86 give us the 'real deal'. It's great that there are people willing to do this and, particularly, willing to document it to the level of detail that gives great confidence that the numbers are accurate.
Cheers,
Gary.