So what you're proposing is more like a Xeon Phi Knight's Landing made by nVidia and sold at reasonable prices? (it's got x86 cores with large vactor units, which would be roughly equivalent to making each SM as smart as a full CPU core, so that it can boot its own OS and schedule & issue tasks on some reserved cores)
Peter, judging by your current numbers the GTX1070 would produce 45% more credits at a bit over 50% more shader power. This looks reasonable, especially since GTX1060 has more bandwidth per flop.
So what you're proposing is more like a Xeon Phi Knight's Landing made by nVidia and sold at reasonable prices? (it's got x86 cores with large vactor units, which would be roughly equivalent to making each SM as smart as a full CPU core, so that it can boot its own OS and schedule & issue tasks on some reserved cores)
I doubt that Nvidia would try to invade Intel's space to that degree. Intel could fight back with a surface-mount substrate themselves that integrated cores (whatever they are) and memory with a high-bandwidth memory path between them. I would think that Nvidia would just expand CUDA functionality more, but that is for their marketing geniuses to decide. If I could make those decisions better, I would be running their company, which I think is not likely.
Did the memory leak problem occur with NV 9 series cards and early drivers or did it come with later drivers? And 10 series?
I'm not confident of any of the scoping. There seems to be a bug somewhere, but a variety of conditions seem to be able to enable or disable the circumstances under which the bug has this substantial effect.
I have personally observed the problem in full form on two different Windows 10 systems, with a GTX 1070 card running recent (obviously) drivers, and a GTX 970 card running two different drivers from before 2016. On the very same systems, running the same drivers, I have seen the problem NOT manifest when the same Einstein code was only running on a GTX 750Ti card (i.e. jobs for the troubled card suspended). A third Windows 10 system, also built by me so at least somewhat similarly set up, does not show the problem running the same Einstein code (BRP6/CUDA55) on a pair of GTX 750Ti card. So for some reason there appears likely to be a degree of card dependence.
But results across my small observation base are inconsistent on application dependence (currently on my GTX 1070/1060 running system, the BRP4G/CUDA32 application does not suffer the problem, while the BRP6/CUDA55 does, but on my GTX970 system, currently, both applications suffer, but before a driver change that one also lacked the problem on BRP4G).
Some similar observations from other users are in hand, also with somewhat inconsistent results regarding application and driver dependence.
As the offending excess memory consumption is in a paged pool, I think doctrine is that a driver must be involved, as I think applications don't run at sufficient privilege level to allocate pool memory. But that does not settle whether it is a pure driver bug, or whether the application is asking the driver to do something it should not.
It has been dormant for a while, but in Problems and bug Reports there is a thread on just this topic.
Hi archae86,
Drat, win10 can be awkward, and unfortunately I have 2 rigs running it:-(
However my main rig runs win7 ult and I've not seen any problems, but then it gets rebooted daily.
I'll make sure both my win10 rigs get daily reboots, which should I 'think' work around the problem.
I'll make sure both my win10 rigs get daily reboots, which should I 'think' work around the problem.
Better check if you have the problem at all before you take such "drastic" steps. It's more work for you and steals some of your precious crunching time.. would be a shame if it's not needed.
I noticed early on that my GTX 1060 was actually labelled as GTX 1060 6GB, hinting that another memory size 1060 might be coming.
There was, and a 3GB version has arrived. They are a bit trickier to find, as NowinStock so far does not have a listing for them, and in a quick search on Amazon I did not spot any, but here are a couple of examples at NewEgg.
The claimed MSRP base for the 1060 3GB is $199.99 (down from $249.99 for the 6GB), and these two meet that. However one must be aware that the relationship to the 6GB 1060 is not only memory size, but also a modest reduction in compute units. In Nvidia terminology, the number of core compute units comes down from 1280 to 1152, and of TMUs from 80 to 72. But the memory provision appears unchanged save for the factor of two reduction in capacity. As Einstein cares a lot about memory, I'd guess these will be down less than 10% in Einstein performance from the 1060 6GB, so an appreciable improvement on an already good value proposition.
I'd think the cheaper 1060 3GB cards would be pretty nice price performers at Einstein, and very nice power performers. Compared to "serious gamer" cards of the older days, even the regular 1060 is pretty modest in real power supply requirements, so I suspect many people could swap a 3GB 1060 into an existing moderate PC without a power supply upgrade. Also, the two cards I've linked are "shorties" which should fit easily in most cases that find the longer cards awkward or even impossible.
Were it not for the unsolved memory leak problem, I, myself, would immediately purchase one thinking it a good fit in one more more slots in my three systems. But with that problem outstanding, I cannot in good conscience recommend Nvidia card purchases except for the modest 750/750Ti to Windows 10 PC targets with Einstein as a primary use.
NVidia has scheduled the release of gTX 1060 3 GB for Q4 2016
Source? Sounds like a rumor mill number to me. Nvidia has not been making official release date disclosures anywhere near that far in advance for the Pascal cards.
Quote:
sounds like availability is still far away
The EVGA card I listed above shows in stock at NewEgg as I type, although the Zotac shows out of stock.
Quote:
Peter, have you tried the new 372 driver?
That one released three days ago and I missed it. Nothing in the release notes hints it will help me, but I'll plan to update my 1070/1060 PC very soon. Thanks.
So what you're proposing is
)
So what you're proposing is more like a Xeon Phi Knight's Landing made by nVidia and sold at reasonable prices? (it's got x86 cores with large vactor units, which would be roughly equivalent to making each SM as smart as a full CPU core, so that it can boot its own OS and schedule & issue tasks on some reserved cores)
MrS
Scanning for our furry friends since Jan 2002
Peter, judging by your
)
Peter, judging by your current numbers the GTX1070 would produce 45% more credits at a bit over 50% more shader power. This looks reasonable, especially since GTX1060 has more bandwidth per flop.
MrS
Scanning for our furry friends since Jan 2002
RE: So what you're
)
I doubt that Nvidia would try to invade Intel's space to that degree. Intel could fight back with a surface-mount substrate themselves that integrated cores (whatever they are) and memory with a high-bandwidth memory path between them. I would think that Nvidia would just expand CUDA functionality more, but that is for their marketing geniuses to decide. If I could make those decisions better, I would be running their company, which I think is not likely.
Hi archae86, Did the memory
)
Hi archae86,
Did the memory leak problem occur with NV 9 series cards and early drivers or did it come with later drivers? And 10 series?
Regards,
Cliff,
Been there, Done that, Still no damm T Shirt.
cliff wrote:Did the memory
)
I'm not confident of any of the scoping. There seems to be a bug somewhere, but a variety of conditions seem to be able to enable or disable the circumstances under which the bug has this substantial effect.
I have personally observed the problem in full form on two different Windows 10 systems, with a GTX 1070 card running recent (obviously) drivers, and a GTX 970 card running two different drivers from before 2016. On the very same systems, running the same drivers, I have seen the problem NOT manifest when the same Einstein code was only running on a GTX 750Ti card (i.e. jobs for the troubled card suspended). A third Windows 10 system, also built by me so at least somewhat similarly set up, does not show the problem running the same Einstein code (BRP6/CUDA55) on a pair of GTX 750Ti card. So for some reason there appears likely to be a degree of card dependence.
But results across my small observation base are inconsistent on application dependence (currently on my GTX 1070/1060 running system, the BRP4G/CUDA32 application does not suffer the problem, while the BRP6/CUDA55 does, but on my GTX970 system, currently, both applications suffer, but before a driver change that one also lacked the problem on BRP4G).
Some similar observations from other users are in hand, also with somewhat inconsistent results regarding application and driver dependence.
As the offending excess memory consumption is in a paged pool, I think doctrine is that a driver must be involved, as I think applications don't run at sufficient privilege level to allocate pool memory. But that does not settle whether it is a pure driver bug, or whether the application is asking the driver to do something it should not.
It has been dormant for a while, but in Problems and bug Reports there is a thread on just this topic.
Hi archae86, Drat, win10 can
)
Hi archae86,
Drat, win10 can be awkward, and unfortunately I have 2 rigs running it:-(
However my main rig runs win7 ult and I've not seen any problems, but then it gets rebooted daily.
I'll make sure both my win10 rigs get daily reboots, which should I 'think' work around the problem.
Regards
Cliff,
Been there, Done that, Still no damm T Shirt.
cliff_9 wrote:I'll make sure
)
Better check if you have the problem at all before you take such "drastic" steps. It's more work for you and steals some of your precious crunching time.. would be a shame if it's not needed.
MrS
Scanning for our furry friends since Jan 2002
I noticed early on that my
)
I noticed early on that my GTX 1060 was actually labelled as GTX 1060 6GB, hinting that another memory size 1060 might be coming.
There was, and a 3GB version has arrived. They are a bit trickier to find, as NowinStock so far does not have a listing for them, and in a quick search on Amazon I did not spot any, but here are a couple of examples at NewEgg.
Zotac 1060 3GB mini
EVGA 1060 3GB mini
The claimed MSRP base for the 1060 3GB is $199.99 (down from $249.99 for the 6GB), and these two meet that. However one must be aware that the relationship to the 6GB 1060 is not only memory size, but also a modest reduction in compute units. In Nvidia terminology, the number of core compute units comes down from 1280 to 1152, and of TMUs from 80 to 72. But the memory provision appears unchanged save for the factor of two reduction in capacity. As Einstein cares a lot about memory, I'd guess these will be down less than 10% in Einstein performance from the 1060 6GB, so an appreciable improvement on an already good value proposition.
Videocardz page including 1060 6GB to 3GB particulars
I'd think the cheaper 1060 3GB cards would be pretty nice price performers at Einstein, and very nice power performers. Compared to "serious gamer" cards of the older days, even the regular 1060 is pretty modest in real power supply requirements, so I suspect many people could swap a 3GB 1060 into an existing moderate PC without a power supply upgrade. Also, the two cards I've linked are "shorties" which should fit easily in most cases that find the longer cards awkward or even impossible.
Were it not for the unsolved memory leak problem, I, myself, would immediately purchase one thinking it a good fit in one more more slots in my three systems. But with that problem outstanding, I cannot in good conscience recommend Nvidia card purchases except for the modest 750/750Ti to Windows 10 PC targets with Einstein as a primary use.
NVidia has scheduled the
)
NVidia has scheduled the release of gTX 1060 3 GB for Q4 2016.. sounds like availability is still far away,. but Q4 is almost around the corner.
Peter, have you tried the new 372 driver?
Scanning for our furry friends since Jan 2002
ExtraTerrestrial Apes
)
Source? Sounds like a rumor mill number to me. Nvidia has not been making official release date disclosures anywhere near that far in advance for the Pascal cards.
The EVGA card I listed above shows in stock at NewEgg as I type, although the Zotac shows out of stock.
That one released three days ago and I missed it. Nothing in the release notes hints it will help me, but I'll plan to update my 1070/1060 PC very soon. Thanks.