Silent_Buddha
Legend
It does if the they expect to make greater inroads to the high margin HPC market. Not saying this is the case or that they would do this, of course. But this whole line is a "hypothetical" thought exercise anyway.
There's likely a minimum run of wafers for any order if they wish to be profitable keep per chip costs down. Only Nvidia knows where that breakeven point is where additional wafers don't significantly lower the per chip cost.
I'd imagine even with greater penetration into the HPC market it's still going to represent a low percentage of product compared to the professional market which could also benefit from this. And any remaining product can be sold to the consumer/prosumer market.
In other words, to prevent the per chip cost from being astronomical, it would make sense to have a chip capable of being sold in other markets than just HPC.
Enthusiasts have shown they are more than willing to spend a 200-700 USD price premium for 10-20% more performance (video cards and CPUs), so even if it's extremely large compared to the performance gain, it could still be priced such that there's some margin.
It may be counterintuitive from a consumer standpoint but it's business as usual for corporations which also sell product to business/professional customers.
Heck just look at IVB-E or SB-E. Both products offer extremely minimal gain for the consumer market (almost everything added is mainly of benefit in the professional/business market), yet it is still available to consumers and as you can see even on this forum, Enthusiasts are more than willing to pay that price premium.
Regards,
SB
Exactly, hence it can't be just a GK104+HPC. Nobody would buy it.
Blazkowicz
Legend
I was thinking just that, GK110 could be essentially the same as 104 but with possibly even a 512bit bus, to enable 8GB on a card.
all the gaming stuff gets used on a Quadro, too. we've seen nvidia using a GF100 with only 256SP rather than a cut down GF104 by the way.
I understand skepticism but that idea is not all too different conceptually from the GF100/GF104 situation.
"nobody" buys a gtx 580 either, an order of magnitude more gamers buy a 560 ti instead.
another idea : maybe the GK104 can eat all the power budget already if full-specced and clocked high enough.
a GK110 wouldn't do much better for gaming except if it's sold as a monster 300W card.
Last edited by a moderator:
It does if the they expect to make greater inroads to the high margin HPC market. Not saying this is the case or that they would do this, of course. But this whole line is a "hypothetical" thought exercise anyway.
There's likely a minimum run of wafers for any order if they wish to be profitable keep per chip costs down. Only Nvidia knows where that breakeven point is where additional wafers don't significantly lower the per chip cost.
I'd imagine even with greater penetration into the HPC market it's still going to represent a low percentage of product compared to the professional market which could also benefit from this. And any remaining product can be sold to the consumer/prosumer market.
In other words, to prevent the per chip cost from being astronomical, it would make sense to have a chip capable of being sold in other markets than just HPC.
Enthusiasts have shown they are more than willing to spend a 200-700 USD price premium for 10-20% more performance (video cards and CPUs), so even if it's extremely large compared to the performance gain, it could still be priced such that there's some margin.
It may be counterintuitive from a consumer standpoint but it's business as usual for corporations which also sell product to business/professional customers.
Heck just look at IVB-E or SB-E. Both products offer extremely minimal gain for the consumer market (almost everything added is mainly of benefit in the professional/business market), yet it is still available to consumers and as you can see even on this forum, Enthusiasts are more than willing to pay that price premium.
Regards,
SB
Once you have GK104 taped out, it's going to cost an insane amount of time and money to remove the graphics pipeline from GK110, even if no Tesla consumer is going to need those.
Last time I checked, GF110 had 33% more CCs than GF104/114 and had greater overall performance. You don't need a huge performance delta to cash in on a much higher price, which is exactly the point. It is not a significant investment in die size cost for that extra performance, so it is always worth doing.Blazkowicz said:
Please explain.
Because you might also want to sell it in the CAD/CAM/CAE area where you might need a little texturing, but then, maybe you're right! But that (maybe I've hidden the link too well in my earlier posting) would probably mean even less gaming prowess, not "only a little more" due to higher memory bandwidth.
If it weren't for higher bandwidth and larger framebuffers and thus probably just a little more punch in high-res, high-antialiasing gaming. But if they really use what I've linke above, it would be even GK104-minus-something-plus-HPC...
Could be there a solution to beef up geometry processing and tessellation on GK106 and GK107?
The tiny Cape Verde has a tessellation performance like GF114:
http://www.hardware.fr/articles/855-8/performances-theoriques-geometrie.html
http://www.tweakpc.de/hardware/test...on_hd_7770/benchmarks.php?benchmark=heaven2ex
This would be far above a 1 GPC GK107 and on a level with a 2 GPC GK106, if they follow the Fermi geometry/tessellation performance scaling.
Or could be Cape Verde just be overpowered in this terms, especial if you compare it with Tahitis performance, which could be driver problem or broken design?
But on the other side, I think geometry should not be scaled so much. GK104 and above should render nearly the same geometry like GK106/107 and invest the higher performance in resolution/anti-aliasing/better lighting.
The tiny Cape Verde has a tessellation performance like GF114:
http://www.hardware.fr/articles/855-8/performances-theoriques-geometrie.html
http://www.tweakpc.de/hardware/test...on_hd_7770/benchmarks.php?benchmark=heaven2ex
This would be far above a 1 GPC GK107 and on a level with a 2 GPC GK106, if they follow the Fermi geometry/tessellation performance scaling.
Or could be Cape Verde just be overpowered in this terms, especial if you compare it with Tahitis performance, which could be driver problem or broken design?
But on the other side, I think geometry should not be scaled so much. GK104 and above should render nearly the same geometry like GK106/107 and invest the higher performance in resolution/anti-aliasing/better lighting.
The GTX460 in that comparison is based on GF104 that have much lower base clock and hence lower primitive rate than GF114. Additionally, GF104 omits one of eight multiprocessors -- geometry performance in Fermi scales also with the # of multiprocessors (culling, clipping, tessellation, etc.).
chavvdarrr
Veteran
GCN 2.0 ?
That should come end of 2012.
That should come end of 2012.
therefore you believe that this slide is just a sort of "strategic statement"?
I would say it's just a fake.
Silent_Buddha
Legend
Wrong guy, Ninelven was making the claim that a HPC only part should just remove the graphics pipeline. I'm arguing that including the graphics pipeline is much more cost effective in a part that is tailor made for the HPC market yet still viable for the professional market with some appeal to enthusiasts in the consumer market.
The same hypothetical idea that Carsten proposed.
Regards,
SB
GK107 GT640M (unknown specs) tested: http://translate.google.com/transla...er+timeline+ultra&cd=7&hl=zh-TW&ct=clnk&gl=hk
Comparable game benchmarks with GT 630M (GF108 96SPs/128-Bit @ 675MHz and 1.6Gbps DDR3): http://translate.googleusercontent....-630m/&usg=ALkJrhhmL_M2Rb5zhatsr4E4PUM50QeHhA
GT 640M seems to be arround ~50% faster.
Comparable game benchmarks with GT 630M (GF108 96SPs/128-Bit @ 675MHz and 1.6Gbps DDR3): http://translate.googleusercontent....-630m/&usg=ALkJrhhmL_M2Rb5zhatsr4E4PUM50QeHhA
GT 640M seems to be arround ~50% faster.
Could you possibly be bothered to actually use the link I provided? Because it is exactly about part of what you said in the post I was answering to. In short: doing away with separate texturing, using shaders to do that.
Then maybe I did not understand, sorry.
Man from Atlantis
Veteran
good find
IF we consider GT640M has same bandwidth with GT630M, GK104 may be roughly 85-106% faster than GTX560Ti(GF114) in a perfect world..
