G92, that gen was different due nVidia living off rebrands, and the 21x can be explained by the fact that it was clearly "newer generation" than 20x, having DX10.1 support and all.
G92, that gen was different due nVidia living off rebrands, and the 21x can be explained by the fact that it was clearly "newer generation" than 20x, having DX10.1 support and all.
GT215 only supported DX10.1 over GT200. A pretty minor improvement.
GK110 supports dynamic parallelism and Hyper-Q over GK104. An improvement that's much more significant for compute people.
I should have used "refresh" rather than "new gen", but same holds for GK110 features, except that they were most likely in GK100 too before it got canned, they're just useless in gaming chip like GK104, and won't probably be on GK114 either.
R580 was only 352mm², and R480 was only 297mm². R360 was smaller than either of those, and R200 was smaller than R360.
AMD didn't cut any features from RV670 or RV770 due "small die strategy", GK104 sacrificed crapload compared to GF100/110, and nVidia is having another monolithic chip coming (and canned one) so there's no "small die strategy" for nVidia here. GK104 is just based on the same philosophy as GF104/114.
Linky-link pls?
Oh good luck getting anything for that. All that you get is a colourful mix and match of any possible desparate theory because things have to be at any price as they want it to be. A complete waste of bandwidth and time.
Blazkowicz
Legend
couldn't nvidia have designed GK10x and GK110 at the same time, before anything taped out?
I believe so, and don't see the need for a GK100 or the rationale behind it. is it a "bad" version of the GK110? did nvidia plan to waste time on a huge chip and replace it six monthes later? (and force $BIG_OIL_COMPANY to junk their brand new Tesla racks and re-certify everything if they want to have the 10% power improvement or whatever the difference would have been)
it seems to me that nvidia didn't choose between "small die" and "big die" strategy. their new strategy is to follow both strategies
.
I believe so, and don't see the need for a GK100 or the rationale behind it. is it a "bad" version of the GK110? did nvidia plan to waste time on a huge chip and replace it six monthes later? (and force $BIG_OIL_COMPANY to junk their brand new Tesla racks and re-certify everything if they want to have the 10% power improvement or whatever the difference would have been)
it seems to me that nvidia didn't choose between "small die" and "big die" strategy. their new strategy is to follow both strategies
.Considering that they do design architectures around ~3 generations ahead, yes.
Blazkowicz
Legend
that was a rhetorical question by the way, but sure. we can safely assume they have worked a lot on Maxwell and the ARM CPU already
It's all be gone through before, if nVidia didn't have GK100 project that got canned, GK110 would be called GK100.
I would like to see a timeline about when people think this GK100 chip was cancelled. The theory here seems to be that it was cancelled because it couldn't be done because it was too big for the process at the time.
For 28nm, there are no major hiccups like 40nm: it's pretty smooth sailing (capacity issues are not process hiccups.) This means that it's very predictable how good your yields are going to be for a particular die size without the need to have silicon in your hands.
Tahiti and GK104 hit the shelves with only ~2 months difference. This suggests that it was dictated by process readiness. It takes at least 4 months to go from first silicon to full production (more for new processes?), so in August 2011, 28nm must have been very immature yield wise.
The theory by a few that GK100 was cancelled due to low yielding silicon is just stupid: everything must have been low yield before August 2011. Do we all agree on that?
That means it must have been cancelled before there was silicon, based on theoretical numbers. At that point, you rely on your yield models. Since those were known long before, that decision might as well have been made 2 or 3 years ago. That long before tape-out, these things become marketing decisions, not engineering decision.
So, yeah, maybe it existed, on a slide somewhere, maybe not. But is it important in terms of things that matter (e.g. resources, forum blatter doesn't count) ? Nah...
For 28nm, there are no major hiccups like 40nm: it's pretty smooth sailing (capacity issues are not process hiccups.) This means that it's very predictable how good your yields are going to be for a particular die size without the need to have silicon in your hands.
Tahiti and GK104 hit the shelves with only ~2 months difference. This suggests that it was dictated by process readiness. It takes at least 4 months to go from first silicon to full production (more for new processes?), so in August 2011, 28nm must have been very immature yield wise.
The theory by a few that GK100 was cancelled due to low yielding silicon is just stupid: everything must have been low yield before August 2011. Do we all agree on that?
That means it must have been cancelled before there was silicon, based on theoretical numbers. At that point, you rely on your yield models. Since those were known long before, that decision might as well have been made 2 or 3 years ago. That long before tape-out, these things become marketing decisions, not engineering decision.
So, yeah, maybe it existed, on a slide somewhere, maybe not. But is it important in terms of things that matter (e.g. resources, forum blatter doesn't count) ? Nah...
Blazkowicz
Legend
let me play a bit with names :
as for a gk104 refresh I would say, maybe there's none. maybe there's one, far off, but it could be GK114, or a GK124 instead if there are new tweaks. we know nothing anyway.
I believe nvidia may use the GK104 till next generation. the successor would be GM104, as in Maxwell.
as for a gk104 refresh I would say, maybe there's none. maybe there's one, far off, but it could be GK114, or a GK124 instead if there are new tweaks
. we know nothing anyway.I believe nvidia may use the GK104 till next generation. the successor would be GM104, as in Maxwell.
Last edited by a moderator:
iMacmatician
Regular
How does that explain chips like GF117, GF119, and most of the GT21x lineup? Unless what you're saying applies just to the GXyz0 chips.
They changed their naming with GT2xx-series to G(eForce)X, in that case GeForceTesla200
21x-lineup was notable refresh over 20x-lineup with DX10.1 support and HD audio support.
GF117 and 119 were refreshes of GF108, they decided that it's worth it to make even smaller chip instead of disabling half of the SMs like they did with GF108, ending GF117 to laptops (same config as 108) and 119 on lowest end desktop (halved 108 except ROP/mem))
--
Of course if next gen top-end GPGPU monster will be x10 with no signs of x00, I take all my words back, but I think it's highly unlikely scenario.
I dont know there's many thing we will surely never know, untill maybe someone from Nvidia put a word about it in an interview in 2 years.
( If ofc, as many time with Nvidia, what he said he's just not marketing for the product they will do at this time )
GK110 is quite a different can of tuna then the current GK104/GK107. It's not just that it's bigger which some around here seem to think is everything. Copying SMX a few more times on the die is not that complicated. Making GPU kernels able to launch new GPU kernels is however quite an architectural advancement. And if you think that say a year ago Oak Ridge was sold on something that was essentially just a bigger GK104...
But since someone mentioned in this very thread about 200 pages ago that the next highend chip will be GK100 this means of course that it existed at some point in time right? There also never was a "G90" or was that also canned?
It's a useless line of discusion really.
But since someone mentioned in this very thread about 200 pages ago that the next highend chip will be GK100 this means of course that it existed at some point in time right? There also never was a "G90" or was that also canned?
It's a useless line of discusion really.
GK110 is quite a different can of tuna then the current GK104/GK107. It's not just that it's bigger which some around here seem to think is everything. Copying SMX a few more times on the die is not that complicated. Making GPU kernels able to launch new GPU kernels is however quite an architectural advancement. And if you think that say a year ago Oak Ridge was sold on something that was essentially just a bigger GK104...
But since someone mentioned in this very thread about 200 pages ago that the next highend chip will be GK100 this means of course that it existed at some point in time right? There also never was a "G90" or was that also canned?
It's a useless line of discusion really.
GF104/114 vs GF100/110 weren't quite just "paste more SMs and it's done" -cases
