Do you think that Tesla based on Fermi will use this application based ECC checking? I reckon that it will support ECC for all applications without requiring program modification.
NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
Do you think that Tesla based on Fermi will use this application based ECC checking? I reckon that it will support ECC for all applications without requiring program modification.
If they cannot hit their expected power and/or clock targets, then performance compared to Cypress will not be where Nv would want it to. And far worse, if their presumably big fat die will not be within a healthy margin out of current-issue cypress's reach, then AMD might do to them, what they did with RV790 to (most of) GT200b.
They don't have the option of just throwing a sub-expectations performing part on the market again with no immediate successor being in the pipeline, assuming they cannot go much bigger, denser or higher clocked on the current 40nm process and it's options (of whom i am not aware).
If yield is a critical issue for them and their coarse grained redundancy (which builds upon improvin process maturity over time or on the availability of optical shrinks), then they're in it even deeper - at least from where I am sitting.
So they might need to do an A4 or, as Ailuros suggests, even a Bn, to improve on either or both fronts if A3 isn't up to scratch in terms of delivering their intended performance. With G80, they could afford to wait with a new stepping/model/respin in the form of the Ultra until just before R600 launch. This time, a mediocre product, which might be on the market about six to eight months after the competition (I'm talking about Cypress here), will cost them dearly.
Groo The Wanderer
Regular
Basically, yes. If they can't get the clocks up after 3 tries for metal layer tweaks, a fourth probably won't produce miracles. It could, but I doubt it will in practice.
I am betting on a serious revamp of the architecture before a 28nm part, either that or NV will just suck it down in the benchmarks while proclaiming a crushing lead in some really odd benchmark that they do well.
-Charlie
Groo The Wanderer
Regular
You don't think Intel would make a custom CPU (fused/specified, not silicon changes) for Google? You don't think they have special bins of Nehalems for HPC/dense server farm operations? You don't think they would make a special mobo for the same uses that have higher temp operations, more efficient components, and in general higher specced everything?
/me looks to his left and sees a HPC Nehalem rack server with all of that other than the specially fused Nehalems.
-Charlie
That certainly seems quite plausible. So basically what you're saying is that you would agree that that other guy's claim that most server farms typically rely on budget single socket desktop components is unlikely?
Still, even lower TDP, higher efficiency components and the like don't really do much to change the probability that any Nehalem based solution will still be a factor or so off when it comes to computational density and power efficiency compared to a Fermi based Tesla setup.
A4 or B1 doesn't make sense at all. When A3 is crap nvidia would concentrate on the refresh chip and release a "Benchmark version" of A3 with limited quantities.
GF100 was planed for November 2009, so the refresh should be planed for about summer already.
If Fermi is equal to Cypress it will cost almost the same.
If Fermi is faster than Cypress it will cost more.
GF100 was planed for November 2009, so the refresh should be planed for about summer already.
If Fermi is slower than Cypress it will cost less.
If Fermi is equal to Cypress it will cost almost the same.
If Fermi is faster than Cypress it will cost more.
Hopefully NV will do the smart move and price it the same making Cypres obsolete, at least until AMD lower prices. I think forcing them to do so earns a lot of mind share though.
But will they be able to do that? If GF100 launches in March, by then the HD 5870 could very well be priced around $300.
Have they never sold any GT200(b) card for under 300$ in the usa?
And there will be a GF100 card directly target HD5870. They did the same with the 8800GT (vs 3870) and the GTX260-192 (vs 4870). I don't see why they won't do it with GF100.
For yield, speed or power, an A4 won't do much. (I'm a bit on a mission to kill the idea in people's brain that metal spins have a lot of impact on those.
) A B spin is something else entirely, of course, but that would be a much longer term solution.
FrameBuffer
Banned
Do you still stand by your GT215 metal spin claim to reduce leakage?
No, they gave GT218/16/15 a spin just for the heck of it.
It's interesting looking at the clocks on the various nvidia 40nm parts(in rough release order):
GT218 589/1402
GT216 625/1360
GT215 550/1340
GT215(OC) 585/1420
GF100 A1(Hotboy) 495/1100
GF100 A2(Various) ???/1200-1350
GF100 A3(Nvidia) ???/1250-1400
They seem very clustered around 1300-1400Mhz for shaders(excluding GF100 A1 obviously). Assuming they keep the same ratio as GF100 A1 core clock is looking around 550-630 maybe?
Contrast with the G9x generation which covered 1375->1836Mhz
GT218 589/1402
GT216 625/1360
GT215 550/1340
GT215(OC) 585/1420
GF100 A1(Hotboy) 495/1100
GF100 A2(Various) ???/1200-1350
GF100 A3(Nvidia) ???/1250-1400
They seem very clustered around 1300-1400Mhz for shaders(excluding GF100 A1 obviously). Assuming they keep the same ratio as GF100 A1 core clock is looking around 550-630 maybe?
Contrast with the G9x generation which covered 1375->1836Mhz
