NVIDIA Kepler speculation thread

The 40nm process is of exceptional quality right now, without monkey business. Yes, I am sure about that. As are tons of companies who are have rolled out their 40nm chips.

As for the exact timing about when that happened: I don't know first hand, but see various press releases and financial filings put it at end of Q4 2009/Q1 2010. It was certainly not much later.

There has never been an initially broken process that wasn't fixed eventually. 40nm was no different, it just took a little longer.

OK, thanks.
 
Isn't it at about 550 sqmm?
It's not reticle limited - it's cost limited. It certainly couldn't have been 800mm² or something crazy like that, but it could have been a fair bit bigger without hitting the reticle limit.

This whole reticle limit thing started when David Kanter asked John Nickolls about it for GT200 at the Tesla Editor Day. I was in the same room (along with Damien Triolet and iirc Theo Valich - I'm not kidding) and he basically didn't deny it, so David reasonably interpreted it as meaning it was indeed reticle limited. It wasn't - so presumably we just misinterpreted it and frankly John wasn't really the right person to know that anyway. Nobody, no matter how smart they are, can know everything for a company the size of NVIDIA or a field as extensive as 3D Graphics. And he was definitely very nice in person, RIP :( (see RecessionCone's post below, here's an article link)

But let's not forget the other side of the case: there is no question that TSMC fixed the issues completely early 2010 and 40nm became quickly a very stable high yielding process without any via doubling monkey business.
Ding ding. Thank you! I'm tired of people forgetting about that - TSMC clearly said so at the time, but people didn't believe them, so it apparently slipped out of everyone's mind. It happened in January 2010 according to one of my sources (not sure if that's what they applied the fix or when they got the first fixed wafers back - presumably the latter).

The 40nm process is of exceptional quality right now, without monkey business.
Exactly - and the best proof of that is its massive commercial success. TSMC 40nm is now one of the most successful processes in the entire history of the foundry industry. It has become TSMC's fastest process to surpass its predecessor in volume - and in fact, their fastest process to become their largest revenue contributor!

rpg.314 said:
People's focus on yields. They see considerable disparity and latch onto it.
NVIDIA probably had fairly low margins on GT21x before that but it is noteworthy that they shipped in high volumes as early as Q2/Q3 2009 when AMD was producing a lot less RV740s (if any at all for all I know). So the idea they had 40nm-specific problems on those that AMD did not is rather absurd. Yes, GT21x probably yielded worse than RV8xx before TSMC fixed the process, but it didn't turn out as badly as some people think it did.

However there is one more thing: NVIDIA did have one 40nm-specific problem with GF100 and the metal-heavy 'fabric' around the chip (as explained in that golem.de article). They were more than 75% responsible for that problem and admitted as such - while the tools did not properly simulate the fabric, NVIDIA should have known they couldn't be sure it would work, and they're the ones who took the risk anyway. What they should have done, given the low maturity of the tools at the time, is either: 1) do something more conservative, or better: 2) make a test chip for that specific aspect to make sure it'd work in advance.

In the end, they fixed the fabric, and at about the same time TSMC fixed the process. GF100's mass production yields and margins were fine. It was behind schedule which hurt them somewhat but they apparently survived. And ironically if didn't have the fabric problem they'd have mass produced before TSMC fixed the process and would have had much lower initial yields so they'd have been screwed either way. Then later they made incremental improvements with GF11x which had relatively little to do with process problems. End of story.

(BTW, just to make the cynics happy given how little there is for them to like here otherwise: GF10x uses 2 transistor Vt levels whereas GF11x uses 3 transistor Vt levels. That will allow NVIDIA to reduce power consumption but it also increases wafer price slightly. It's still a good trade-off and AMD might actually be doing the same, but it's not entirely free)
 
Isn't it at about 550 sqmm?
A typical retical size is 26x33mm (104 x 132 mm mask size which gets reduced 1:4 by the projection to the wafer). Some kind of "effective" reticle limit for a single die would probably be 1" x 1" equaling about ~650 mm². So if nvidia plans to make even larger dies, they have to start going rectangular at some point. ;)
 
More next-gen GPU codenames revealed
There have been no shortage of code name / branding leaks for next-gen GPUs in the last few weeks. To add to the list, not surprisingly, GK104 is in the works. This chip will form the basis for N14E mobile Quadro GPUs.

We are also aware of GK107, which is already sampling, and will feature in N13P-LP, N13P-GS, N13P-GT and N13E-GE GeForce mobile SKUs and N14P-Q1 and N14P-Q3 mobile Quadros.

GK104 is expected to be the successor to GF114, serving as performance GPUs on desktops and flagship GPUs on mobile SKUs. On the mobile side, GK104 could power the successors to GTX 570M/580M, most likely codenamed under the N13E series. On the mobile workstation front, GK104 will feature in N14E-Q1, N14E-Q3 and N14E-Q5 mobile Quadro SKUs. These Quadro GPUs will play one segment higher than the GK107 based N14P Quadro.
+ some AMD codenames (I don't even bother to try and remember them) through the link.
 
"Tahiti Woody and Tahiti Buzz on workstations under the AMD FirePro brand"

I don't know if Toy Story codenames are the most appropriate choice for the pro market, but OK. :D
 
It was indeed! Too subtle for anyone to actually get it at the time, it seems. :smile:

While you're dropping hints, you wouldn't happen to have a launch date, would you? Or any comment about the rumors saying desktop chips won't be released until 2012?

I can say that the supposed launch date flying around now (Early December) is not for something AMD
 
Haven't heard of that before, and neither seems google - at least with a credible source instead of some random forum post, which indeed show up when googling "Fermi GF100 "double vias"".

If I remember correctly I read about it at AnandTech. Their focus group member asked NV about that and they said yes. Never mentioned the extent though.
 
JHH just confirmed that 28nm Kepler products are currently in production during the Q2 conference call.

He didn't make it clear whether they were mobile or desktop parts.
 
JHH just confirmed that 28nm Kepler products are currently in production during the Q2 conference call.

He didn't make it clear whether they were mobile or desktop parts.
It doesn't seem like he confirmed that. Also, I assume you mean the Q3 call.

Jen-Hsun Huang; said:
And then there's a portion of it that is related to 28-nanometer tape out of new processors that will go into production shortly.
http://seekingalpha.com/article/307194-nvidia-s-ceo-discusses-q3-2012-results-earnings-call-transcript?part=qanda

I take this to mean production of shipping parts has not commenced, but they have parts in house.
 
He also said they got more design wins for the Ivy Bridge generation than any other before, due in part to the perf/watt of the Kepler design. That must mean they have validation cards already running in near full spec in OEM's hands right? He seemed very confident in Kepler, but we all know he can exaggerate.
 
From 4Gamer:

http://translate.google.com/transla...www.4gamer.net/games/120/G012093/20111124085/ (original: http://www.4gamer.net/games/120/G012093/20111124085/)

005.jpg


with interesting tidbits on TDPs and FLOPS (Google Translated):

OEM official said, "GK104 power exceeds 250W" also points out. In fact, NVIDIA is in a developing stage of PCI Express Gen.3, and has sought to incorporate into standard PCI SIG power 400W.
Furthermore, according to the specifications that are currently exposed to the OEM relationship, the amount of graphics memory or 3GB GK104 is 1.5GB, the memory interface (the same GF110 and GF100) 384bit. Single-precision floating point performance is expected to greatly exceed 2TFLOPS.
GK110 GK104 is equipped with two groups, but the clock kept to the problem of low power consumption, processing performance is still coming over the 4TFLOPS. And they mentioned his name GK112 GK102 Furagushippuchipputaru or earlier, is expected to be adopted by 512bit memory interface.
 
What is Furagushippuchipputaru ? My first guess would be some kind of Pokemon, but that doesn't really fit in the context.
 
Back
Top