Graphics Do you feel the Nvidia 9xx series in general has thus far been a bit of a let down?

We have been at the stage of lower than expected yields from silicon for the last 4 years, why do you think GPU makers haven't made the transition away from 28nm ?

It's because when they transitioned to 28nm they had lower than expected yields, it's why they are skipping 20nm because the yields were so low (lack of experience in making 20nm), TSMC has more experience with 16nm FinFET so rumors suggest that is going to provide better yields.

Yes, yes they were.
Maxwell is the architecture not the fabrication process, they have been using 28nm since 2012 when Kepler made the transition from the 40nm being used on the 600 series cards.

If rumors and past history are an accurate guide i would expect we will see the same transition from the 28nm Maxwell cores used in the 900 series cards to 16nm Maxwell cores.

It's taken them almost 5 years to get the yields up where they want them.

For someone who says they understand a fair bit about the chip manufacturing process you seem to have very little understanding about the business of GPU manufacturing, GPU manufacturers don't make their own chips, they are customers of the major fabricators.

If it's anti consumer then perhaps you need to look at the chain of supply, and who is the consumer.

 

I'm aware of that, and you're right, TSMC are really bad at making 20nm cores, and so unsuprisingly they've launched GM204 on a 28nm process node. My point is they've (TSMC) had 3 ~ years of experience with 28nm for Nv cores now, they are good at making it, it's evidenced by the fact Nvidia are able to clock GM204 as high as they are, with so little voltage. Given that it seems perfectly reasonable to expect that they'd still get perfectly usable yields right now if they made the larger version, as I said even if there were issues they could just of released a 5 cluster version.

You're also right that 16nm is going to provide better yields than 20nm would, but it's still going to be lower than the current 28nm process. It would surely make sense to have launched a higher power, higher performance maxwell 6 cluster design on the 28 nm process they're experienced with, then launch a smaller 2 or 4 cluster design on the 16nm process node when it's ready.

No, no they're not. That's an article from 2012, before the 600 series launched, which launched as 28nm cards by the way. The 480 and 580 were both 40nm and so the 600 series launch represented a new architecture and a new process node. Given that it's understandable that they launched with a 4 cluster design but this time they've got no excuse, they're using the same 28nm process node that TSMC have gotten very good at producing and they got to test the architecture with the 750Ti. There is no excuse not to launch the bigger version. As it is the GTX 980 has actually even gone backwards on transistor count (5.2B down from 7.1B)

You're right history would tell us that they would have the architecture for 2 series of cards, They'd launch a slower clocked, slightly cut down, full size GPU and then die shrink it for the second series, and fully enable it, giving more performance for the top end card. It's what happened with the 8800 GTX (128 cores, 90nm to 9800 (128 cores 65nm), the 280 (240 cores, 65nm) to 285 (240 cores, 55nm) and the 480 (480 cores, 40nm) to 580 (512 cores, 40nm)

What we've been given here is 780 TI (2880 cores 28nm) to 2048 cores (28nm) It's unprecedented to have a backwards step of architecture and the only reason it can still keep up with the 780 is the slightly improved architecture and clock speed. A new architecture should debut with at least a doubling of ALU's, that's what we've always seen. And I know I'm generalizing a bit here but surely you see the point?

What they've done here will allow them to drag maxwell out over 3 series of GPU's, They have this series, the Next series can be a 6 cluster Maxwell and a 3rd series can be 16nm Maxwell. For that reason they may even stretch it out over a fourth, with the 3rd being a 16nm 4 cluster and the 4th series being a 16nm 6 cluster.
So the result of launching the 980 as a 4 cluster design means that if the competition doesn't push them, we could be sat here with Maxwell for longer than we'd expect rather than getting Pascal sooner. That's bad for progress, and bad for the consumer.

I am perfectly well versed in chip manufacturing thank you very much, If I ever imply that Nvidia are "manufacturing chips" it is in so much as they do actually commission someone to make them, and they tell them which ones to make, rather than GPU designers like Imagination Technologies who literally design IP and never actually get their GPU's manufactured, their customers do.

 

You argue your points well, but for all but a tiny fraction of us enthusiasts, somewhere between one and three 970/980s will be absolutely massive cards for the next few years at least.

It's not great shock that Nvidia (or indeed any CPU/GPU manufacturer) would hold back some of their developments until they're needed, or only make them available to professional users. That's always been the way. They are here to make extraordinary amounts of money, not to give a couple of thousand extreme enthusiasts around the world a marginally better product - that's definitely not where the money is.

 

A larger version like the rumored Titan II ?
Sure they could release a non Titan'esc card with 5 or 6 clusters but that's not how you run a business that has to cope with varying yields, you use the seasons when you have high yields to weather the times when you don't.

You seem to have missed that i said "were", past tense.

The only reason the 600 series launched with the 680, and for that matter other Nvidia series GPU's is because those cards need lower stock levels so it's easier to get the needed amount from the fabrication process.

The 600 series wasn't entirely made up of a new architecture and a new process node, it was made up of new architecture but it was made on both old and new process nodes.

I do, but as has been mentioned we're at the stage that yields are becoming a problem so the gains from one generation has to fund the losses from the next.

In a perfect world or in the past better yields could be passed onto the consumer in the form of bigger or cheaper chips, nowadays a die shrink can lead to higher per chip price and companies have to smooth out those rises and falls, iirc TSMC now charges per wafer not as they have done in the past per usable chip.

From my perspective the exact opposite of what you describe would happen, if the competition was pushing them we would have Maxwell for even longer, you only have to look at how long AMD have been using old fab and old architecture because they need to earn as much profit as they can from what they have.

Like i said, you say you are versed in chip manufacturing but seem to have very little understanding of the business side of GPU manufacturing.

It's all well and good understanding how chips are manufactured, but unless you understand the business model of how to deal with varying yields you will struggle to understand the decisions made on how the wafer is divided up into certain configurations.