[Sweclockers] 20 nm Nvidia Maxwell delayed - to 2015

[Mondozei]

Sweclockers, who are always careful with leaks and who nailed the 750 Ti announcements, are now out today saying we have to wait until 2015 for Maxwell on 20 nm.

Source(Swedish).

This naturally raises the question: will AMD be true to their recent history and get to 20 nm this year, ahead of Nvidia?
Since they rely on GloFo for their fabs, I say unlikely, despite recent history.

 

[TreVader]

Sweclockers, who are always careful with leaks and who nailed the 750 Ti announcements, are now out today saying we have to wait until 2015 for Maxwell on 20 nm.

Source(Swedish).

This naturally raises the question: will AMD be true to their recent history and get to 20 nm this year, ahead of Nvidia?
Since they rely on GloFo for their fabs, I say unlikely, despite recent history.

Probably not, but its possible Glo Fo might end up actually being an asset here. Idk why people thought 20nm was going to be here soon anyway. They just released high end 28nm a couple months ago.

 

[Mondozei]

I'm Swedish so here's a better translation than Google Translate.


------------------------------------------------------------------

Those who wait for the 20 nm generation will have to wait a while further. According to information which Sweclockers has received, the Maxwell architecture will stay at 28 nm all year this year.

In the middle of February this year, Nvidia launched the first graphics processors with the architecture Maxwell. The new generation was initiated with the budget class cards Geforce GTX 750/750 Ti, which means that the older architecture Kepler still reigns in the higher performance classes.

Despite this, the new architecture will still use 28 nm.
Although TSMC is in production using 20 nm, it does not look like it will be applicable to high performance ciruits required for graphics processors(This means that we're looking at 28 nm for all classes, not just high end GPUs - Ed). This has affected Nvidia's future plans.

Taiwanese partners are now telling Sweclockers that it will take a while before Nvidia can introduce GPUs with the Maxwell architecture. In best case scenario(my italic) we may see a very selected few models, but more likely is the case that we'll have to wait until some time into 2015.(Note: not right away, but "some time into", meaning we may be looking at Q2 in 2015 -Ed)

A transition from 28 nm to 20 nm is viewed as required(my italic. So no "soft launch" then) for Nviida to use Maxwell in the higher performance classes. This is especially the case for the successor to the monster GK110, which is the foundation for 780, 780 Ti and Titan/Titan Black.

That 20 nm will be delayed can also explain why the first GPUs with Maxwell did not constitute a new series but are now forced to cohabitate with the current Kepler 700-series.

 

[NTMBK]

This lines up with what NVidia said all the way back in 2012:

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

Still, I'm surprised, I was certain that they would fill out the rest of the lineup with 28nm Maxwell to match the 750ti.

 

[Mondozei]

Clarification: what I meant with "soft launch" was basically a 28 nm launch this year but with new architecture as some had speculated.
The sources to Sweclockers say this is ruled out, as Maxwell requires 20 nm in the middle-to-high performance classes.
(Obvious exceptions are the very, very low-end like 750/750 Ti)

 

[Mondozei]

This lines up with what NVidia said all the way back in 2012:

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

Yeah, that was my first thought, too.

Interestingly enough, Sweclocker's sources weren't able to say why the delay happened, the most glaring omission. (I've double-checked the news article, nothing in there as to why the delay).

But I'm guessing you're right to point out the costs.

Everyone is struggling these days, even almighty Intel, although in their case its more complicated than just costs?

 

[TreVader]

This lines up with what NVidia said all the way back in 2012:

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

Still, I'm surprised, I was certain that they would fill out the rest of the lineup with 28nm Maxwell to match the 750ti.

That's not what nvidia said, look at where those yellow circles line up on the curves. They revised it, on the 28nm curve the yellow circle is not in the same place as on the 20nm curve. All they did was move the circle so it LOOKS like that's what they had always predicted.


On the 28nm curve it bottoms out right at 2012. On the 20nm curve it bottoms out at Q2 or Q3 2014 but they then circle 2015.

 

[wand3r3r]

With the titan(ic)-z (due to price) and r9 295 just being released I suspect 20nm is farther out then some uninformed posters like to claim/guess. I will not be surprised at all if we are looking a year out.

 

[witeken]

This naturally raises the question: will AMD be true to their recent history and get to 20 nm this year, ahead of Nvidia?
Since they rely on GloFo for their fabs, I say unlikely, despite recent history.

This has nothing to do with recent history, but rather the fact that TSMC's struggling with new nodes. Yields probably aren't great yet, which is certainly a bad thing for bigger dies. The business model of independent semiconductor foundries is breaking apart.

I wonder if Nvidia might switch to Intel. Intel won't release dGPUs, so it could be possible, although Intel's also competing in that market with their APUs/IGPs.

 

[rtsurfer]

No suprise here.
I already knew it.

Even if by some magic we get 20nm this year, it wont be anything faster than a 770.

 

[NTMBK]

That's not what nvidia said, look at where those yellow circles line up on the curves. They revised it, on the 28nm curve the yellow circle is not in the same place as on the 20nm curve. All they did was move the circle so it LOOKS like that's what they had always predicted.


On the 28nm curve it bottoms out right at 2012. On the 20nm curve it bottoms out at Q2 or Q3 2014 but they then circle 2015.

Are we looking at the same graph? The crossover doesn't happen until late Q4 2014...

 

[witeken]

Everyone is struggling these days, even almighty Intel, although in their case its more complicated than just costs?

Intel isn't struggling with new nodes. Maybe it it, but at this point I'd say the delay of 14nm was just an exception. Intel already stated that it plans to start volume production in Q4 2015, exactly 2 years after when 14nm was originally was supposed to be started.

What struggling really means is that new nodes aren't profitable enough, so you'll have to wait longer before you can launch a new node. TSMC's new nodes are sponsored by their older nodes. But Intel's whole business model is based upon their 2-year manufacturing lead. From what I've heard, TSMC's yields seem to be dwarfed by Intel's, nor does Intel have any financial problems (also look at the links my post above).

 

[Mondozei]

One last clarification.

The phrase that Sweclockers used was, in Swedish, "en bit in".
This is a clever phrase that has no immediate English translation because you need to know Swedish to know the context it is used in. I used "some time into" as the best proxy.

"En bit in" is usually used by Swedes when we are trying to downplay expectations of something happening fast. It isn't readily seen in a literal translation, but it is most often used that way. If you still end up doing something faster, you have plausible deniability.

Sweclockers are very careful with their language throughout their article, and since they used that phrase, they are conveying to you that they believe that you shouldn't get your hopes up to a flying start right out of the gates in 2015.

But, of course, if Nvidia/TSMC surprises to the upside, Sweclockers have not tied themselves down to any specific timeframe. (Just as intended).

 

[NTMBK]

This has nothing to do with recent history, but rather the fact that TSMC's struggling with new nodes. Yields probably aren't great yet, which is certainly a bad thing for bigger dies. The business model of independent semiconductor foundries is breaking apart.

I wonder if Nvidia might switch to Intel. Intel won't release dGPUs, so it could be possible, although Intel's also competing in that market with their APUs/IGPs.

Intel are competing with NVidia in the HPC market- Xeon Phi is competing against Tesla, and letting NVidia fab GPUs on their process would screw the Phi.

Anyway, it's not just the independent foundries who are struggling; Intel's fabs are seriously underutilized, and there's a reason why they've delayed equipping their new 14nm megafab. They need insanely massive volumes to fund the development of new nodes, and between the shrinking PC market and their limited success in mobile I don't see where that volume growth is going to come from.

EDIT: And don't forget that Intel has also delayed their rollout of 450mm wafer technology.

 

[witeken]

Intel are competing with NVidia in the HPC market- Xeon Phi is competing against Tesla, and letting NVidia fab GPUs on their process would screw the Phi.

Yes, so the chances are low indeed. I wonder why Intel has no plans for dGPUs. There will always be a market for those products, and with their 2-3 year advantages, they probably could make that quite profitable.

Anyway, it's not just the independent foundries who are struggling; Intel's fabs are seriously underutilized, and there's a reason why they've delayed equipping their new 14nm megafab. They need insanely massive volumes to fund the development of new nodes, and between the shrinking PC market and their limited success in mobile I don't see where that volume growth is going to come from.

EDIT: And don't forget that Intel has also delayed their rollout of 450mm wafer technology.

The new 14nm fab is delayed because of cost savings. Intel also denied your second claim.

 

[ShintaiDK]

Sweclockers realiable? Hah.

Lets see what happens in H2. Specially Q4.

 

[NTMBK]

Yes, so the chances are low indeed. I wonder why Intel has no plans for dGPUs. There will always be a market for those products, and with their 2-3 year advantages, they probably could make that quite profitable.

Their plan is to gradually kill it with improved iGPUs, as far as I can tell. But don't forget that the Xeon Phi started life as a discrete GPU called Larrabee.

The new 14nm fab is delayed because of cost savings. Intel also denied your second claim.

If the capacity is there to fill a massive new fab, then the costs of installing equipment would be paid off- that's how fabs work. The problem is that Intel's demand is nowhere near where they projected it would be, meaning that they would not be able to pay off those costs- and there's no point opening a massive new fab if your existing ones aren't full, and can keep up with demand.

 

[piesquared]

Yes, so the chances are low indeed. I wonder why Intel has no plans for dGPUs. There will always be a market for those products, and with their 2-3 year advantages, they probably could make that quite profitable.



The new 14nm fab is delayed because of cost savings. Intel also denied your second claim.

intel doesn't have the ip for dGPUs, nor the experience, drivers, developer relations, or just about anything that is required for a dGPU. It would be a no return money pit for them. They have no chance against AMD or even NV for that matter, i'm sure they realize that.

14nm has been delayed greatly, that much is known. What happened to tic toc? It vanished into thin air. 450mm has also been delayed, that much is also clear regardless of intel PR.

 

[witeken]

If the capacity is there to fill a massive new fab, then the costs of installing equipment would be paid off- that's how fabs work. The problem is that Intel's demand is nowhere near where they projected it would be, meaning that they would not be able to pay off those costs- and there's no point opening a massive new fab if your existing ones aren't full, and can keep up with demand.

You forget that Fab 42 is also 450mm ready, so it isn't a waste of money at all.

 

[NTMBK]

You forget that Fab 42 is also 450mm ready, so it isn't a waste of money at all.

They've built a massive facility which is sitting vacant because they don't have the demand to fill it. That's a waste of money.

 

[witeken]

intel doesn't have the ip for dGPUs, nor the experience, drivers, developer relations, or just about anything that is required for a dGPU. It would be a no return money pit for them. They have no chance against AMD or even NV for that matter, i'm sure they realize that.

Fair enough, they might not see a good return on invest, but Intel can certainly afford to become serious about dGPUs. A process lead of 2 years is especially helpful for GPUs.

14nm has been delayed greatly, that much is known. What happened to tic toc? It vanished into thin air. 450mm has also been delayed, that much is also clear regardless of intel PR.

14nm's been delayed by about 1Q as far as we know. They might have changed the regular release schedule to have 14nm products faster to market where they are needed most for Intel (Cherry Trail). I've already disproved your other statements in other comments.

 

[witeken]

They've built a massive facility which is sitting vacant because they don't have the demand to fill it. That's a waste of money.

Have you read the article I linked in my post?

 

[NTMBK]

Have you read the article I linked in my post?

Yes, and it changes nothing. The building may be used eventually, sure, but they planned to kit it out and use it for 14nm because they anticipated suitable demand. That demand never materialized, so they're mothballing the facility.

 

[janii]

1 April.

For ze clicks

 

[Grooveriding]

Not surprised. I don't think Glofo even comes into play for AMD. Aren't they only for CPUs ? I think AMD makes their high end GPUs all at TSMC just like nvidia.

We're still hearing new high end 28nm SKUs being announced right now. No surprise 20nm could be delayed till 2015. When has TSMC not been late on a new node ? Never AFAIR.