[Sweclockers] 20 nm Nvidia Maxwell delayed - to 2015

[witeken]

But they would've needed such as fab anyway so building it a few year earlier than needed isn't that terrible, right? So I don't think the Fab 42 delay is a big deal at all.

“Building semiconductors is like playing Russian roulette. You put a gun to your head, pull the trigger, and find out four years later if you blew your brains out.”

 

[NTMBK]

But they would've needed such as fab anyway so building it a few year earlier than needed isn't that terrible, right? So I don't think the Fab 42 delay is a big deal at all.

“Building semiconductors is like playing Russian roulette. You put a gun to your head, pull the trigger, and find out four years later if you blew your brains out.”

In of itself it's not that big a deal (though I'm certain that the money spent on Fab 42 could have been used better elsewhere). The problem is what it indicates for Intel as a whole- that they are struggling to fill their fabs, and volumes are nowhere near where they predicted.

 

[tviceman]

Not at all surprising. But given GM107's power characteristics and perf/mm^2, Nvidia can launch 28nm GM106 and GM104 based chips to replace the 760, 770, 780, and 780ti if 20nm is ultra delayed. Will they do so? I doubt it. I think GM107 was pushed out asap on 28nm to keep AMD's APU's and Intel's ever-increasing (yet still crappy) IGP at bay in the mobile space (IMO mission accomplished). If 20nm truly is delayed until next year, even for AMD, I wonder what AMD's plans in the interim will be. I assume to stick with it's r200 series, but who knows.

I wonder if 20nm woes are going to affect TSMC's finfet rollout? It seems like TSMC's roadmap to 16nm (aka 20nm finfet) is scheduled to launch aggressively close to after 20nm.

 

[blackened23]

Well that sucks if true. But I agree with the sentiment that nvidia could do more Maxwell chips on 28nm considering the excellent performance per watt. At 60Watts they doubled the performance (looking at the prior 65W TDP amd and nvidia parts). Maybe they could release a 120W Maxwell part. Or 150. 175. Who knows. Now the truly high end Maxwell parts may be 20nm. But it'll be interesting to see if NV takes an apple type approach - staying on 28nm and using their engineering expertise and efficient design to create more chips. I'm sure they could do it if they wanted. IF the 20nm delay is true. Who knows.

Certainly hope the news isn't true though. I'd love to have new GPUs this late summer or fall to coincide with the apparent release of Haswell-E and X97/X99/etc. Should be an interesting summer/fall since we're about to get a ton of new products, including Haswell refresh, Haswell-E, and DDR4 memory. Normally GPU manufacturers are all too anxious to release GPUs along side new GPU releases. I guess we'll see.

 

[tviceman]

But it'll be interesting to see if NV takes an apple type approach - staying on 28nm and using their engineering expertise and efficient design to create more chips. I'm sure they could do it if they wanted. IF the 20nm delay is true. Who knows.

I don't think there is enough die space left to create a meaningfully new halo chip on 28nm. It took 25% more die area and 44% more transistors to get about 80% more performance(GM107 vs. GK107). If GM110 were to scale exactly like GM107, it'd have to be 688 mm^2, which is simply not possible.

I have no idea how much r&d goes into releasing a chip on one node, then optimizing/shrinking it down on a more sophisticated node later. If it's financially worth it, both Nvidia and AMD may do it for their mid-range dies (assuming they have the engineering man hours to dedicate as well).

 

[raghu78]

Sorry but I have to say that article doesn't look credible. This JP Morgan report says Apple 20nm SOC is at 50% yields and the 20nm process is in a much better condition than 28nm at the same time in its lifeycle

https://markets.jpmorgan.com/research/email/-kjegkq4/GPS-1336259-0

20 SOC brings full node like density increases but only half node like power/perf improvements. So designs will be power constrained. But it should not affect GPU design manufacturability. Clocks will have to be a bit lower to account for power limitations.

http://www.eda.org/edps/EDP2013/Papers/4-4 FINAL for Tom Quan.pdf

a 3072/3200 cuda core Maxwell GM104 running at 1 Ghz will easily be 40 - 45% faster than GK110. GM104 would also draw lesser power than GK110. say around 225w. a holiday 2014 release for GM104 is definitely on the cards.

 

[rtsurfer]

I don't think there is enough die space left to create a meaningfully new halo chip on 28nm. It took 25% more die area and 44% more transistors to get about 80% more performance(GM107 vs. GK107). If GM110 were to scale exactly like GM107, it'd have to be 688 mm^2, which is simply not possible.

I have no idea how much r&d goes into releasing a chip on one node, then optimizing/shrinking it down on a more sophisticated node later. If it's financially worth it, both Nvidia and AMD may do it for their mid-range dies (assuming they have the engineering man hours to dedicate as well).

I think both Nvidia & AMD would be better of dedicating their engineering resources to future GPUs & stacked memory stuff then wasting time to eek out more from the 28 nm node.

I am sure enthusiasts will be disappointed if there is no 20 nm high end GPUs this year.

But they would be more disappointed if what comes out on 20nm after all these years of wait doesn't offer a substantial performance increase.

 

[f1sherman]

I don't think there is enough die space left to create a meaningfully new halo chip on 28nm. It took 25% more die area and 44% more transistors to get about 80% more performance(GM107 vs. GK107). If GM110 were to scale exactly like GM107, it'd have to be 688 mm^2, which is simply not possible.

And 25% more die area is needed because anything bellow 80% is not meaningful?
You've done the math - I can easily see them staying on 28nm with +50% brought with Maxwell (skipping 20nm)

That +50% could keep them for 1 1/2 year. Or 20nm will indeed happen.

Too bad we are just guessing when all this has already been decided. Although there has been some roadmaps shuffling recently, so there could be some last minute changes.

As for R&D that goes into GPU design, I have no idea and would like to hear if anyone knows anything... ten(s) of $M?

 

[tviceman]

And 25% more die area is needed because anything bellow 80% is not meaningful?
You've done the math - I can easily see them staying on 28nm with +50% brought with Maxwell (skipping 20nm)

Of course not, but with what GM107 brings (80% over GK107), I don't think a Maxwell flagship die offering 50% more performance than GK110 leaves enough room for the additional GM106 and GM104 to exist simultaneously and offer enough differentiate without severely cannibalizing each other. Simultaneously, I think that leaves too much of a gap to cover with just a GM106 or GM104 - Nvidia would be forced to bin way too much, possibly using quite a bit of fully functioning dies with fused parts just to meet demand. Nvidia is all about margins these days, so I doubt they'd do that.

It's quite logical to predict that there will be flagship maxwell die on 28nm.

 

[wand3r3r]

Certainly hope the news isn't true though. I'd love to have new GPUs this late summer or fall to coincide with the apparent release of Haswell-E and X97/X99/etc. Should be an interesting summer/fall since we're about to get a ton of new products, including Haswell refresh, Haswell-E, and DDR4 memory. Normally GPU manufacturers are all too anxious to release GPUs along side new GPU releases. I guess we'll see.

It will be a huge missed opportunity if they can't launch around the same time. Who wouldn't want a shiny new 20nm card with the new enthusiast level parts coming. X99, DDR4, all the "new" technologies and then good ol' 28nm doesn't seem right.

 

[moonbogg]

Going to be a loooong wait.

 

[Keysplayr]

Doesn't matter what manufacturing process they're on as long as the performance is there. GTX 750Ti is a testament to what could be accomplished on the same 28nm node compared to 650Ti.

 

[AnandThenMan]

Doesn't matter what manufacturing process they're on as long as the performance is there. GTX 750Ti is a testament to what could be accomplished on the same 28nm node compared to 650Ti.

That's nice, but this thread is about 20nm Nvidia product delays. Which should be a surprise to no one, TSMC is always having issues going to a new node, even to the point where in 2009 AMD had to use all kinds of tricks just to get chips out the door.

 

[krumme]

Have you read the article I linked in my post?

As for 450mm delayed intel nowhere explicitly says its not. It is as the pr guy say constantly "aligned with partners" "for the second half of the decade". Everyone and his brother know its delayed and that also what is said of the analyst in the link. For good reasons. Intel is adapting capex eg reducing investments in eg hsml. No wonder. Find me someone outside of at forums and intel pr that says otherwise. Probably 99.999% of Intel employees will say its delayed. Its old news.

Gf tsmc intel is adapting.

 

[krumme]

In of itself it's not that big a deal (though I'm certain that the money spent on Fab 42 could have been used better elsewhere). The problem is what it indicates for Intel as a whole- that they are struggling to fill their fabs, and volumes are nowhere near where they predicted.

How much is the fab anyway? Is it finished?
And is it more than say 10% of the total cost with tools working?

 

[Mand]

How much is the fab anyway? Is it finished?
And is it more than say 10% of the total cost with tools working?

Last I checked a fab ran you in the range of two or three billion. Keep in mind that was a while ago, however.

 

[witeken]

How much is the fab anyway? Is it finished?
And is it more than say 10% of the total cost with tools working?

Yes, it's finished, ready to be used. The numbers are out there and I'm not going to verify it, but the building cost about 1.5B$. Equipment adds about ~4B$.

 

[krumme]

Aparently nv (tsmc) dont think we will pay for the performance 20nm gives. They have all the data and i am pretty sure they are right. If they can sell a gfx card for 3000usd on 28nm tech they got all segments covered. No need for expensive new nodes then.

 

[krumme]

Yes, it's finished, ready to be used. The numbers are out there and I'm not going to verify it, but the building cost about 1.5B$. Equipment adds about ~4B$.

The 1.5+4b is it the fab working then eg for 14nm atom whatever?

 

[Mand]

The 1.5+4b is it the fab working then eg for 14nm atom whatever?

Those numbers are upfront capital, to build the building and fill it with equipment. Operational costs are not included.

 

[krumme]

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.

I find it difficult to accept the graph. Perhaps its because it brings our gfx development to a near stop with 20nm forward. Gaaar. If this graph is true - in light of the saturation of pc market and probably in a handfull of years the mobile market - one can understand why intel and tsmc is pushing the brake. it looks like this area of technology development is slowing drastically. Damn.

 

[krumme]

Those numbers are upfront capital, to build the building and fill it with equipment. Operational costs are not included.

Thats what i thought. And how is operational cost defined?

I ask because as i remember wagely the process work to install and get the equipment to work was like 2-3 times as much as the equipment itself for Intel last i read?

When Intel is reducing investments in hsml its perhaps because all those extra cost is what matters? (And the market is not there - look eg at atom 22nm tablet sales - its sold at a massive loss)

 

[thilanliyan]

(And the market is not there - look eg at atom 22nm tablet sales - its sold at a massive loss)

Intel...selling at a loss? I didn't think they ever did that. :hmm:

 

[blackened23]

Going to be a loooong wait.

Maybe, maybe not. The way I see it, nvidia has an excellent architecture on their hands with Maxwell and while the truly halo part might need 20nm for sheer transistor count - it is also very possible that NV may scale Maxwell up a tad for more desktop GPUs.

Remember, NV's 60W TDP maxwell is twice the performance of the 65W Kepler or 65W Bonaire (I think, R7-250). So if they scale that up to maybe.....120W TDP, 150W TDP, they could still be well in the 28nm transistor window and have an appreciable performance gain over Kepler still.

I dunno. I just have a difficult time accept that NV will just go all year without releasing a new product especially when they have a capable new uarch on their hands; certainly 20nm is a constraining factor but that doesn't mean gains aren't still possible on 28nm. Now like I said, the true halo part may be 20nm if the delay IS true. We don't know, I sure hope not. I would like new products in tandem with the new intel platform, even if it arrives a few months or a quarter after. Haswell-E sounds pretty tempting, or even Haswell refresh if the overclocking rumors are true. In conjunction with a new GPU? Should be a cool full system upgrade. But I don't know, i'm just speculating here. I'm hoping it happens whether 28nm mid range Maxwell or 20nm high end. I want new toys, basically. But we'll see.

 

[Silverforce11]

Sorry but I have to say that article doesn't look credible. This JP Morgan report says Apple 20nm SOC is at 50% yields and the 20nm process is in a much better condition than 28nm at the same time in its lifeycle

Entirely a different process, LP 20nm on a TINY SOC die with 50% yields is HORRIBLE.

Saying its better isn't much because 28nm early on was atrocious.

TSMC is not stupid they know where the money is and its the vast bulk of mobile SoCs, so their priority must be in meeting the demands on 20nm low power first and foremost before expanding into high power production. This was noted a year ago, how mobile SoC makers Apple and QC pretty much bought out all TSMC production capacity leading to delays.. and here we are a year later, with info of such delays going public.