Exclusive: TSMC raises prices on AMD and Nvidia
Apple and low yields to blame
Sep 8, 2011 in Channel, Chips, Finance, Graphics, Microprocessors, Opinion, Rumors
TSMC is doing some thing that are sure to make AMD (NYSE:AMD) and Nvidia (NASDAQ:NVDA) grumble, and grumble a lot. Moles tell SemiAccurate that they are raising prices on 28nm wafers.
Since we exclusively reported that Apple had taped out CPUs at TSMC, there has been the looming question of capacity. Apple uses a lot of chips, and they are not exactly small chips either. TSMC has limited capacity on their cutting edge processes, and if you depend on that, things can be a little problematic at times. Ask AMD about the 5xxx shortages when 40nm had hiccups.
Apple wants lots of chips, and because of this, they want lots of cutting edge wafers. If you have ever seen an Apple supplier contract, you know that they demand first priority on everything, and are pretty ruthless negotiating every detail. This is the long way of saying everyone else but Apple will lose out when there is a shortage.
One more thing to add, moles, ehem sources, say that TSMC’s yields on 28nm are not all that great. While we don’t know if the contracts with Apple specify a minimum number of wafers or minimum good die, either way, the initial guesstimates are unlikely to match the parts coming out. This means Apple will want, nay need, more wafers. What Apple wants, Apple gets. When Apple gets, everyone else looses.
That brings us back to AMD and Nvidia. They seem to want 28nm wafers as well, and in high volumes. They may not need quite as many as Big White, but they need a lot, and it is unlikely that TSMC has that many to offer. To lessen demand, TSMC came up with a cunning plan, raise prices to make the wafers less attractive. Either that or just rake in more cash.
In the end, we hear that prices for 28nm products just went up by somewhere between 15-25%, lets call it low 20′s. If 28nm wafers cost about $5000, this is around a $1000+ bump. You can probably understand why this would make GPU makers very unhappy, especially if you look at the 28nm lineups we exclusively published a few weeks ago (Nvidia here, AMD here).
The short story is that Nvidia is looking to capitalize on 28nm as a cost down measure for laptop chips, AMD less so. Given the margins on low end products, one has to wonder if this move just made initial profits on the new line just go *POOF*. Either way, you can bet both companies are not going to be happy. One has to wonder if this is why AMD is trying to move later SI/HD7000 production back to GloFo. Happy times continue.S|A
Updated: typo corrected 4xxx series to 5xxx series.
73 Responses to “Exclusive: TSMC raises prices on AMD and Nvidia”
Trackbacks/Pingbacks
-
-
Sep 20, 2011
[...] però potrebbe essere ostacolata anche dalla “forza contrattuale” di Apple. Secondo il sito SemiAccurate, TSMC avrebbe aumentato il costo di wafer a 28 nanometri fino al 25% per ridurre i volumi di GPU [...]
-
-
Sep 15, 2011
[...] информации, которой обладают авторы SemiAccurate, крупнейший в мире контрактный [...]
-
-
Sep 12, 2011
[...] Wolfgang Andermahr Quelle: SemiAccurate « Vorige News | Nächste News [...]
-
-
Sep 9, 2011
[...] to unnamed sources cited by Charlie Demerjian of SemiAccurate, TSMC has taken this preference given they see a lot of direct for their 28nm [...]
-
-
Sep 9, 2011
[...] to unnamed sources cited by Charlie Demerjian of SemiAccurate, TSMC has taken this decision since they see a lot of demand for their 28nm [...]
-
-
Sep 9, 2011
[...] to unnamed sources cited by Charlie Demerjian of SemiAccurate, TSMC has taken this decision since they see a lot of demand for their 28nm [...]
-
-
Sep 9, 2011
[...] AMD NVDA TSM AAPL – SemiAccurate reports TSMC is raising prices by 15-25% on 28nm products. Link [...]
-
-
Sep 8, 2011
[...] [...]
Does this mean that A6 will be ready before kepler in bulk? Or is this a whole different animal?
And if prices go up doesent apple pay the premium as well? But apple can afford it they’ll just raise prices on their toys while Nvidia and AMD less so I guess.
“TSMC’s yields on 28nm are not all that great” TSMC’s yields seem to be always shit. What a crappy company. They can’t get anything done right. I hope GloFo finally squashes that bug.
I don’t know who I hate more apple or tsmc.
Well there you go it was sure that Apple sueing Samsung will backfire on them sooner or later!
An analysis of Apple’s past purchasing history will show that they buy huge quantities of materiel and manufacturing capacity and want priority access… to the detriment of other parties.
I saw it coming when it was first announced when Apple went to TSMC but figured that AMD is smart enough to recognize the pattern and plan for it.
Shocking news about the price increase is if both AMD and Nvidia were taken by surprise. I think not, however.
Think about it. Every process shrink costs the foundry a huge amount of dollars. It must be recovered plus a profit. Customers wanting this new shrink will pay for it. This is an old pattern anyway and foundry customers are aware of it.
What the article alludes is the ability of both AMD and Nvidia to compete in the 28nm market, with this “new” development. AMD, I think, has a leg up with GloFo and so is probably insulated.
What have we learned today? “Woe for the old customer when Apple becomes the new customer.”
Was counting down the days till some moron tried to connect Apple’s move to TSMC to process issues suffered by Nvidia and AMD. Tech journalism at its very worst, but sadly 99.9% of tech journalism has less ethics than print tabloid journalism.
The only Apple story is that Apple gets the lowest prices wherever it takes its business, and that Apple’s profits on its hardware are vastly higher than almost any of you may guess. Agreeing to build for Apple is a VERY risky business (just ask Samsung), but the rules of capitalism means that no-one can afford to turn down their biggest potential customer.
The real problem with Apple at TSMC is that TSMC has fixing AMD and Nvidia chip problems as an even lower priority than before. Apple gets the service, and everyone else gets the shaft, due to finite on-site engineer resources.
dave id like to show u picture.
dave are u read.y. its about paul.
http://4.bp.blogspot.com/-CSvguIaleBs/Tmm-a4vXaaI/AAAAAAAAAVU/lJY4zaBDgzg/s1600/DSCI0211.JPG
paul real name is patatoe, as u can see. dave are u listing. patatoes are cheaper. ‘ole patatoe read, look at that mix, golden chipper if ever saw. dave are u gernam, too.
Whoosh….
vondrashek HAL8000….
Every die shrink wafer costs more since more dies can be produced.
That is why many still use 90nm and 60nm.
So the question is if this is the regular price increase for a new technique or if TSCM raised prices because they can.
TSMC have today no competition in 28nm. They have of course not shipped one single chip today.
Like explanation, sounds credible. heres incredible story just came out. Intel new 2011 pin enthusist cpu is coming in Nov ’11. Called Z79 Chipset, one problem, if anouncement is correct & graphics seem to done by:Intel. Z79 with 2011 Pin Has 8 Pci-e2 lanes.
Cough, Could be prostrate problem, yet kept selling way before system actually works properly. Can anyone imagine $1,000 for cpu ,that is best, yet only 8 pci-e2 lanes. Needs further investigation, Pray for misprint, as might just be worst Defeat for numero uno, yet. at least BIG Hold on imediate buying.
drashek Laughable….
Pcie is only a part of the puzzle. How many PCIe lanes does a server chip need? CPU power-efficiency is king and queen and the whole court.
“loses” not “looses”
Wait, wasn’t it just a month ago that TSMC was saying that they were delaying their ramp because of a lack of demand? “TSMC said in a conference call that its 28nm ramp had experienced delays due to customers deferring migration of chips to the next node due to the global economic conditions, rather than due to any fabrication issues.” Something doesn’t jive here.
I am probably missing something here, but with Apple going to TSMC, and Samsung recently building a new fab (or nearly completed with it), does that open the door to allow someone else fab their chips with Samsung? I doubt Samsung would use all that capacity themselves (it was essentially built for Apple, correct?), so maybe someone can move some of their chips over there (not sure what process they are on…).
Maybe that “Secret AMD/Samsung Joint Project” was really a hint at something to come.
Well I heard that Samsung also provide some 28nm LP process isn’t it possible for NVIDIA to
move it’s Tegra line there ?
Beat me to the punch, but that was the kind of thing I was wondering.
Complete crap. Wafer prices are negotiated well in advance. There is no last minute price increases. nVidia pays what they signed up for months ago. AMD is moving projects over to TSMC which may get price bumped but not the projects that are already there.
Altera and Xilinx are already shipping TSMC 28nm. No price hikes there either.
Charlie = douchebag
Dude did you read his conjecture? LATER SI/7000 production. That probably wouldn’t have been negotiated yet. So anything not already set in stone is going up, he didn’t say anything about current deals.
However that does F’up the long term plans of both companies. Figure it out.
Charlie’s no saint, but calling him a douchebag for pointing out the obvious consequences and the why for it, makes PorcheDan == douchebag && Charlie != douchebag.
LTFR
lawls, good sir.
Your honest criticism is highly appreciated. Hope to read more of it.
You want to hear more from TSMC’s paid social media guy? rrrarwl. TSMC flies him out every quarter. Pays him to troll. Reminds me of “fair and balanced” but no science included.
Meanwhile you troll for free.
Glofo can’t get good fast enough, normally on a new process at TSMC its mostly AMD/Nvidia fighting it out for wafers, add a 3rd big player for very early production and you go from 1/2 capacity to 1/3, which isn’t insignificant. If GloFo can get anything up and running, it would actually be best for AMD to keep running every wafer they can at TSMC, while having essentially exclusive access to Glofo(if they can use all the capacity), because they’d increase market share simply by having more capacity available to them than Nvidia. You can’t buy Nvidia, if they can’t make enough.
Short term, not great, longer term, could work out quite well for AMD. Leave Nvidia fighting for capacity at TSMC with Apple on just about all their products, while AMD buy out glofo capacity. Likewise with Apple going big at TSMC, capacity for Tegra can be hurt aswell….. is that one of the reasons they went for a quad core 40nm Kal-el, rather than move to 28nm dual with the same speed as soon as possible….. because they simply wouldn’t have had any capacity to produce them?
nVidia can go and manufacture at GloFo too. Intel might finally open it’s foundries…
Chipset business is long gone, now the only thing with some margins that Nvidia does now is graphics. If they go to Intel, i’m certain reverse engineering to favour Intel’s graphics development will be facilitated by greater degree than Nvidia would like.
Yeah right, but thats is only possible if GloFo can get its stuff in order.
Although i guess that if GloFo solves is 32nm problems 28nm will be no prob since its only an optical shrink.
Lol, you have no idea what’s going on.
GloFo 32nm SOI is a completely different process from GloFo 28nm bulk. No “optical shrink” here; it’s a complete re-design.
according to charlie (yeah, I know), solving 32nm will solve almost all 28nm problems…
can’t be that different if that’s the case.
Compared to who ?
I am pretty sure TSMC would KILL for Glofo yields on 32nm GATE FIRST.
Compared to TSMC 28LHP GLOFO’s 32nm SOI HKMG strained is like comparing 64nm to 40nm.
It is just night and day.
Oh, don’t forget Glofo shipping in vollume for 2nd quarter while TSMC just getting first runs out of the fab …
err 65nm to 40nm
If Apple quits Samsung, won’t Nvidia just go to Samsung? I think Nvidia has pretty good relationship with Samsung now. Samsung uses Tegra 2 in some of their devices, and they will use Kal-el,too, maybe even for GS3.
Nvidia and AMD have never used Samsung for any fabbing, and Samsung has zero experience with discrete graphics products. Nvidia might switch to Samsung in the long-term, but Samsung uses different process technology than TSMC as Samsung is in the IBM foundry consortium (as is GloFo). Even a long-term switch seems unlikely, as TSMC will likely try to open/expand fabs to increase capacity for Apple. It’s in TSMC’s best interest to make as many wafers as possible, as long as the demand exists.
nVidia did used another foundry in the past, IBM. The Geforce 6800 Ultra used the IBM 130nm bulk manufacturing process, Im not sure which other GPU’s used IBM though.
If AMD has returned to their small die strategy wouldn’t this impact them a lot less than a larger die Nvidia strategy?
Nope, it’ll still make their wafers, thus chips, 20% more expensive. That they can get more chips from 1 wafer is nice, but that just means they need less wafers than Big Green.
Besides, they may have 15% smaller dies or so, but they still need LOTS of chips to supply their ever-hungry markets.
15%? Where did you get that number? At least on the high end at 40nm AMD’s Cayman at 389 is more like 26% smaller than the 580. Heck the 560 is only 10% smaller than Cayman.The 6800 series is even more dramatic. If 28nm follows suit AMD is still in a better position as they can produce more chips for the increased cost versus Nvidia (assuming the die size strategy holds true at 28nm).
Considering that yields are already low and we are talking APPLE now wants a share, there’s only so much the small-die strategy will save AMD. Low yields, less chips, higher cost per wafer bodes badly for us AMD and Nvidia fanboys. I’m sure there’s a LOT of interest from AMD and Nvidia in GloFo now. I’d bet they would send engineers to help get 28nm yields up if they could at this point :)
Gee, i sure hope so. These friggin manufacturers are the biggest hurdle for progress in GPU’s and CPU’s.
You can have everything done and ready, but the chip-manufacturers just ain’t that effective. They’ve just become more and more expensive, which, in the end, hurts the customer.
Dumbass. How easy do you think it is to make 28nm transistors??
What hurts the customer is design engineers not understanding semiconductor processes well enough to design circuits on them.
I’m sure 28nm is hard. But before you go blaming this on the design engineers, remember that they get their design rules from the fab. When the fab is constantly changing the rules, it’s a tough target to hit. And “recommended” design for manufacturing rules can be even worse since it’s not always clearly WHICH recommended rules actually improve yield. Especially since the fab guys probably don’t know either. With a new process, it just the blind leading the blind. Eventually the fab has enough data to give accurate guidance, but that’s usually after you’ve taped out your bleeding edge design.
Good point. I was unnecessarily harsh…
I was coming from the 40nm GPU battle between AMD and NVidia, where AMD managed to get working silicon out the door, while NVidia was blaming TSMC.
The comment before mine sounded like somebody blaming the foundry because they didn’t know how to do the design well (but since other guy managed to design working silicon, maybe manufacturing wasn’t at fault)
AMD had fabs, and sold them as I recall.
“What hurts the customer is design engineers not understanding semiconductor processes well enough to design circuits on them.”
If that was the case, then there wouldn’t be any working silicon.
Do you see working 28nm silicon anywhere?
Don’t forget 7xxx isn’t going to have the same, space efficient VLIW architecture as previous generations. But as it can’t really be too much worse, I guess you’re right.
Fortunately for AMD, SI chips are probably going re-take the high end market, dethroning Fermi and Nvidia until kepler comes around. “Enthusiast” GPUs have a huge pricetag, and hence, high margins to make up for low volume. Bumped up production costs are less of an issue.
The mobile and low end GPU market, on the other hand, is a cut-throat industry. Production costs are EVERYTHING, as margins are small and volume is high.
Well, well. So the HD7k is moving to where the Bulldozer is manufactured? Damn. I dont know better, but I would call it ‘crysis’ for AMD.
And: What about the Korona launch? Are there any new facts about it? Will the BD V2 support PCI-E 3.0? And will the Radeon 7k support PCI-E 3.0? And what about GCN? Charlie! Can you please clear the fog?
Pci-e3.0 has hit quirk, Both msi 7 gigabyte make pci-e3,0 mainboards, yet two manufacturers pci-e3.0 are incompatible.
msi claims gigabyte uses incorrect controllers., over 40 mainboards from each manufacturer are tainted or mismade.
later pci-e3.9 has not had much news as result nor product to fill mainboards. got to be story in there somewhere.
drashek pci-e oppsee’….
No.
32 SOI and 28 bulk are separate fabs/tools.
“Ask AMD about the 48xx shortages when 40nm had hiccups.” Are you sure you don’t mean 58xx?
Wafer isn’t just size, where bigger holds more. Wafer is also series ofimages, not always same, yet ,each is imbeded one at time. so maybe get X% more for smaller nm, yet that takes X% more time to complete expousure of wafer, thus more cost, especially with high failure rate. Thus 40% longer than already proven process.
drashek md pile=Data Bank….
Your reply is completely unrelated to my comment. Good job mate, keep it up.
Don’t worry, what thomasxstewart says is never really related to anything.
besides 2mucruaoo replyed2 below, Present state has two cores run by 1 floating point & second branch prediction unit, so Floating point can split software up to know best routes thru L1 cache & specific software item. However, testers complain that one floating point unit lower actual number of tranee’ firing, which is standard score.
Well, by Thanksgiving, on Actual Game Tests, value of Bulldozer core will Be Better understood. How well do branch Prediction, & stoeage into work of L2.L3 as guide into two cores work, is it better & what is hand. who needs full yieds when engineering is bit on testbed.
mainbly results are:iLOVE Me.Drashek Dahologist….
Is this one a bot? Some sort of experiment at AI? People just discuss how off-topic his replies are and he goes on creating another one without even touching the observations?
Well, if this is an AI, somebody taught it grammar pretty well. Never seen an AI compose such sentences before. Can I get sources?
Well, Eye Nev’r. Thought Pricing & wafer size Vs. Yields. OK comment below on pci-e3.0 might help explain why 7XXX series are still pci-e2.1.
On 4xxx/5xxx rumble thru Jungle of 45 nm to 40 nm. about few days ago, mention’d lack of normal HKMG to LKMG then Mirror of die to lower nm process as HKMG then final as lkmg. That normal route did not take, as module took some time to develope. as need to add more sontrol thru out die, just stream core processors had problem of meaning to software. so slowly went back & forth, needing more time. up till 4xxx from 1xxx,2xxx,3xxx problems with compatibility with Vista 32 NT6 overwhelmed any idea of module. now both Vista as -=7=- 64 bit & module are in, if only pci-e3.o where standardized, to complete process.
drashek md
No PCIe 3.0 mobo yet so its useless at this moment. It’s like making DDR4 RAM while no mobo supports it.
Any 3.0 card will be backwards compatible so it’s really nothing like your comparison (you can still make us of the part, it’s not wasted).
That almost makes sense, keep it up….
Dude, i really love you.
Long Live Drashek!
Qick reply to merkulass, after -=7=- worked thru All hardware, especially expanding bandwith for larger footprint, problems with mlticore to reach 100% made TRY ANYTHING game, keeping working stuff & redoing, until, now, nearly completely 100% best 3 core module ever.
TS vondrashek RED’r.
The name is right there – in text – and in the “Leave a Reply to x” text above the reply box – and you STILL get it completely wrong.
This is why I use your text in my RNG. It’s about as good as any other entropy source.
Dude, you can’t expect anything but jibberish from drashek.
If it’s gibberish to you, maybe you’re not looking hard enough.
Ah AMD, how quickly you forget one of your old CEO’s famous sayings: “Real men have fabs…”
Jerry Sanders, where are you now and are you smiling?