Even if they reacted immediately (unlikely) and even if it's only 3 1/2 months order to store shelves (unlikely as well, the fab alone for regular lots is 3m+), then the question remains: where did they immediately get additional fab capacity?
The AMD Execution Thread [2007 - 2017]
- Thread starter overclocked_enthusiasm
- Start date
-
- Tags
- amd
- Status
overclocked_enthusiasm
Regular
http://wccftech.com/amd-carrizo-apu-excavator-cores-28nm-claims-leak-shift-glofo-spotted/
"An extremely important leak just arrived on our doorstep and (for the first time I think) the leak is provided by Extremetech.com which noticed a very important piece of evidence posted on SA forums. The leak more or less confirms that Excavator cores and the Carrizo APUs will be on the 28nm Process but as always I would recommend that pinch of salt. More interestingly however it also hints that part of the production will shift back to Global Foundries and away from TSMC.
~ Porting 28nm TSMC to 28m Glofo.
~ Porting 28nm Glofo to 20nm Glofo.
Very very interesting. So if this leak is correct then AMD is working on porting existing 28nm TSMC design to Glofo 28nm design. Now a couple of things could fall into this category. The most obvious choice would be Hawaii. Since we know part of the Excavator production of the Carrizo APU will take place in Glofo it only makes sense that the 28nm Hawaii GCN cores are made in the same fab design. And exactly is AMD doing with a 28nm-to-20nm port research. Are they going to pull a Maxwell with Hawaii ? or is it for the ARM sector which will naturally benefit immensely from a die shrink."
Getting back to my original fly in the ointment supposition that the 2014 WSA is perhaps THE driving force behind AMDs current fab decisions and may be impacting supply as a result. I think that AMD is being compelled to move their 28nm GPU business from TSMC to GF sooner than they wanted to because of the $190 million miss to the 2013 WSA and the $250 million 1Q 2014 obligation as part of the 2013 WSA. I am guessing that GF forgave this $190 million miss and as a result AMD agreed to much higher 2014 WSA numbers than they would have liked. I have no proof...this is my suggestion. This article merely supports my idea that AMD is at least considering moving GPU production from TSMC to GF. Perhaps this transition caught them with their Litecoin pants down and that is why they have been unable to respond appropriately to the supply issues.
In Q3 2013 AMD had inventory of $922 million and $884 million in Q4 (9 quarter average is $704 million). That is a TON of inventory and way way too high for AMD. With the PC market soft, especially at their traditionally strong high volume low end laptop segment, AMD can't satisfy their commitments to GF with just CPUs and some APUs. As John Byrne alluded to in December, some of the semi-custom console APUs will move to GF. As he also explained, AMD wants to keep "60% core ip consistent from gen to gen", so perhaps having these common parts in the same house is more cost-effective. AMD simply can't buy more CPUs from GF and MUST transition their wafer obligations to other parts or run the risk of hundreds of millions in "take or pay" fines like the $380 million they paid from defaulting on the 2012 WSA, or bloating their already massive inventory with more unwanted parts, risking write-downs.
We need more info:
~ 2014 WSA details
~ 28nm semi-custom moved to GF?
~ 28nm GPU moved to GF?
~ Who is likely to get to 20nm first...TSMC or GF?
~ Excavator and Carrizo rumored on 28nm so again why not GF?
The anecdotal evidence is stacking up that AMD may have been nailed to the wall by GF and forced to move a significant amount of their runs from TSMC to GF or risk hundreds of millions in fines they cannot afford. This may have served to exaggerate the Litecoin induced supply issues as AMD got caught in transition from TSMC to GF and perhaps couldn't secure unanticipated fab space at TSMC while waiting for GF production to be ready?
GF has leverage over AMD...the question is how hard are they squeezing and what does AMD have to do to end 2014 with a clean slate unlike 2012 and 2013? $190 million in missed obligations and $250 million in just Q1 obligations puts AMD on the hook for $440 million with 3 quarters left to negotiate. At a VERY conservative $250 million for the other 3 quarters that puts the 2014 WSA at $1.19 billion. That means AMD has to buy an additional $240 million in wafers compared to 2013 just to satisfy their obligations to GF. That is a TON of wafers and must come from new parts GF hasn't made for them in the past.
Again, I go back to AMD wants to fix the current supply issue (as any good business would want to) but can't because of this transition to GF and the issues I suggest.
"An extremely important leak just arrived on our doorstep and (for the first time I think) the leak is provided by Extremetech.com which noticed a very important piece of evidence posted on SA forums. The leak more or less confirms that Excavator cores and the Carrizo APUs will be on the 28nm Process but as always I would recommend that pinch of salt. More interestingly however it also hints that part of the production will shift back to Global Foundries and away from TSMC.
~ Porting 28nm TSMC to 28m Glofo.
~ Porting 28nm Glofo to 20nm Glofo.
Very very interesting. So if this leak is correct then AMD is working on porting existing 28nm TSMC design to Glofo 28nm design. Now a couple of things could fall into this category. The most obvious choice would be Hawaii. Since we know part of the Excavator production of the Carrizo APU will take place in Glofo it only makes sense that the 28nm Hawaii GCN cores are made in the same fab design. And exactly is AMD doing with a 28nm-to-20nm port research. Are they going to pull a Maxwell with Hawaii ? or is it for the ARM sector which will naturally benefit immensely from a die shrink."
Getting back to my original fly in the ointment supposition that the 2014 WSA is perhaps THE driving force behind AMDs current fab decisions and may be impacting supply as a result. I think that AMD is being compelled to move their 28nm GPU business from TSMC to GF sooner than they wanted to because of the $190 million miss to the 2013 WSA and the $250 million 1Q 2014 obligation as part of the 2013 WSA. I am guessing that GF forgave this $190 million miss and as a result AMD agreed to much higher 2014 WSA numbers than they would have liked. I have no proof...this is my suggestion. This article merely supports my idea that AMD is at least considering moving GPU production from TSMC to GF. Perhaps this transition caught them with their Litecoin pants down and that is why they have been unable to respond appropriately to the supply issues.
In Q3 2013 AMD had inventory of $922 million and $884 million in Q4 (9 quarter average is $704 million). That is a TON of inventory and way way too high for AMD. With the PC market soft, especially at their traditionally strong high volume low end laptop segment, AMD can't satisfy their commitments to GF with just CPUs and some APUs. As John Byrne alluded to in December, some of the semi-custom console APUs will move to GF. As he also explained, AMD wants to keep "60% core ip consistent from gen to gen", so perhaps having these common parts in the same house is more cost-effective. AMD simply can't buy more CPUs from GF and MUST transition their wafer obligations to other parts or run the risk of hundreds of millions in "take or pay" fines like the $380 million they paid from defaulting on the 2012 WSA, or bloating their already massive inventory with more unwanted parts, risking write-downs.
We need more info:
~ 2014 WSA details
~ 28nm semi-custom moved to GF?
~ 28nm GPU moved to GF?
~ Who is likely to get to 20nm first...TSMC or GF?
~ Excavator and Carrizo rumored on 28nm so again why not GF?
The anecdotal evidence is stacking up that AMD may have been nailed to the wall by GF and forced to move a significant amount of their runs from TSMC to GF or risk hundreds of millions in fines they cannot afford. This may have served to exaggerate the Litecoin induced supply issues as AMD got caught in transition from TSMC to GF and perhaps couldn't secure unanticipated fab space at TSMC while waiting for GF production to be ready?
GF has leverage over AMD...the question is how hard are they squeezing and what does AMD have to do to end 2014 with a clean slate unlike 2012 and 2013? $190 million in missed obligations and $250 million in just Q1 obligations puts AMD on the hook for $440 million with 3 quarters left to negotiate. At a VERY conservative $250 million for the other 3 quarters that puts the 2014 WSA at $1.19 billion. That means AMD has to buy an additional $240 million in wafers compared to 2013 just to satisfy their obligations to GF. That is a TON of wafers and must come from new parts GF hasn't made for them in the past.
Again, I go back to AMD wants to fix the current supply issue (as any good business would want to) but can't because of this transition to GF and the issues I suggest.
I expect that "simply" moving the console SoCs to GloFo would solve any WSA issues for the foreseeable future. Discrete GPUs could then be manufactured at either fab, depending on the quality and cost of their respective process offerings. Future semi-custom designs could also be made at GloFo.
I don't expect AMD to move any other existing designs (such as Hawaii) to GloFo: that would just be too costly and would only have potential benefits for a few months at best, after which Hawaii and its little GCN friends would be replaced by whatever is coming up next. Console chips should be manufactured for another couple of years (until they're shrunk to a smaller node) and at much higher volumes, so that's a different story.
I don't expect AMD to move any other existing designs (such as Hawaii) to GloFo: that would just be too costly and would only have potential benefits for a few months at best, after which Hawaii and its little GCN friends would be replaced by whatever is coming up next. Console chips should be manufactured for another couple of years (until they're shrunk to a smaller node) and at much higher volumes, so that's a different story.
The article laid the blame on two factors, miners and the new year. I think either the writing or the writer were muddled, but no direct quotes were provided to know what exactly whoever it was at AMD was saying.
The other thing that comes up here is that AMD as a GPU chip maker does not have the ability to conjure complete cards in MSI and Sapphire boxes onto shelves, which would mean that if AMD was going to massively ramp GPU production before the Litecoin peak, it would then need to somehow convince every party down the line to ramp as well for a reason they could provide no evidence for.
You have to order too much before knowing of a steep demand spike, one that appears to have measurably distorted the market. If you start reaching the limits of other parts of the pipeline, you incur higher costs and complexity if you need to bring in additional sources and your partners need to negotiate for an atypical product run. There are limits to how much inventory most companies these days are willing to stockpile, and in general many supply chains do not maintain significant inventory.
99 times out of a 100, you'd incur costs, write down inventory, and screw over partners.
AMD, as investors and its partners know, has probably experienced 99 of those times.
If you had the responsibility of planning product and managing the risk in millions of dollars of transactions and obligations, what are you going to tell everyone in August, September, or October that would convince them that a demand spike composed of GPU owners buying 4, 8, 16 units who don't play games in November?
You sell 20 bottles of a specific pale ale on a good day, and you receive bi-monthly shipments on a contract you negotiated several months back.
Apparently a new social media craze pops up on Thursday that involves using said ale in squirt gun fights because its bottle apparently plugs right into some new model.
How did you know to treble your order of that ale all the way back then?
Did it displace any popular brands in your back room?
How do you like your employees having to parkour over alcoholic squirt gun ammo to get to customers?
This is the case where they are not bringing in that kind of customer.
It's possible some number of buyers might start with an AMD GPU and just put more things in the AMD bucket. That sort of is the point of their trying to form a platform brand.
It's not a particularly rational thing when going into the upper performance range, and AMD has absolutely no halo in the CPU side of the equation. I would argue in some aspects that the GPUs are hurt by the association.
I think they may be in trouble anyway should GCN become economically inferior to other mining alternatives or the coins seriously devalue. All those cards are going to be dumped, and unless the next round of product is superior enough to Tahiti and Hawaii cards being dumped 4 to 8 at a time at cut-rate prices, AMD's still going to be hurt.
It would be nice if AMD could count on mining as a long-term use case that could be used to their advantage. Their compute line would have full mining capability, but their consumer cards would not, up until it was time to bring out a new round of cards.
Unlock the last-gen cards and have a non-gaming market hoover the inventory and prod reluctant upgraders to move up.
However, I do not see enough rhyme or reason to the GPU mining craze to rely on that. There's too many ways for mining in general to falter, or specific coins to falter, or specific architectures to become competitive or cease being so.
The opinion that a corporation not do things that have lack a rational basis is not the same as one requiring that AMD roll over and die.
However, they operate in an industry that works in a certain way, so stamping one's feet and demanding actions that require AMD be in a different industry, a master of its own, or the master of time and space is not a mark against them.
Just for kicks:
here's the Bitcoin hashing rate for the past year: http://blockchain.info/charts/hash-rate
I believe it was still profitable to do BTC GPU mining last May, or so, when the global hashing rate was at 80k GH/s. It's now at 24000k GH/s. Hashing rate went up by a factor 300 in 9 months. I don't know when BTC mining became unprofitable for GPUs, but it is obvious that the road from profitable to not profitable was extremely quick.
I see that the first scrypt ASIC miners are being announced at 5MH/s. That's a factor 5x better than an R290X, for a lower power consumption. scrypt will be harder to scale up than SHA, but I don't think there's any doubt that ASICs will soon take over.
So how large is the window really for AMD to ramp up and get a larger piece of the action?
There's ASICs promised for H2 this year. And if larger H2 Maxwells are anything like GTX750Ti in terms efficiency, it's a given that those will seriously hamper R290 popularity as well. Anybody with a bit of smarts will try to sell their R290's while they still can at a high enough price. August seems about to be the latest possible time to pull the trigger.
So we're talking a window of max 6 months, best case. A very risky play, if you ask me.
here's the Bitcoin hashing rate for the past year: http://blockchain.info/charts/hash-rate
I believe it was still profitable to do BTC GPU mining last May, or so, when the global hashing rate was at 80k GH/s. It's now at 24000k GH/s. Hashing rate went up by a factor 300 in 9 months. I don't know when BTC mining became unprofitable for GPUs, but it is obvious that the road from profitable to not profitable was extremely quick.
I see that the first scrypt ASIC miners are being announced at 5MH/s. That's a factor 5x better than an R290X, for a lower power consumption. scrypt will be harder to scale up than SHA, but I don't think there's any doubt that ASICs will soon take over.
So how large is the window really for AMD to ramp up and get a larger piece of the action?
There's ASICs promised for H2 this year. And if larger H2 Maxwells are anything like GTX750Ti in terms efficiency, it's a given that those will seriously hamper R290 popularity as well. Anybody with a bit of smarts will try to sell their R290's while they still can at a high enough price. August seems about to be the latest possible time to pull the trigger.
So we're talking a window of max 6 months, best case. A very risky play, if you ask me.
overclocked_enthusiasm
Regular
Who is likely to get to 20nm first...TSMC or GF?
TSMC is already mass producing chips at 20nm
overclocked_enthusiasm
Regular
Why did Nvidia use 28nm for Maxwell?
The used it for a mid/low range part because it's available now and cheaper, 20nm will be used on everything above it for the 800 series in several months.
While 20nm is technically available now, it's in limited quantities and comes at a premium.
That and I think it's still unclear whether TSMC will have 1 or more 20 nm processes, the current one is dubbed "20SOC", and it might not be optimal for GPUs
20nm is probably not ready yet for low redundancy designs. With the architecture clearly providing huge benefits even for 28nm, there's no reason to wait for 20nm to get ready. One can sell a lot of chips when the competition doesn't show up, as is expected to be the case for notebooks.
And also: contrary to the past, the cost per transistor for 20nm is not expected to be much improved over 28nm. So for a small chip that isn't in anyway die size limited, the better question for this Maxwell chip would have been: why bother with 20nm at all?
And for all we know, so is GlobalFoundries.
Some component lead times are longer than ASIC leads.
It's a good thing Bitcoin was already well beyond GPU mining, given the closure of Mt Gox and the devaluing resulting from the ensuing panic.
I suppose we'd figure out how much independent value scrypt-based coins would have, but they and Litecoin in particular tended to make their case as a complement to Bitcoin.
Even if this is not fatal to the whole endeavor, if the overall value ceiling these cryptocurrencies drops to more modest levels, it already cuts into the ROI for a GPU rig for this. Areas with higher energy costs were already making it less and less compelling, and a serious drop would be enough to hurt returns significantly, if not make it a loss in some places.
The prices of AMD's GPUs might be moderating somewhat sooner. Hopefully they didn't order a hundred thousand wafer starts this week...
I suppose we'd figure out how much independent value scrypt-based coins would have, but they and Litecoin in particular tended to make their case as a complement to Bitcoin.
Even if this is not fatal to the whole endeavor, if the overall value ceiling these cryptocurrencies drops to more modest levels, it already cuts into the ROI for a GPU rig for this. Areas with higher energy costs were already making it less and less compelling, and a serious drop would be enough to hurt returns significantly, if not make it a loss in some places.
The prices of AMD's GPUs might be moderating somewhat sooner. Hopefully they didn't order a hundred thousand wafer starts this week...
UniversalTruth
Veteran
Oh, any ideas when AMD is going to build any CPUs using GF's latest process?
Sorry but I don't follow very closely what is going there.
Do they plan to introduce new server chips, hi-end desktops and also lower APUs?
Or what exactly are they doing? APUs everywhere? Now?
They're likely to skip 20nm for high-performance APUs and go straight to 14nm FinFET. There is now a solid amount of evidence supporting this.
For low-power APUs it's less clear. These don't necessarily reach very high frequencies, so 20nm might make sense in 2015. But Mullins/Beema remains on 28nm.
There will be new "pure" CPUs (SoCs without graphics, really) this year, but they will be ARM-based. AMD might not design any more "pure" x86 CPUs, no one knows.
So are people now exchanging their Bitcoins for Dogecoins?
It makes sense for them to skip 20nm for the high frequency desktop parts with 16nm being only 1 year away.
I agree that 20nm makes sense for the low power parts, but only if it is early 2015, mid to late 2015 and they should just go 16nm.
UniversalTruth
Veteran
I cannot agree with you - anyways they are so late everywhere that it really doesn't make sense to skip in order to be even later..
Skipping 20nm(for high powered APUs) allows them to focus their limited budget and resources onto 16nm, which hopefully will translate into a quicker release of said products.
- Status
