New AMD Roadmap for 2013 (donanimhaber.com)

Discussion in 'CPUs and Overclocking' started by jones377, Nov 2, 2012.

  1. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,153
    Likes Received:
    39
    Gaming for example will see a moderate benefit, also socalled multimedia applications. Assuming its an exe compiled for it. But the benefit is that you can take existing SSE code and just recompile it for the AVX2 benefit and 256bit execution. So the work (and cost) from the developer side is basicly 0. So there is no excuse not to use it in all newer applications and games.

    AVX2 is radically revolutionary due to providing vector equivalents of every scalar instruction.
     
    #126 ShintaiDK, Nov 7, 2012
    Last edited: Nov 7, 2012
  2. blckgrffn

    blckgrffn Diamond Member

    Joined:
    May 1, 2003
    Messages:
    6,644
    Likes Received:
    0
    Hopefully we see some awesome then :thumbsup:
     
  3. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,153
    Likes Received:
    39
    Plus remember unoptimized code brings another 10-15% on top of that due to other improvements. Branchprediction, double the L1 and L2 bandwidth and etc.
     
  4. inf64

    inf64 Platinum Member

    Joined:
    Mar 11, 2011
    Messages:
    2,257
    Likes Received:
    7
    Well those 10-15% is what the benchmarks will show( except maybe sisoft sandra). Intel needs to "help" developers use even AVX,let alone AVX2. It takes time when it comes to new ISA extensions ,it's just how the software market works.
     
  5. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,153
    Likes Received:
    39
    AVX2 you just need compiler support if you code is already SSE. And compilers already support AVX2.
     
    #130 ShintaiDK, Nov 7, 2012
    Last edited: Nov 7, 2012
  6. NTMBK

    NTMBK Diamond Member

    Joined:
    Nov 14, 2011
    Messages:
    6,995
    Likes Received:
    33
    You do know that that isn't true, right?

    If you have code which uses SSE intrinsics, the compiler will not automatically map those to 256-bit wide AVX2.
     
  7. inf64

    inf64 Platinum Member

    Joined:
    Mar 11, 2011
    Messages:
    2,257
    Likes Received:
    7
    Yes but you need to recompile the code,hand tune the parts that cannot be just recompiled,then do test/verify process and at the end ship it to market. It takes time and you will see how "fast" the software market will react. Maybe in 2015 we will see some big software suits that rely on integer vector(a lot) gain support for the AVX2 ISA. That's my estimate at least.
     
  8. pelov

    pelov Diamond Member

    Joined:
    Dec 6, 2011
    Messages:
    3,512
    Likes Received:
    0
    GloFo has claimed that tape outs are set to happen in 2014 with products in 2015, so AMD will still be a year behind. Granted, Intel will be on 14nm for at least two years so we'll see how that works out ;P

    I also think GloFo and AMD both may have a bit of trouble with the 14nm-XM node, as GloFo has been working very closely with ARM to tune it for their SoCs whereas AMD has been left out in the cold, so to speak. If Steamroller is cancelled/delayed then an Excavator chip at 14nm-XM wouldn't likely see release until late 2015
     
    #133 pelov, Nov 7, 2012
    Last edited: Nov 7, 2012
  9. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,153
    Likes Received:
    39
    Remember GloFos 14XM is low power only and is the size of 20nm.

    If they tapeout the first one in 2014. Then products is more likely to come in 2016. Just look at 28nm. tapeout in 2010, products in 2012.

    I am pretty sure Excavator is completely cancelled.
     
    #134 ShintaiDK, Nov 7, 2012
    Last edited: Nov 7, 2012
  10. pelov

    pelov Diamond Member

    Joined:
    Dec 6, 2011
    Messages:
    3,512
    Likes Received:
    0
    That's not exactly true, as the 14nm-XM uses 14nm FinFETs but with 20nm elements

    Well, they expected customer tapeouts in 2014, given a year that would mean products in 2015 would be a safe bet. Whatever AMD's high powered chip release dates are, they'll certainly be well behind the ARM SoC's on the same node. You're right, the 14nm-XM node at GloFo is being tuned for ARM rather than AMD.

    And that's why
     
  11. Idontcare

    Idontcare Elite Member

    Joined:
    Oct 10, 1999
    Messages:
    21,130
    Likes Received:
    0
    GloFo claims 14nm-XM tapeouts are happening in 2013:
    Production is expected to hit in 2014 with the goal being to compete against Intel:
     
  12. pelov

    pelov Diamond Member

    Joined:
    Dec 6, 2011
    Messages:
    3,512
    Likes Received:
    0
    Ha! Wow, then they've sped up? They only recently taped out 20nm late last year :/ Looks like they were serious about becoming profitable in the coming near future.

    How viable is that 14nm-XM for AMD though? With regards to high power chips that is. Their Jaguar-or-whatever-follows SoCs could potentially benefit, but what about their high end Opterons and laptop parts? GloFo and AMD both have been pretty hush about that as of late. AMD roadmaps show absolutely nothing past 28nm next year
     
  13. mrmt

    mrmt Diamond Member

    Joined:
    Aug 18, 2012
    Messages:
    3,976
    Likes Received:
    0
    From what I get 14XM isn't a full fledged 14nm process, but a kind of hybrid process using 14 and 20nm elements, so not as good as a 14nm process, but slightly better than others 20nm process. Also while transistors are smaller, the size of the chip won't shrink because the chip structures will be 20nm, which means that AMD will lag in die size for the time being. At least 14XM should yields gains in power consumption.

    If AMD can ship 14XM products by H214 they might be able to offer some competition to Intel at slightly better terms than they are offering now, if they somehow improve Steamroller faster than Intel is improving Core, but I think it's a stretch to bet that they will be able to run for the performance crown.

    That said, they might have APUs ready by that time frame, but it is fairly safe to say that their server business will be toast by that time, as they won't be able to field something barely competitive at least until 2015. They can't stay two years without a barely competitive product. I wouldn't really have an idea of how their other business will fare against Haswell, but everything is pointing to a shrinking share on notebooks and desktops once Haswell arrives. In 2014 they will have to fight an uphill battle to break Intel momentum and claw back OEM share by the time Intel will be pushing Broadwell on the market. Not a good prospect here.

    Here's the cause of the big if. GLF screwed 32nm SOI, then screwed 28nm bulk and according to the info you brought here they screwed 20nm bulk too. Their track record does not show the bare minimum execution capacity, they are the AMD of the foundry world, always hyping, always delivering less than it should. Also remember that a very cash strapped company decided to pay a hell of a money to get out of a contract with them, something they wouldn't do if they were assured of their capacity of execute the proposed road map.
     
  14. Idontcare

    Idontcare Elite Member

    Joined:
    Oct 10, 1999
    Messages:
    21,130
    Likes Received:
    0
    I don't expect 14nm-XM to enable higher GHz over that of 20nm for high-performance stuff, but I do expect it to dramatically lower power so that the performance/W goes up.

    I base this on what I observed with the 32nm->22nm transition for Intel.

    Those finfets don't seem to be all that great in improving Fmax at the top end, but they do enable much lower voltages for the same clockspeed which in turns dramatically lowers the power consumption.
     
  15. Idontcare

    Idontcare Elite Member

    Joined:
    Oct 10, 1999
    Messages:
    21,130
    Likes Received:
    0
    Yeah, I agree. I was just trying to be hopeful. I'm a process development engineer myself so I can convince myself that there is upside potential behind these press releases, that management is holding back just a bit to keep some positive news should all the toast land jelly side up in 2014.

    But I can't deny the more probabilistic outcome here being that which you spell out above.
     
  16. ShintaiDK

    ShintaiDK Lifer

    Joined:
    Apr 22, 2012
    Messages:
    20,153
    Likes Received:
    39
    I will just leave this :p

    [​IMG]
    [​IMG]
     
    #141 ShintaiDK, Nov 7, 2012
    Last edited: Nov 7, 2012
  17. Khato

    Khato Golden Member

    Joined:
    Jul 15, 2001
    Messages:
    1,003
    Likes Received:
    2
    Yup, it's always a possibility that we won't see a repeat of their track record with the 14XM process... but I'm not going to believe it 'til they have actual products rolling through the fabs in mass production. They're just now getting 28nm under control and want us to believe that they're going to have 14nm finfets in 2 years? It's not impossible, but it certainly does sound like it's something that they just finished cooking up in their labs and are now running it through the fabs to see if it actually translates to mass production - note the fact that when asked about the status they simply stated that test silicon is running through their fab without mention of whether or not it's looking good.
     
  18. mrmt

    mrmt Diamond Member

    Joined:
    Aug 18, 2012
    Messages:
    3,976
    Likes Received:
    0
    Not likely. Globalfoundries was always seen as a long term investment, and not something they were going to have returns tomorrow. ATIC is sinking a lot of money there and I don't doubt that they will wait a few years before throwing the towel, that's the way that ATIC run its businesses.

    But I do agree with your observations that this speed up is very suspect. They bought factories that acquired manufacturing processes from others, not a full fledged foundry, there is a lot to learn and expertise to develop internally before they are able to develop a custom process.

    I don't think they should be aiming to compete against Intel or beat TSMC, but try to become a safe, reliable company to compete against TSMC and only then try to take on Intel. So far their current strategy generated a lot of negative cash flows and a shattered credibility, and once credibility is lost, it lost, as some smaller players cannot afford a 12 months delay in their product launch, they won't tie up with an unreliable partner.
     
  19. pablo87

    pablo87 Senior member

    Joined:
    Nov 5, 2012
    Messages:
    374
    Likes Received:
    0
    intel's cost structure is going up, not down. if you look at capex, it has literally doubled the past 2 years and as a result, depreciation has increased, and depn charge is the base cost of all intel processors. without growth, depreciation has to catch up to capex, or there will be a massive one time charge...Even if by some minor miracle capex returns to $5B annual, Intel would still have to depreciate $32B over a 4 year period, $8B annually, over 300MM processors, that is $26.67 per unit...More likely is capex will be much higher than this, intel will stretch the depreciation period to 5+ years, but the result will be similar.

    If you add actual fab operating costs and packaging and overhead its easy to conclude Intel cannot make money selling parts for $50, they can only continue to do so by way of subsidy justified by the perfectly legal meetcomp. So already the balance between the subsidized parts and the subsidizing ones has to tilt in favour of the latter, if profits are to be maintained.

    And then there is the grim competitive reality (as opposed to self inflicted wound as above) that ARM based tablets and smartphones are taking share from x86 and now you have a real problem: volume is going down which means based on simple overhead incl. depreciation allocation, costs have to go up. Ad so first things first, its better AMD cry than they do and so a gear shift from meet comp to meet volume (75MM units per quarter) has to be made and IMO, is being made.

    Eventually, AMD will probably go out of business ($450MM per quarter + interest + whatever _____you have to take from Global Foundries becomes very very expensive and unsustainable when you're shipping less and less units at lower and lower prices - just try to model 10MM units per quarter at $30 and see how ugly the bottom line quickly gets).

    But will x86 annual volume settle down at a 300MM per year equilibrium? If not, then Intel must replace that volume with tablet/smartphone biz which is much lower ASP so while they're doing that, they'll be raising prices on 286, Celeron, Pentium parts they previously had to sell below cost all these years) until there is no such thing as an x86 processor under $100.

    Last but not least, what happens when high end corporate PC's that contain almost exclusively $200+ Intel processors become just another device to access corporate data (which I am told, is the IT vision these days)? More pain for Intel and, higher CPU prices for the masses still.

    All that being said, I agree with MRMT that Intel is not doing anything unethical or untoward here, its just the market reality - there's been an inflection point with several aftershocks - first ARM, then iphone, then ipad, then Android, then Amazon and Google's business models.

    The irony: intel's most lucrative and stable business after Server: high performance computers with DISCRETE video cards.
     
  20. MisterMac

    MisterMac Senior member

    Joined:
    Sep 16, 2011
    Messages:
    779
    Likes Received:
    0

    Interesting viewpoint - but don't you think Intel has:

    A. Plans to lower cost per wafer, per working CHIP etc etc - to loose some Gross margins but also lower costs to make the drop less noticeable.

    b. If IT wants just a flat device to access corporate data - corporate needs big boxes to store data and handle all those simoultaneous compute requests in both networking and direct resources.

    Which benefits intels cash cow in servers - which then in turn could make more use of MIC outside HPC for coprocessing of certain user workloads.


    I can't believe Intel doesn't have several plans in order - with only a few required to succeed in order to continue to gain or stand neutral at ground in terms of revenue\profit\margins.


    You tell me how many users can work on a E5 - and the cost of it versus X hp\dell boxes for the enterprise.
    Intel would sell more expensive chips at a greater margin than they're business enterprise crappy pentium box workstations.

    Result: Enterprise saves money - intel gains more revenue.
    It's just moved from low end celeron\pentium to expensive xeon.

    Something i'm quite sure intel would like.
     
    #145 MisterMac, Nov 7, 2012
    Last edited: Nov 7, 2012
  21. mrmt

    mrmt Diamond Member

    Joined:
    Aug 18, 2012
    Messages:
    3,976
    Likes Received:
    0
    That's rather obvious, as revenues are going up too. The ratio for operating profits is around 50%, not a bad rate by any standards.

    I think there are a lot of wild assumptions here. First you are assuming that all Intel investment is in machinery that will depreciate in five years, which it isn't true. The fab itself and some other machinery are depreciated over far longer periods. Second, even if you have your asset fully depreciated but it still has value on it, you can go the impairment route. Third, the Capex isn't factories only. Acquisitions, patents and shares in other companies are also Capex but they don't necessarily depreciate at all.

    So while your analysis has some merit in pointing out a risk, that investing too much might lead to overproduction risks, your numbers are overshot. You probably got that number from a trend that Intel depreciates 80-90% of the last FY Capex, but because of the factors I mentioned in the last paragraph this isn't a golden rule.

    In 2011 they acquired McAfee and were building a new factory (the second item should impact this year CAPEX too), so to get a better estimative of how much they have to depreciate to keep the business running you have to exclude those two items.

    Intel does not disclose the amount of money they are using in the factory, but we might have a very good idea looking their cash flows fillings in Q213. As a ballpark I would use 6.5 billion.

    Intel fundamental unit cost isn't cost per chip, but cost per waffer, so Intel fundamental measure is cash flow per waffer. If they can manufacture one chip worth 100 out of a 50USD waffer, they are going to be in the same situation if they are selling two 50USD chips or 25 4USD chips. In the previous example ASP fell but Intel cash flow and gross margins were untouched. So the shift to smaller processor should allow smaller ASP for a given cash flow. This is where Atom, smartphones and tablets fit.

    But if we are talking here about smaller cash flows, then things can become a little different, but with current volumes Intel could go as low as 46% in gross margins and still break even.
     
    #146 mrmt, Nov 7, 2012
    Last edited: Nov 7, 2012
  22. NTMBK

    NTMBK Diamond Member

    Joined:
    Nov 14, 2011
    Messages:
    6,995
    Likes Received:
    33
    Wafer. It's spelt wafer.
     
  23. mrmt

    mrmt Diamond Member

    Joined:
    Aug 18, 2012
    Messages:
    3,976
    Likes Received:
    0
    Thanks :)
     
  24. krumme

    krumme Diamond Member

    Joined:
    Oct 9, 2009
    Messages:
    3,579
    Likes Received:
    18
    Well they fit, but there is no way the can generate the same profit as fx. the server market, or highend desktop market for the simple reason the good server cpu literaly brings money in your hand, being a cathalyst for the entire investment , while the Atom - well does the same as A15 just giving say a better experience in angry birds. How much worth is that? What exactly is it people is missing playing or using the A15?

    Price per wafer is just going down. TSMC and Samsung is making sure of that. This is different from competing to ppt slides from GF :) - and that is reflected in the share price for Intel.

    It is red ocean market, if you dont have any other technology or brand to build on after the die is cut. There is profit, but no where near the usual level Intel is used to. Intel will have to adjust and differentiate their cost structure.
     
  25. mrmt

    mrmt Diamond Member

    Joined:
    Aug 18, 2012
    Messages:
    3,976
    Likes Received:
    0
    I'm not sure if I'd call CPU market a red ocean yet. The bottom market certainly is, nobody wants to develop a custom chip for a blu-ray player, but in cars, phones, tablets, convertibles, there is still a lot of space for differentiation. If you look at the biggest ARM players, Apple, Samsung, Qualcomm, Nvidia, every one of them is trying to grow by differentiation, not by making their operation more efficient as they would have to do in a true red ocean environment.

    I see a weakness on the ARM business model. ARM chain is composed from an IP researcher, then a lot of people customizing designs with overlapping efforts and then a few foundries, some of them spending money on the same node but not sharing research... and everyone here wants higher margins.

    There might be a lot of inefficiencies looking at the chain, and when facing other chains like MIPS the bigger player wins, but Intel is a different game. Intel as a vertical player can be very focused and deliver a fine tuned solution for a certain niche, and there isn't such thing as fight for margins internally, the final result is what counts.

    But even if Intel can't get enough margins to keep its current business model, they can adapt. Intel overcame the 180nm debacle, then overcame netburst, then moved to tick/tock, then went for power efficiency.... every one of these events generated a different Intel. If business change again, they have both the cash and the technology to adapt.
     
Loading...