Bringing Competition To The Server Processor Market.

Applied Micro leads the charge to infiltrate the $12 billion server processor market with ARM-based ICs.

This is not a trivial task. The $54 billion gorilla standing in Applied Micro’s way is Intel with a 90% plus share of the server processor market.

So what, if any, are Applied Micro’s selling points compared to Intel’s?

First and foremost there’s the business model.

“Competition is what we’re bringing,” says Gaurav Singh, vp of technical strategy at Applied Micro, “in most other markets there is very healthy competition with multiple silicon customers.”

The ARM model allows for multiple silicon vendors to compete against Intel’s proprietary architecture, sole-sourced chips. In addition to Applied Micro, there are AMD, Broadcom and Cavium which are trying to get ARM-based server chips to market. Samsung and Qualcomm may also dive in. Opening up the server processor market to competition should rid the market of $1,000 ICs.

Server manufacturers, and end-users who design their own servers, know that, currently, Intel has a lock on the server processor market and can name its prices.

So how does Applied Micro’s processor, called X-Gene 1 currently in production at TSMC, stack up against Intel’s server chips?

“Intel’s Xeon (its top end server chip) dissipates 100W and costs $1000 whereas X-Gene 1 dissipates 30-40W, delivers 80-90% of the computer performance of Xeon and sells in the hundreds of dollars,” replies Singh, vp technical strategy at Applied Micro.

Singh is reluctant to be more precise about the expected price level for X-Gene 1.

“People are looking for an alternative to Intel because their margins are very high – 80% + on Xeon – and Intel has been inflexible in its product offerings,” said Singh.

In what respect inflexible? “For instance in-memory computing has become very important – people want to store processed data in DRAM rather than disc,” responds Singh.

This means that a server processor has to be capable of accessing large amounts of DRAM. Intel’s Xeon chips can access enough DRAM for in-memory computing but Intel’s cheaper Atom-based server chips can’t.

“With Intel, if customers want sufficient memory for in-memory computing they have to go to Xeon,” says Singh.

So what’s stopping Intel putting enough memory interfaces on its Atom-based server chips to provide sufficient access to DRAM for in-memory computing?

“If Intel allows Atom to have as much memory as Xeon then they lose the margins of Xeon,” replies Singh.

By contrast, X-Gene was designed to have high memory density with four channels of DDR memory.

Applied Micro’s next selling point against Intel is that: “Xeon needs other chips like a network interface card and external storage controller, whereas X-Gene has all this integrated on the die,” says Singh.

How about Intel’s argument that, while ARM-based chips can perform many functions in servers, they can’t perform the heavy lifting as effectively as Xeon?

“That was true,” says Sigh, “but we’re working on that. At the very high end people will still use Intel for some time because the benefits of power efficiency diminish at these very high levels.”

Is Intel’s argument valid that its process advantage means its chips perform better?

“Absolutely not,” replies Singh, “X-Gene 1 on 40nm is competing with Intel on 22nm and delivering 80-90% of the compute performance.”

Moreover, Applied Micro’s second generation server processor, X-Gene 2 made on a 28nm process, will be sampling in Q1 2015.

“We’re also working on X-Gene 3 on 16nm Finfet which will have between 20 and 64 cores,” adds Singh.

After 16nm, Singh expects process to become less of a differentiator.

“As you go from 16 to 14 to 10nm the performance is staying the same but the cost per transistor is going up,” says Singh.

As well as challenging Intel, Applied Micro has challenges from other companies bringing out ARM-based server ICs. What advantages does it have over these?

“Our advantage is early mover advantage,” replies Singh, “we started five to six years ago with Power PC and did a 64-bit chip then decided that ARM was the way to go. We have already been through all the stringent server OEM qualification cycles and we’ve shipped $1 million worth of product. We’re making headway.”


Comments

22 comments

  1. Whatever next? FPGAs, GPUs and multiple ARM vendors joining SPARC and Power – you’re making me feel quite sorry for Intel.

  2. A good article here today from EETimes on where server design is going. FPGAs are of course faster but more hassle to program than GPUs so I suspect this will only take off in mega-installations but it shows that even a top of the range Xeon is still far too slow for many applications and that nowadays the x86 device has to be thought of more as a ‘station controller’ for the real horsepower devices attached to it.

    http://www.eetimes.com/document.asp?doc_id=1323500&_mc=NL_EET_EDT_EET_programmablelogicdesignline_20140814&cid=NL_EET_EDT_EET_programmablelogicdesignline_20140814&elq=1f289dcb2d3d47f3baae16610ae72d2b&elqCampaignId=18577

    I suspect there is a market somewhere in servers for the ARM 64 bit cores, unlike the A15 which I upset an ARM director by calling an upgraded phone chip. But of course even phones are now 64 bit and I suspect most licences and hence future design focus will go on the consumer end of the link, simply because ARM get more royalties at that end. That’s not to say the A57 could not be redesigned into something to challenge Xeon, but it’s roughly an order of magnitude off at the moment.

  3. I bow, of course, to yours, and Mrs B’s, vastly superior knowledge of servers, Mike, I just never realised that the term ‘heavy lifting’ was a term of art applicable only to single core processing and could never be applied to the work done by multiple core processing.

  4. I’m sure the managers and programmers working on such systems (e.g. my wife) would love your expertise in converting these ‘heavy’ threads to processor arrays. Trust her (and Gene Amdahl) – it’s almost impossible.

    And the ones that they do parallelise they use nVidia GPUs for.

  5. I think heavy lifting means compute intensive tasks like processing, analysing or searching data which can be done just as well by an array of processor cores as by a single core.

  6. I think you are misunderstanding what heavy lifting means. The Cavium and Applied Micro devices are capable of large amounts of ‘small lifting’ but they cannot provide the single thread performance that many applications require.

    Instead these devices are aimed more at providing large numbers of separate connections to the server and there is a definite market for such a product. However it has to be worth the effort of compiling ALL applications on the server to ARM architecture and often there is one that isn’t available.

    There are also applications where parallelised performance is useful but the latest servers are beginning to use GPUs for these applications which far out-perform any current ARM device for most parallel processing targets.

  7. Does the market for servers really grow without limit? Are competitors not a bit late to this party?
    Surely a farm of motherboards does not suddenly outlive their usefulness?
    Is there a second hand market?

  8. Yes that does look a big beast, Rupert, sampling q4, a bit behind Applied Micro, presumably in volume production in H1 2015. Should be capable of this “heavy lifting” which people say can only be done by Xeons.

  9. I think you ought to look at Cavium’s Thunder X product

    They are the only ones using architecture license.

    A very impressive device, with 48 custom ARM cores on one die.

    (Everyone else stops at eight because that’s all that ARM’s interconnect can support)

  10. It is a particularly interesting struggle this one, Stooriefit, not least because it will be decided by technical excellence and responsiveness to customers rather than the dark arts of marketing.

  11. ‘the best products invariably rise to the top’

    Well was the 68000 better than the 8088?

  12. Well nobody forces anyone to buy an Intel engined server. Or an iPhone. Or anything else that has ever had a large market share. As long as there are alternatives available then all well and good. You could think of IBM as the Samsung and AMD as the Nokia of servers – they’re there but you need a special reason to use them ?

    The best products invariably rise to the top and however much one may not like some of Intel’s marketing methods – which as far as I’m aware were never applied to servers anyway – Xeon is a fantastic product line best suited to the task in hand.

  13. With IBM & Oracle at the top end ARM won’t be able to touch their volumes or margins.

    It will be interesting to see what the effect of ARM eating away at the low end of Intel’s business is – it could actually narrow the gap between volume % and revenue % as ARM cores pick off Intel’s low margin business.

    A sensible manager would restructure the business and lap up those increased margins at lower volumes – no more busy fool business. I suspect that won’t allow for overall growth however, so barring a big divvy Wall St won’t wear it.

    I wonder what will happen in the supercomputer space? If any market can could handle more lower power cores and heterogeneous architectures it would have to be that one.

  14. My fault, I should have made it clear that 92% was units not revenues. But I think my main point still stands – that 75% market share is too much for anyone – if you’re in favour of healthily competitive markets.

  15. OK now I see what you are doing. Don’t you think revenue is rather more important than units ? Our numbers are very similar then.

    IBM and Oracle’s units are expensive very heavy lifters that Intel can’t compete with so they get far more per processor but if you count units you are making these equal to an Atom based unit used in a home server.

    And no way do AMD have 5% of server revenue anymore. Oracle were their major customer there but they scrapped that range.

  16. Well Gartner says Intel has 92% of the units and 75% of the dollars.

    Gartner: 2013 stats

    Market share. Revenue share

    Intel. 92. 75

    AMD. 7. 5

    IBM, Oracle,
    Fujitsu etc.
    1. 20

  17. David – As I said at the start of this thread Intel DO NOT HAVE >90% of the server processor market. I really do not know where you get this number you keep repeating.

    To be precise Intel have over 97% of the x86 server market but this only represents 76% of the total server market, making Intel’s share 73%, or maybe 74% if we allow for a few Itanium sales. Still a long way from 92% !

    And if Applied Micro, AMD or any of the others were to succeed Intel would still have 97% of the x86 server market, though the 73% might drop a little, maybe all the way to 72% before people realised their mistake.

  18. ‘A lot of competition’ is a subjective phrase and I wonder if it can be applied accurately to a market in which one company has 92% of the market and half a dozen others share 8%

  19. I’ve no idea if the current pope is a catholic, nor care.

    But there’s already a lot of competition in the server market and I suspect Applied and maybe AMD as well could be casualties in it.

  20. And these guys also repeat the 92% market share figure Mike. It’s a Gartner figure
    http://seekingalpha.com/article/2403475-intel-losing-dominance-in-the-server-segment

  21. Well see this:
    http://www.fool.com/investing/general/2014/06/09/amd-vs-intel-will-arm-server-chips-tip-the-balance.aspx:
    Is competition a good thing? Will it bring the price of server processors down? Is the Pope a Catholic?

  22. I’m sure Intel would love to have 90% plus of the server market. Unfortunately there’s these two small little companies called IBM and Oracle that in addition to using Intel processors have their own processors in full production.

    As for delivering 80-90% of the compute performance of a Xeon I won’t be holding my breath. These current ARM solutions rely on being able to use large numbers of cores equally and we all know there are few server applications in which this is the case. ARM need something akin to Intel’s Turbo Boost before they can compete effectively, though I think it’s safe to assume they are working on it.

Leave a Reply

Your email address will not be published. Required fields are marked *

*