I recently had the pleasure of speaking to Andrew Feldman, founder of SeaMicro which was eventually acquired by AMD (AMD). At AMD, Mr. Feldman not only runs the SeaMicro division (which is seeing some real traction, as evidenced by the recent wins at Verizon (VZ)), but is also in charge of the company's server processors from low power ARM (ARMH) based chips to the highest end Opteron parts for 2 socket/4 socket, more traditional "server" applications. While AMD's battle ahead looks tough, I believe that the company has a solid strategy and, more importantly, a management team that is likely to be able to execute on that strategy.
People And Strategy Matter
I am now significantly more positive on AMD's server prospects for two major reasons. The first is simply that Mr. Feldman's vision/strategy is bold but at the same time realistic. He realizes that it's not about taking on the 800 lb gorilla head on in the areas in which that gorilla is very strong, but instead about fighting more concentrated battles in which a smaller - and perhaps more nimble - player can really "win".
On top of this, the strategy itself is actually quite good and plays right into AMD's strengths. So, what are those strengths you may ask? Well, let's take a look at what AMD has had tremendous success at, even as it continues into the later stages of its turnaround, so far.
Server Experience And IP
There's no doubt that AMD has plenty of experience building general purpose server processors. It is among the few shops still operating today, along with Intel (INTC), IBM (IBM), and Oracle (ORCL), that has server chips in its DNA. Of course, things really haven't been easy for any player aside from Intel, which has gobbled up the vast majority of the market. It's tough to forget, though, that AMD's Opteron once ruled the roost and gave Intel a good whipping back in 2003, particularly as AMD's market share in servers is now about 5%.
That being said, the game has been "reset" so to speak. While it is unlikely that in the near term AMD will make meaningful inroads in the 2S and above space with its current lineup of multi-socket Opterons, the hyper-scale space is still new and, frankly, fair game. Intel is off to an early start in this space with its "Avoton" parts (based on the 22 nanometer "Silvermont" Atom), but with the "Avoton" lead architect now working for Samsung (SSNLF), and with Qualcomm's (QCOM) ex-CMO now working on "enterprise related initiatives), it's clear that this space is going to heat up pretty quickly. AMD intends to be a major player here, leveraging what it believes to be key advantages over the nascent ARM players like Cavium Networks (CAVM) and Applied Micro (AMCC), as well as even over the behemoth, Intel.
Semi-custom For The Win?
What are those advantages? Well the first, Mr. Feldman points out to me, is that AMD has a very rich heritage of building server chips and has established relationships with key server OEMs and even the "roll-your-own" server guys like Facebook (FB). But more importantly, Mr. Feldman believes that a key differentiator is that AMD not only has the right IP to build server chips, but the flexibility to develop custom-tailored solutions for the biggest cloud guys.
So, for example, if Facebook needs a chip with a special, perhaps Facebook-designed IP block to accelerate a certain, highly used function (for example, modern networking chips today often sport dedicated packet processing hardware), then AMD could integrate it into an SoC specifically for Facebook. It's the same "semi-custom" playbook that AMD used to win over Microsoft (MSFT) and Sony (SNE) for the latest game consoles (as well as the other "major" wins that Lisa Su hinted about on the most recent call).
AMD's flexibility is that it can use its own internally developed X86 cores, off-the-shelf ARM cores, and - if the ARM market really blossoms - there's a chance that AMD might do a full custom ARM core. Indeed, Mr. Feldman indicated to me that even a ground-up, custom ARM core (and not an off-the-shelf ARM design) costs significantly less to build than a full custom X86 core, owing to the fact that the ARM instruction set (and in particular the 64-bit subset) is apparently simpler to implement (although a true apples-to-apples comparison of development costs for a custom core with the same features/functionality would be very difficult to quantify given that so many other factors play into this).
The trade-off, of course, is software compatibility. While quite a lot of software enablement has been going on in the ARM ecosystem, X86 still has a software compatibility advantage. For those that "roll their own software", it may be cheaper to move to ARM for some parts of the data-center than for, say, the enterprise market, but it's still a non-trivial cost to recompile and validate software. This means that the ARM guys need to offer a tangible advantage in performance per watt, per dollar (also known as "total cost of ownership") in order for the ARM ecosystem to gain material traction here.
Not Giving Up On X86
Something that has always bugged me about AMD's server roadmap has been the fact that AMD outright replaced its own low power X86 core based products with the ARM-core based products. However, Mr. Feldman was quick to point out that the company absolutely had not given up on its low power X86 chips even in the server space and that further details would be announced at a later time.
So, for example, if the ARM ecosystem fails to gain traction in this space, this would actually position AMD as the only second source for low power micro-server chips. While this hasn't worked out with the "big cores" (sub 5% market share), Intel's low power cores are first and foremost designed for smartphones and scale up from there. So, this suggests that if AMD were to simply target its low power X86 processor cores at a slightly different design point on the performance/watt curve (or performance/mm^2) it could still have something that's pretty competitive if it's willing to make the right trade-offs.
It's Not All Roses, Though
The biggest roadblock here is likely to be, as it generally is for anybody trying to play in servers, Intel. The chip behemoth is feeling the pressure from the hordes of players trying to break into the market, and as a result has responded with a very aggressive roadmap, as shown here:
Not only is Intel using its traditional Xeon E3 parts as well as its new Atom based parts to play in micro-servers, it's even building a highly integrated, purpose-built based chip based on its "big" cores. This is likely in a bid to counter what Cavium Networks and Applied Micro are doing with Project Thunder and X-Gene, respectively, as these two have signaled that their cores would be much more along the lines of "big" cores than the smaller Atom/ARM A57 type chips.
That being said, we will have to wait and see how AMD's "Seattle" system-on-chip sells. If AMD can leverage its ownership of SeaMicro to gain a foothold in this space, and if it can offer competitive silicon at decent gross margins (remember, AMD is likely willing to accept lower margins than Intel is), then it could still see a real degree of success - but, of course, it won't be until late 2014 that we'll really know how well AMD will do here.
AMD has truly "cleaned house" and brought in management that seem to really understand the situation that the company is in. There are no guarantees for anybody in this space, but for the first time in a long time, I find myself legitimately excited to see if AMD can succeed in the cloud server market. While its competition is still fierce and the landscape unknown, I do know that I walked away from my conversation with Mr. Feldman very impressed. And, at the very least, that opens the door to situation where "this time it'll be different" really could apply.