ARM – Possible Beginning of the End for AMD? – Part 1
Last month's move by AMD to license the initial 64-bit ARM cores for its server use could signal a brave attempt towards a new direction, as well as finally giving up any plans to competitively fight Intel on the X86 performance front. Aside of the still relatively healthy GPU business, what is the impact on AMD futures? This two parter focuses on the problem and possible solutions.
After covering AMD – among many other CPU vendors, of course – over nearly two decades, and seeing quite a few of these CPU vendors fall by the wayside, I thought of the company as a kind of 'survivor', the feisty underdog or even 'Luke Skywalker' freedom fighter which, even in spite of the 'Darth Vader'-esque Intel shadow of the past, and, more importantly, repeated top management blunders, still soldiers on to provide the PC community an alternative CPU and system platform source.
However, this past year's happenings, for the first time, make me feel that AMD's time as a serious CPU contender may be coming to an end. Why?
Firstly, the accumulated, consecutive top management blunders, which now cannot just be blamed on Hector Ruiz era any more, have become just too much for any company to survive, even if able to take heavy blows. Yes, the troubles started well over half a decade ago with 'Barcelona', but what followed was even worse. Even though the decision to go with the Bulldozer, then strip it down and release a huge lemon, was a huge mistake, an even bigger mistake was not doing a quick U-turn on this the moment it became clear that both the raw performance, and power/performance are going to suck. Abandoning that microarchitecture on time, and refocusing on an early replacement, would have been better, despite the resulting say a year lull in CPU sales – which ended up depressing anyway till now. Simply, Intel was still big enough to swallow one Pentium 4 disaster, and still moved on just on time to the Core – but AMD doesn't have the stamina to continue this way.
(somebody even bothered to make a viral video about the Bulldozer)
The continued series of delayed minor CPU core improvements in the shape of Piledriver and follow-up Steamroller, if it gets released, didn't help the cause, as Intel continued to widen its performance lead despite slowing down its own tick tock sequence by over a year at this moment – this is what lack of true competition does to the market, like it?
Secondly, the surrounding platform infrastructure – read uncore, chipsets and so on – followed the CPUs in being below par across the board. Slower memory controllers on top of truly bad cache latencies, which you can see from VR-Zone and others's benchmarks in plenty of cases, lower performing off-chipset PCIe v2 instead of on-CPU low latency PCIe v3, then the delays in introducing new stuff like USB3, SATA3, Thunderbolt etc. AMD is still stuck in 4 years old socket infrastructure for its CPUs, both desktop and servers.
Thirdly, the relationship with the vendors, channels, enthusiasts and yes, even media. We in Asia Pacific will still remember how, a decade ago and before, AMD fetted the Western press with first class paid press conferences while the Asian press, except Japan, got, literally – nothing. Only way later, when the crisis struck and everything became uncomfortable coach, did they remember the Asia Pacific. But OK, forget about the usual journalists' junket noises, handling the vendors and channels is still far more important.
And oh, all those 'early adopter' vendors and integrators who supported AMD in the early Opteron days when no big name wanted to, still do remember how AMD quickly abandoned them once they got the biggies like Dell on-board in those happy times prior to the first major fiasco, the 'Barcelona' – including even upping the prices post-launch. And, did that Dell keep AMD later on, once that crisis struck? Of course not, they just used them for better negotiating terms with their true parent – Intel. Now, no major vendor at all takes AMD seriously, at least not in the CPU business of any kind. And the smaller SI's, who depend on the unique performance or power usage advantages of the CPUs they peddle, are more likely to design something based on Chinese 'Loongson' MIPS or 'Shenwei' Alpha, which we covered here in detail, then on AMD, in the future – at least Loongson can use AMD I/O chipsets over HyperTransport, but too bad, they are also doing their own superior ones now.
Next point, even though not directly coupled with the CPUs, are the professional GPUs. With GCN GPU series, AMD actually has a superior GPU design to Nvidia Fermi for any serious GPU computation, with much higher SP FP and comparable DP FP, not to mention the time to market advantage they had till this month prior to the Tesla K20 launch. Was it ever seriously used? No – the workstation card was darn late, and the software stack is still behind Nvidia – not to even mention the smooth ramp up of Intel MIC aka Xeon Phi. This opportunity could have been used by AMD to bundle clusters of their GPGPUs with their own server CPUs as a good sales approach in HPC, even if the CPU performance itself is below par. Now, well, it is gone, at least till the high end Sea Islands GPU parts arrive.
(FirePro S10000 – missed the boat?)
Final point, of course, is the management replacement – kicking out Dirk Meyer now seems to have been yet another mistake, even though, yes, he was a better engineer than a commander. The continued loss of the top engineering brains is a great pity – till last month, AMD had the most of the old key DEC Alpha team in one place, for instance. That CPU architecture, till today, likely stands as the best there was. Rory is a much more jovial type, but the 'SuperBowl sports announcer' style of CEO talk has both pluses and minuses here – the Chinese in Lenovo would have loved it, though.
(AMD's current CEO, Rory Read)
Then, in parallel with this, comes the long prepared – since Dirk's departure, it seems – ARM announcement.
Now, as pointed out by many websites, this is not AMD creating its own custom cores, but just licensing the initial ARM Cortex 64-bit cores as they are, and putting them into whatever uncore environment they can, likely something tied to the SeaMicro interconnect they acquired for an obnoxiously high sum of money, likely the brainchild of then interim CEO Thomas Seifert and his connections. What does this mean?
AMD is now entering into a field where, instead of one single huge competitor, Intel, they will have half a dozen of huge and less huge competitors, ALL of them with a headstart over AMD – both in the ARM experience, custom core and uncore design for performance and/or power focus, customer relationships and, of course, influence within the ARM community. Besides the mammoth Samsung and not much smaller Qualcomm, the old gang like TI, the server focused Marvell and Calxeda (with their own server interconnects matching or exceeding SeaMicro one), and the prospective Chinese ARM server CPU vendor HiSilicon, a part of Huawei, there's also AMD's other nemesis there – NVidia.
Like them or not, their have both CPU and GPU headstart over AMD in the ARM world that may be extremely difficult to overcome – ever. And, the AMD announcement would just give Jen-Hsun Huang a good enough reason to deny AMD easy entry in ARM market, any segment. Think Nvidia can't do a good ARM server CPU? Think again…
Yet, for the money and resource tight AMD, focusing on ARM means de-emphasising X86 core competition. What then? Read the Part 2 for the possible scenarios.