原创 ARM in the data centre: Separating myths from reality

There has been a lot of industry hype about the market for microservers as we close out 2012.

While the prospect of the ARM architecture making it into the data centre is tremendous fodder for the tech press, the majority of the coverage seems to be viewing the topic through rose-tinted glasses.

The hype machine has been amply fed with a constant stream of bold pie-in-the-sky predictions from ARM CEO Warren East, Simon Segars and other ARM executives on how much market share the firm will inevitably take from Intel.

It's interesting that a company which doesn't yet make a product is feeding this hype cycle more than the customers who are actually tasked with bringing those products to market. Indeed, we've not heard much on this front lately from any of the ARM licensees to validate such grand aspirations, perhaps because they are the ones who live in the world of product reality.

Taking a broader view of the segment, there is very little competition for Intel today. AMD continues its painful death spiral, and no other companies seem to have the capability of shipping products into the data centre with scale.

Nvidia is a peripheral bit player in HPC and has had some success but not enough to lessen its dependency on consumer and workstation discrete graphics for the foreseeable future.

Now to be clear, I think competition in the data centre is a good thing and sorely needed. Customers need it, and companies like Intel need it to make better products. But let's face it, Microservers is not even that big of a market anyway, with only a handful of important customers like Facebook, Google and Amazon.

If and when the ARM architecture makes it into the data centre it has the prospect to offer this competition, but a lot of things need to happen and a number of questions need to be answered first.

Deliver 64bit first, talk after
Firstly, ARM has to deliver its new architecture on time with its first ARM V8 implementations with Cortex A53 and A57. 64bit computing is a baseline requirement to enter the space and today it is, unfortunately, MIA in ARM world.

If ARM's products don't have 64bit, they can be considered dead on arrival in the data centre, and the reality is that it will be at least two years before we see any ARM products that have this capability. Fact.

Two years is a long time in product cycles, which means that as this new category of microservers grows, customers will only have one company they can buy product from for the foreseeable future, Intel.

ARM also has to deliver on time and with scale. While Cortex A15 is a tremendous new architecture it has largely limped out of the gate in 2012 and we won't see the majority of designs until next year. Hopefully, for ARM's sake V8 will have a smoother birth in 2014.

Secondly, ARM will need to build an eco-system in record time. The server world is a complex one, driven by a broad eco-system of applications and workloads that have been optimised over many years to run exceptionally well on X86 processors. Moving these workloads over to a competing architecture is not a trivial matter, just look at the launch of Windows RT on the consumer side to see the pitfalls of lack of application compatibility and low performance.

Nobody can argue that low power in the data centre isn't a tremendous driver today, but applications compatibility remains essential. In this area ARM will need to step up and create an eco-system that minimises the deficit of compatible applications and workloads for its architecture rather than simply leaving this to customers.

Management will also be a huge issue that needs to be addressed. A rather large number of companies have aspirations to participate in the ARM server market. Calxeda, Applied Micro, Samsung and AMD are all committed to bringing products to market in the coming years and more will follow.

With the exception of AMD, very few of these companies have any real expertise in data centre management which could be problematic as they try and convince enterprise IT executives to deploy a mixed environment of X86 and ARM.

All that effort for such low ROI?
The ROI will have to be compelling and the risk to disruption minimal. At this point I've not heard anyone articulate how exactly this will happen.

In addition, one has to question the logistics of how competing ARM vendors will work together in a mixed customer environment, given they all have their own unique implementations and fabric strategies.

"If you look at the number of companies looking to do microservers, the market starts looking crowded in a hurry," said analyst David Kanter, noting that at best there were 30 potential customers for Microservers, though he estimated the real number may be closer to 15.

Kanter explained that to have any hope of success in the market, companies who wanted in would need a serious breadth and quality of IP for CPUs, caches, memory, interconnects, networking, etc.

Indeed, in 2014, servers are going to need large L3 caches, DDR4 memory controllers, 10GBE or Infiniband networking, coherent interconnets (e.g., QPI), PCI-E gen 3, not to mention comprehensive power management and RAS features.

"Some of those pieces you can license, but quite a few you cannot, and even licensing something may not be quite as good as doing a custom design," said Kanter, noting that most ARM licensees fell short in this respect, and that only AMD and Applied Micro really had what it took at this point in time.

"Applied Micro is interesting because it is using its own custom CPU core that is higher performance than ARM's reference designs," he said, noting that this gave the firm an advantage in terms of performance and cost structure, should they achieve sufficient volume.

This week, we're going to see yet another media cycle around microservers as Intel prepares for the upcoming launch of Centerton, the firm's first server class SoC for the data centre. This will be followed by the company's 22-nm Avoton product in 2013, which will almost assuredly have a follow-on in 2014.

That's a grand total of three Intel product cycles before we even see our first real ARM server product enter the market.

While nobody questions that Intel most certainly has its work cut out for it in the ARM dominated world of smartphones, the firm has proven it can deliver a competitive product, albeit not at scale yet. Now the pressure is on Intel to build market share and customers in the coming years.

The path forward is equally daunting in the data centre and arguably less forgiving for ARM. Enterprise IT managers demand a robust TCO model, have little patience for downtime and ARM is unproven in these areas of real-world deployments.

It may also not be worth the effort.

"If you believe that ARM-based servers can grab 10% market share, how much *silicon* revenue is that? And how many SoC companies can it sustain?" asked Kanter, pointing out that in a market of only about 10 million units, there wasn't room for five to eight companies.

"If you look at it that way, it's starting to sound like a pretty low volume business," he said.

The ARM eco-system is coming to the data centre, that much is certain, but the road traveled to get there will be a long and challenging one with more competitive pressure than the firm has ever felt before. And success may not taste as sweet as the firm thinks it will be.

Sylvie Barak

EE Times