ARM In Servers – Hyperscale Or Hyperbole?

ARM In Server Highlights

• A developing ecosystem for micro and hyperscale servers using SoCs and Open motherboards
• Semiconductor manufacturers beginning to produce products
• Server brands positioning their first commercial offerings
• Research projects demonstrate potential
• Shipments of 783 million mobile device and 148 digital cameras running ARM chips in 2012
• 64-bit chips are a potential disruption
• Self-built hyperscale systems are difficult to count
• Our outlook is positive, but we expect a slower take-off than ARM of course

ARM’s dominance of the mobile space has helped it achieve higher profitability than other microprocessor suppliers, although this is also due to its status as a license-only supplier (see Figure 1).
We listened into a very interesting Webinar recently by ARM, HP, Ubuntu and Red Hat, which raised some important issues about the opportunity for adding ARM processors to servers. You’ll be interested to read more about the potential advantages and pitfalls.

Clear Advantages

Our speakers pointed to the massive success ARM processors have had in the mobile device market: we believe they are used in 783 million of the 906 mobile devices sold in 2012 – to which we could add the 148 million digital camera shipments – which currently fall out of our classification. ARM has several advantages here:

1. It is not itself a manufacturer, taking revenues from licensing only
2. Addressing mobile devices has given it a full understanding of low-power requirements – important for servers where the ‘power per dollar per watt’ is a guiding design principal
3. There is a full ecosystem and massive scale economies in the production of its chips today

However the servers being considered are different from the x86, RISC and mainframe products which make up the whole market today: our presenters talked almost exclusively about Server on Chip (SoC) products being used in hyperscale scale-out architectures and fabric computing. These systems are so different our panel believes we will need a whole new set of benchmarks to assess performance.
Operating systems such as Red Hat and Ubuntu Linux (although the latter is a just a single-user product as far as we understand) are already available and run for ARM processors. Potential applications include Hadoop, public Cloud and scientific/technical computing.
For customer and server analysts the introduction of ARM-based servers will also make the market more innovative and interesting the panel suggested, breaking through today’s 2-vendor (HP and IBM) dominated market.

Potential Pitfalls

As in early server processing the ARM camp is faced with a tricky transition from 32-bit to 64-bit chips – this as a game-changer for AMD verses Intel back in 2002, as it was for Digital’s Alpha processor a decade before. We suspect that significant adoption of microservers will need to wait until 64-bit SoCs are available – there’s little point in experimenting with 32-bit in a market that has yet to take off.
It’s all well and good saying ARM servers will be completely different, but the market for hyperscale computing is currently far smaller than that for general purpose commercial products – Big Data/Agile Data applications such as Hadoop will undoubtedly leverage hyperscale architectures in future, but are themselves a small (though fast growing) sector. As for scientific/technical computing – we’ve never seen it grow above 9% of server spending in the last 30 years and don’t expect it to do so now.

Lots Of Development, Few Sales

We believe the current use of ARM processors in hyperscale servers is currently very minimal, although the vendors are jockeying for position and there’s lots of experimentation. In particular:

• AMD looks likely to add ARM64 SoCs in 2014 – perhaps to supplant the Intel Atom ones used in products by its acquired Seamico microserver supplier; it also plans to add ARM-based modular server motherboards in future like its recently introduced Open 3.0
• Dell is reported last November to have prototyped systems with 64-bit ARM processor from Applied Micro Circuits, in addition to the 32-bit ARM chips from Marvel and Calxeda it had already tested; it already sells Viking microservers
• Freescale Semiconductor announced that it was thinking about using Cortex-A50 series processor license from ARM to extend its digital networking activity into server computing
• HP’s has announced that its Project Moonshot servers will initially be based on Texas Instruments’ KeyStone II SoC ARM Cortex-A15 cores, while it will continue to work with other microprocessor suppliers such as Calxeda to offer customers choice; it has recently launched its first Moonshot production microservers based on Intel Atom chips
• IBM has been working with ARM since 2008 on mobile chips, but we haven’t yet seen anything in the server area – with Power and System z chips it has huge research going on in other areas, it might be better to look at what Lenovo could do
• The Barcelona Supercomputing Center was reported to be using Tegra ARM chips for a quad-core server experiment supported by the EU
• Further back in time the Advanced Processor Technologies Research Group (APT) at the University of Manchester designed the Spiking Neural Network Architecture (SpiNNaker) using ARM chips to simulate the human brain.

However none of the 267 references for Hadoop currently mention using ARM processors. In addition Intel just launched its own Hadoop distribution, which will undoubtedly give it the advantage for now and perhaps push back the ARM tide for now.

The Difficulty Of Sizing Non-Branded Server Markets

The difficulty in trying to size the hyperscale computing market is that a lot of the servers are self-built. A number of large public Cloud, contract manufacturers and scientific users are building their own systems without using branded products, which increases the total number of servers shipped and installed beyond the 15 million and 47.3 million we report in our tracker for 2012. This is important for the types of system proposed by the panel, as SoC and potentially Open motherboards are going to be used more in self-built systems. We’ll do our best to size these activities in coming months, but already suspect that most self-built servers use x86 chips today.

Some Conclusions – Low Power In A Balanced System

2 years ago Intel admitted it had sold very few chips for mobile phones and none for tablets. When it came to describe the opportunities for Atom in servers it made the point that the power envelope advantages of a smaller form factor are eroded as you added the functions needed for servers, which is backed up by IBM’s Deep Blue system used to beat Gary Kasparov at chess in 1997 – it used low power chips, but the savings on the processors were taken up by extra power need for the interconnects. Admittedly Xeon and Opteron include certain overheads for running the x86 instruction set, but this is small in comparison with the fuller set of components needed to build a balanced server.
In fact Intel’s OEMs have taken a dramatically alternative approach, increasing ambient temperature toleration to reduce cooling requirements and get more server bang for electricity bucks – reminding us that there are 3 elements that can be tweaked in the ‘performance per watt per dollar’ equation.
The panel’s enthusiasm is infectious – suggesting that ARM will radically change the server market. In the past there have been similar times, but breakthroughs were slow as Unix overtook mainframes and x86 overtook Unix. In addition MPP servers like those proposed by ARM will be ‘easy to scale, hard to programme’ as usual, verses the ‘hard to scale, easy to programme’ SMP architectures usual in the x86 model.
We totally reject their suggestion that ARM chips might account for 15% to 20% of server shipments in 2015. As an old fogey I think that will take 10 years if they can gain some kind of foothold in the next 2. Whatever happens, this is going to be an interesting market to watch.