ARM Community: ARM In Servers: How Big Could This Be? - ARM Community

This is the week of the annual SuperComputing Show (SC11) in Seattle. About 15,000 people are expected to attend the week-long event. The news that was quite broadly picked up by the electronics press was NVIDIA’s announcement that the Barcelona Supercomputing Center is developing a hybrid Supercomputer that blends Tegra3 (quad Cortex-A9) devices with CUDA GPGPUs. I am equally excited about the development platform mentioned in the same release, and the implications for software ecosystem enablement. The availability of a software development kit (developed by SECO) featuring a board that incorporates a Tegra3 along with a discrete NVIDIA GPU and the NVIDIA CUDA parallel programming tool kit is going to kick start software development for ARM processor based platforms in the high performance computing (HPC) domain.

HPC applications have historically looked purely at throughput for their classification. Much is made of getting onto the “top 500.” But even in the HPC area, a number of organizations are starting to regard platforms as an “energy constrained system.” This measurement will, in many cases, migrate to “how much performance can be delivered inside a particular power envelope.” To that end, ARM is a founding member of the Energy Efficient High Performance Computing (EEHPC) community website. This was kicked off with a “Birds of a Feather” discussion at SC11. There were more than 100 attendees for the event. The HPC market is too small to sustain its own technology development and this is one of the primary reasons why commodity PC technology took over propriety technology about ten years ago. We foresee that a similar shift will occur over time to utilize technology that was originally generated for mobile and embedded applications.

Analysts continue to write up their reflections on recent announcements from Calxeda, HP and Applied Micro. The obvious overriding focus is to trying to size how big this ARM processor server market could be. Well, as someone that previously worked at a mid size semiconductor company and saw the stock rise from $6 to $102 in the craziness of the early 2000s’ technology bubble and didn’t sell (it lies below $6 again today), I am clearly not 100% accurate in my future predictions! And like any marketing guy, we really don’t like being held accountable to specific numbers! But I think in this area, I have some justification for being vague. ARM, along with our partners, is entering the server market with quite a different approach from the incumbent platforms, namely…

• 32-bit processors and platforms in the short-term
• Modest compute, highly integrated, energy efficient system-on-chips

There are too many pieces of analysis to connect to here, but this one from SemiAccurate does portray quite a balanced view of the opportunity and the challenges ahead. I attended HP’s announcement along with other founding partners in their Pathfinder program. Niall Dalton, director of High-Frequency Trading from Cantor Fitzgerald outlined how they could adopt such technology in their trading environment. His statement is that the execution of buy/sell transactions is incredibly sensitive to latency. These platforms need all the GHz they can get. However, this company, like many, has built up sensational amounts of data over the days, week, months and years of trading patterns; the ups and downs of individual stocks, etc. Analysts are constantly developing and modifying models that access this data to make forward looking predictions (they know they need an algorithm more sophisticated than asking me for stock picks!). Analysts are often running many scenarios in parallel. These are less time sensitive, require less compute horsepower and have a compute profile in which machine loading varies over time. A platform that scales the number of nodes being applied to these models and scenarios and can turn off unused processor technology to near zero power draw is an attractive proposition.

I have discussed previously that ARM believes that this market will segment over time, as companies look to deploy optimized devices to address specific server applications. We have been developing some models (Hey, maybe I need to access to that future Cantor Fitzgerald hardware!) to look at the ROI payback period for such an approach. Unfortunately, I cannot share this work publically yet, it is under the waterline I referred to in my last blog. But I can say that ratio of the savings in acquisition cost and running costs of the deployed hardware depends on a number of factors that include workload profile and size (both physically and performance requirements) of the installation, whether the technology is being utilized in an existing chassis or a green field site and the percentage of the day that the systems are fully utilized. ARM will publish more details on this topic in the future.

This is an exciting time to work at ARM. Like many companies, we spend quite a bit of energy at this time of year planning the next fiscal year. Of course, a significant portion of this is standard number crunching based on revenue and cost forecasts. But the market segment teams use this time to hone the market-specific initiatives for the coming twelve months. Some of you may be surprised to note that (with some margin of error regarding timing) we know much of what is coming in the next year. So what you expect to see during 2012? I think you know by now that I don’t predict the future well. But here goes…

• 2012 will be the year when the first ARM Powered® Server platforms ship for revenue
• It will be possible to buy a unit from more than one Original Equipment Manufacturer (OEM)
• Additional silicon companies beyond those discussed in the press up to this point will have announced devices optimized for this area

More soon…

Ian Ferguson, Director of Server Systems and Ecosystem, ARM, has spent years fighting from the corner of the underdog. Most of those scars are healing nicely. Ian is particularly passionate about taking ARM technology into new types of applications that do not exist or are at the very formative stages. Consequently, he is driving ARM’s server program with a view to reinvent the way the server function is implemented in networks as opposed to simply replacing incumbent platforms.