Microsoft’s Catapult is a system that uses field-programmable gate arrays (FPGAs) to almost double the performance of a datacentre, all the while reducing power consumption.
How does it work? Simply put, FPGA boards plugged into server processors take care of special components, applications and services with the result that they are accelerated by almost a factor of two. For instance, Microsoft uses it to accelerate critical components of their Bing search engine to deliver more throughput and lower latencies.
The interesting bit here is that Microsoft used Stratix FPGAs for this particular endeavour – those are Altera’s flagship family of FPGA products where Stratix 10 is built on the Intel 14nm Tri-Gate process. The processors that these boards were plugged into were Intel Xeon server processors too.
“An entire philosophical change”
“To make things clearer, let’s start by clarifying the platform. There are four kinds of chips,” explains Gopi Kumar Bulusu, CEO and Chief Technologist, Sankhya Technologies.
Level One consists of the hard chips like those found in your desktop computer or laptop, which are manufactured by Intel, AMD and others.
Level Two are soft core ones like the ones licensed by ARM to manufacturers like Qualcomm, Samsung and the others.
Level Three are configurable processors that vendors like Tensilica used to work on before being acquired by Cadence Design Systems.
Level Four are synthetic processors by vendors like the Indian Sankhya Technologies.
Intel’s partnership with Altera allowed them to bypass stages two and three above to directly jump to the fourth level. By putting an FPGA next to the hard chip, they have a synthetic compute platform ready. If Intel manages to acquire Altera, it will be a statement that the fourth level is definitely on the horizon and could spark an industry wide movement towards it (potentially resulting in a crosshair over Xilinx, Altera’s primary competitor that relies on Taiwan Semiconductor Manufacturing Company’s (TSMC) foundry).
”In a nutshell, they are bringing the SoC mindset to the server market. It is an entire philosophical change that we are talking about here. Intel has previously never agreed that there is an alternative to their chip, but now this is happening. This is huge for synthetic processors in general,” adds Gopi Kumar Bulusu.
Acquiring Altera thus gives Intel the ability to not only improve its existing products, but also to quickly enter a whole new market without having to start from scratch.
Who will this affect?
The obvious beneficiary is for those designing advanced data centres. Intel’s Data Centre Group is one of its best growing with 11 per cent growth in revenue over the previous quarter. Facebook, Amazon, and Microsoft are potential clients who probably already make use of FPGA enabled processors for their servers, like the Microsoft Catapult we mentioned earlier.
Of course it would be short-sighted to say that this technology would be limited to the server processor space.
Apart from data centres, another area where this would allow Intel to create system-level solutions would be in telecom infrastructure gear. Andrew Feinberg, president and CEO of NetCracker Technology, talked about how network virtualisation (his company’s focus area) could create a more agile business model for communications service providers in the February edition of Forbes. How? He says that critical network functions such as routing and security would be moved from expensive purpose-built hardware to general-purpose servers and software. Reconfigurable chips from Intel could potentially power these.
Intel had also integrated FPGAs from Altera in their Atom series of chips way back in 2010, with an aim to enable embedded applications in industrial, military and transportation applications.
The cost for a design firm to switch programmable logic devices is very high due to the sunk cost in research and development that these firms would have already done, and this means that by the time the life cycle of existing Altera clients are over, Intel might be able to give them a cost advantage due to a deeply ingrained in-house Altera.
It is also safe to assume that prices of FPGAs will reduce if Intel seeks to use its Intel Custom Foundry margins as a tool to win price sensitive customers across from Xilinx. Further more, the chances of Altera gaining early-bird access to the latest process nodes might give it a technological as well faster time to market for both samples and later products. This could put Xilinx in a soup unless it figures out a way to make the most of its “one-team*” approach with TSMC forward.
* The “one-team” approach is from an announcement made in 2013, where the two companies are providing dedicated resources and will work together to co-optimize the FinFET process with Xilinx’s UltraScale™ architecture.