Intel's Investment In FPGAs Could Pay Dividends In Microsoft Data Centers And Mark The Advent Of AI Accelerators
In June of last year, Intel announced a $16.7B acquisition of chip designer Altera, a Silicon Valley bellwether known for producing FPGAs (Field Programmable Gate Arrays), PLDs (Programmable Logic Devices), Embedded Processors, and ASICs (Application Specific Integrated Circuits). At the time of the announcement, Intel CEO Bryan Krzanich said, “Intel’s growth strategy is to expand our core assets into profitable, complementary market segments. With this acquisition, we will harness the power of Moore’s Law to make the next generation of solutions not just better, but able to do more.” That’s a nebulous way of saying Intel had plans to leverage Altera technology to boost its own profits, which is typically the point of acquisitions such as this. How Altera’s technology would be used, however, wasn’t fully explained at the time, but it is becoming more clear by the day.
Xeon processors, which will leverage Altera technology. Those processors are due to arrive later this year and they should allow Intel to market more specialized, configurable hardware for accelerating a multitude of different workloads. FPGAs feature an array of logic gates that can be programmed to perform a myriad of tasks, and they can be optimized and re-programmed on the fly, as new workloads emerge or compute demands and algorithms change.
The flexibility inherent to FPGAs is also at the core of Microsoft’s Project Catapult, which is a code-name for the technology behind Microsoft’s hyperscale acceleration fabric. Microsoft explains that Project Catapult “is at the center of a comprehensive set of investments Microsoft is making to build a supercomputing substrate that can accelerate our efforts in networking, security, cloud services and artificial intelligence.” Efficiently accelerating all of those different workloads is not something easily done with an ASIC – there isn’t some one-size-fits all chip out there to address the specific needs of each use case and data set. An FPGA, however, can be programmed to accelerate the algorithms associated with the specialized workloads and data sets of each specific application.
For the past few years, Project Catapult has been evolving within Microsoft. At first, FPGAs were used to accelerate Bing’s Indexserve engine. A number of servers – over 1600 – were initially outfitted with FPGAs that were connected via a secondary network. The FPGAs were programmed to accelerate specific search-related algorithms, and the end result was in improvement in latency and a 50% reduction in the number of servers required to process workloads. Subsequently, Microsoft decided to put Project Catapult into production. A v2 architecture that enabled configurable clouds was eventually laid out and last year Microsoft ramped-up to large-scale production in Bing and Azure. However, Microsoft now also plans to employ Project Catapult FPGA accelerator boards in “nearly every new production server” and use the technology for virtually all of its services. In the future, these services could also encompass forward-looking AI-assisted, contextually aware computing and search models, similar to what was detailed in a recent Microsoft patent filing.
We’re speculating a bit, but with Intel putting FPGAs on Xeon processors, it should be able to reduce latency even further, since off-chip communications will be minimized due to the coherently attached in-socket FPGA adjacent to the central host processor. Density should be increased as well, and there may be power and efficiency benefits too. Should Microsoft’s Configurable Cloud architecture be adopted by others, and Intel’s future Xeons address the needs of Project Catapult, it could mean very good things for Intel in the lucrative, high-margin enterprise server platform space.