At the GigaOm Structure event in San Francisco yesterday, Intel revealed that it has developed a hybrid server processor that combines a Xeon E5 CPU with an FPGA on the same package. The programmable sidekick can be used to accelerate specific workloads, while the Xeon provides the comforts of x86 compatibility and Intel's existing infrastructure. The two components are connected via the CPU's QuickPath Interconnect. According to Diane Bryant, GM of Intel's Datacenter Group, "the FPGA has direct access to the Xeon cache hierarchy and system memory."
Amusingly, Intel Director of Cloud Computing Raejeanne Skillern told The Register that the implementation is "not the prettiest." "But it's functional and it fits," she added, and the potential gains sound tantalizing. The FPGA reportedly improves performance with certain algorithms by up to an order of magnitude, and using QPI to attach the thing is said to deliver an additional 2X boost.
There are no details on the companion chip, but Intel has worked with numerous FPGA providers as part of its Custom Foundry business. It partnered with Tabula and Achronix on 22-nm parts, and with Altera on a 14-nm one. That Altera chip is actually an ARM-based design, but the company behind it has several different FPGA products in its stable.
In related news, Microsoft recently revealed a "Catapult" project that paired Xeon CPUs with daughter cards loaded with Altera FPGAs. That endeavor reportedly delivered big performance gains in Bing search with relatively little additional power draw. In fact, Catapult was successful enough that Microsoft plans to deploy FPGA-infused servers in its production datacenters early next year. I wouldn't be surprised if those servers had hybrid Xeons onboard.