Intel has introduced its new Xeon 'Phi' co-processor during the International Supercomputing Conference in Hamburg, Germany. This is a new type of processor, reportedly the product of years of research to efficiently use x86 processors for 'High Performance Computing' (HPC) tasks. Up until now, Nvidia and AMD mostly dictated the markets of parallel processing with their high-end graphics cards. Modern GPUs contain several hundred cores, which are perfectly suited for general purpose calculations (GPGPU).
From Larrabee to MIC
Intel saw opportunity in the market in 2008 already, starting development on 'Larrabee'. This GPU was to compete with AMD and Nvidia's graphics card, offering a mix between professional computing capabilities and powerful gaming performance. The project was put on hold indefinitely in 2010. The architecture was developed further and renamed 'Many Integrated Core' (MIC), focusing exclusively on the High Performance Computing tasks. This seems like a step in the right direction as Intel has now finished a co-processor that performs incredibly well on paper. The 'Knights Corner', as the new Xeon co-processor has been named, is equipped with over 50 x86 cores. Like the Ivy Bridge processors, the chip uses 22nm Tri-gate transistors. Intel has added this enormous processor, with over 1 TFLOP of double-precision computing power, to a PCI-Express x8 expansion card with 8 GB of GDDR5 memory. The amount of memory suggests the use of either a 256-bit or 512-bit interface.
Smart co-processor with own operating system
The Xeon 'Phi' co-processor is more than just a powerful multi-core processor on its own PCB, as Intel has developed a complete Linux-based operating system for the card as well. As soon as the card is placed into a PCI-e slot, it will operate as a stand-alone system, processing tasks from the main server and exchanging data at high speeds through the PCI-e interface. The addition of an operating system to the Xeon Phi expansion cards should simplify the creation of a cluster system. A simple driver that would recognize the devices as a cluster would already suffice.
Because the Xeon 'Phi' is based on the x86 processor architecture, existing software can directly benefit from the added cores, while developers for Nvidia's Tesla and AMD's Firestream cards first have to convert their data to a different programming language. This makes Intel's solution suited for a wide variety of traditional x86-based software.
Available late 2012?
Intel has yet to reveal any details on pricing and availability. Production however, has already started, as the supercomputer top 500 list features a system built around this Xeon 'Phi' expansion cards. It's likely to take until the end of the year for the hardware to be available in large quantities. The biggest competitor for Intel's new Xeon Phi is Nvidia's Tesla K20, a card that will be based on the anticipated 'GK110' GPU. AMD will undoubtedly respond with an alternative of its own, based on the high-end 'Tahiti' chip. This chip is based on the 'Graphics Core Next' architecture, which also focuses on GPGPU.