Intel Joins the HPC PCIe Fray with Xeon “Phi”

AMD and NVIDIA are not the only companies to release new “compute” GPUs this week – Intel’s getting in on the fun as well. It’s been a long time coming, but the company has finally announced its Xeon “Phi” (pronounced: fy) “co-processors”, which shares the same goals as AMD’s FirePro and NVIDIA’s Tesla series do: crunching numbers for science. Although Intel’s been in HPC game for many years, this is the first release from the company that focuses strictly on highly-parallel computation.

If you’ve been following along all these years, Phi, given its PCIe base, might sound familiar. It’s a product that began its life as Larrabee, originally designed with more of a graphics focus in mind. With changed plans, Phi aims to compete with AMD and NVIDIA for workstation and server use, and so far, things are looking pretty attractive from the company (except the card itself, which is kind of lackluster in the aesthetics department).

The first released product is the 5110P, which includes 60 cores operating at 1.053GHz and offers 240 threads. The card features 8GB of memory with throughput of 320GB/s and utilizes 512-bit SIMD instructions. Unlike the FirePro and Tesla cards, Phi at the moment is designed exclusively for Linux (score one for Team Penguin!). Performance-wise, the card delivers 1 TFLOPs of double-precision floating-point performance, which is about on par with the competition performance-per-watt as the 5110P is labeled as a 225W offering.

A less-efficient offering will become available in the second-half of 2013. It’ll be known as the 3100, and offer nearly-identical performance but carry a 300W power envelope. The 5110P will hit general availability in January, and will carry an SRP of $2,649. The 3100 by comparison should settle in at about $2,000.

The reason Phi is called a “co-processor” is that it’s designed to be paired up with standard Xeons from the E5 family. This should give its users the best of both worlds computational-wise, though the design really isn’t much different from what the competition can offer. Intel, like AMD, has the unique ability to design both the CPU and GPU / computational card.

So far, the 5100P can be found in 7 of the TOP500 supercomputers in the world, with its biggest touting-point being Stampede, located at the Texas Advanced Computing Center. It currently sits in the #7 spot. In #253 is Beacon, at the University of Tennessee, which Intel says is the most power-efficient supercomputer of the bunch, recording 2.44 GFLOPs per watt.

With Xeon Phi, Intel has just made the HPC market a lot more interesting.