It’s been almost three years since Intel first detailed its upcoming Xeon Phi hardware, codenamed Knights Landing, and the company is now shipping silicon to developers who can pony up the cash.
There are two Knights Landing-based Xeon Phi systems available. One is a liquid-cooled system with six DIMM slots, dual gigabit Ethernet ports, and the Developer Edition of the Xeon Phi — 16GB of MCDRAM, AVX-512, and 72 CPU cores. The other is a higher-end 2U rack with four hot-swappable nodes. Base price on the equipment is $5,000 and $20,000 — and while that’s extremely high by consumer standards, the $5,000 system isn’t necessarily a bad deal if you work in the field and are looking to experiment with high performance computing optimizations.
Knights Landing is the third manycore x86 architecture that Intel has developed. The previous iteration of the hardware, Knights Corner, used a modified Intel Pentium core (P54C). The new Knights Landing, in contrast, is based on a heavily modified Airmont 14nm Atom core — and while Intel hasn’t shared all the details of those modifications, what they have shared point to significant changes.
Knights Landing is made up of 36 tiles. Each tile contains two CPU cores and two VPUs (vector processing units) per core (total of four per tile). Unlike previous Xeon Phi cards, which functioned solely as co-processors, the new Knights Landing is designed to self-boot and can control an operating system as the native processor.
Each CPU can handle up to four threads and the chip’s out of order resources have been substantially improved. Intel is claiming a 3x per-thread performance improvement over older Knights Corner products with an estimated 3 TFLOPS of performance available per chip. The core also uses 16GB of MCDRAM and can be configured for multiple memory configurations that tap both the onboard memory and the available DDR4 memory channels depending on which model is most advantageous for a particular piece of software.
On the one hand, Knights Landing is a straightforward evolution of Knights Corner, meant to keep pace with Pascal and keep Intel competitive in the lucrative high-performance computing (HPC) market. On the other, however, it’s meant to help developers write more parallel software. With single-thread performance almost stalled out in modern CPUs, the industry needs to foster approaches that increase parallelism in common applications. Such efforts often begin at the supercomputing level and hopefully lead to best practices that become commonplace a few years later.
Nvidia’s Pascal will go toe-to-toe with Knights Landing when it debuts, but Intel’s got a substantial development lead on this one, and it’s obviously eager to put the architecture into developer hands.