Intel's next

Intel made multiple computing announcements at the 2014 International Supercomputing Conference, including several disclosures around Knights Landing, the next-generation version of Xeon Phi. Upcoming versions of the microarchitecture will nearly triple its theoretical performance, include support for Intel's next-generation fabric, codenamed Omniscale, and will feature Micron's second-generation Hybrid Memory Cube technology.

According to Micron, the 16GB of DDR4 on-board the upcoming Knights Landing cards (due in the back half of 2015) will feature 5x the bandwidth of sustained DDR4 while consuming one-third the energy per bit and occupying half the footprint. It's the energy-per-bit that's the truly huge advance - one of the most fundamental problems facing the HPC industry is the need for vastly improved energy efficiency in order to hit exascale performance without utterly blowing the power budget. Whether HMC actually delivers that theoretical gain in most use cases is something we won't know until the cards ship, but it's a noteworthy advance. [The image at the top of the story is a Knights Ferry/original Xeon Phi die.]

Knights Landing itself will pack up to 72 Airmont (14nm Silvermont) Atom cores with two 512-bit vector units per core, with estimated peak performance of more than 3 TFLOPs (double precision) per card. Unlike previous Xeon Phi cards, the new Knights Landing will be available as both a standalone card connected over the PCIe bus and as a socketed processor attached to a motherboard via Intel's Omniscale fabric.

Interestingly, the implication of Intel's ISC presentation is that the technologies we see rolling out today in supercomputing will eventually make their way to desktops and other mainstream systems, though it may take some years. The company's timeline shows that it typically takes 6-8 years for a technology debuting at the top spot on the supercomputer charts to make its way to the bottom, and roughly nine years for technology to move from the supercomputing charts to the single-socket market. I suspect that trend may have accelerated in recent years - with Moore's law stalling, there's been talk of moving technologies like HMC to the mainstream computing market after the DDR4 generation, not some 15 years from now.

Omniscale and Knights Landing are both scheduled to debut a year or more from now, in the second half of 2015, alongside next-generation Xeons. One of the interesting trends in the Top 500 list of supercomputers this year is how few new systems have made it onto the list. Ars Technica reports that only one new system has entered the top 10 in the new June 2014 list, and the 384th system in November 2013 is now the 500th system today. In last year's June 2013 list, the 500th system had been the 322nd system previously.

From 1994-2008, the performance of the 500th-placed system grew 90% every year. From 2008-2013, it increased by just 55%, despite the rise of accelerator cards and GPGPUs. Again, this suggests that the rate at which new, performance-boosting technologies roll out across the market may indeed accelerate - especially if those technologies, like Hybrid Memory Cube, deliver much-needed performance and efficiency improvements. Intel has other improvements planned, including the introduction of silicon photonics alongside its Omniscale fabric, but it hasn't disclosed much about how such devices will integrate into existing networks or what kind of performance users can expect.