On Monday, Intel made public its end of life strategy for the Knights Landing “KNL” Phi product set. The announcement makes official what has already been widely acknowledged after Intel scrapped plans for KNL successor Knights Hill and pulled PCIe-based KNL coprocessor cards from the market last year.
In a product change notification [PDF], Intel announced discontinuance of its entire Knights Landing socketed 7200-series Xeon line: part numbers 7210, 7210F, 7230, 7230F, 7250, 7250F, 7290, and 7290F.
The last day to order KNL products is August 31, 2018, and the final shipment date is July 19, 2019.
Marketed as an easier-to-program alternative to GPUs for compute-intensive workloads, the Phi line kicked off in 2012 with great expectations, but never reached the success that its maker hoped for. Knights Landing was the second-generation Phi family, featuring between 64 to 72 Atom Silvermont cores. Launched at ISC 2016, KNL diverged from its coprocessor predecessor Knights Corner in its “self-hosted” form factor and also introduced the AVX-512 instruction set. List price ranged from $1,881 to $3,368.
Knights Landing was the first Intel processor to offer the company’s Omni-Path Architecture (OPA) fabric integrated into the package (the option is denoted by the suffix F in the model number). The Skylake-F Xeon chips that Intel introduced in the third quarter of 2017 are, according to the company’s specification database ARK, still available.
The first instantiation of Phi was a development prototype called Knights Ferry, revealed in 2010. Emerging from Intel’s abandoned Larrabee product, the Many Integrated Core accelerator was made available to select partners in 2011 but was never widely available.
In accounting for the cancellation of KNL, Intel stated only that market demand had shifted to other Intel products. Despite having a number of supporters and several prominent system wins, including TACC’s flagship system, Stampede2, and Berkeley Lab’s CORI supercomputer, Knights Landing failed to achieve broader market penetration. General-purpose GPUs from Nvidia had gotten a several-year head start (CUDA arrived in 2007) and there was also a matter of mismanaged expectations. Intel marketing initially heralded the Phi as being as easy to use as a Xeon, claiming a significant programmability advantage over Nvidia GPGPUs, but many users found getting a meaningful performance increase required extensive code rewrites and tuning. There’s still no easy button for parallel programming.
So is Phi dead? Officially, the AI-themed Knights Mill Xeon Phi is still in play. The parts are listed on ARK, but the line was absent from the roadmap that Intel presented at its AI developer conference in May. There’s also the rumored successor to Phi (spotted by Andreas Stiller of Heise), called “Knights Cove,” that would replace the Phi’s atom cores with up to 44 Ice Lake cores. It would seem this Knights Cove, or a future iteration of it, is likely to play a part in U.S. exascale plans. According to its retooled CORAL contract, Intel is slated to deliver an exascale system to Argonne National Lab in 2021. Last year, Intel’s Trish Damkroger reported that the company would be replacing the next-gen Phi (Knights Hill) with “a new platform and new microarchitecture specifically designed for exascale.”