Intel Knights Landing: Claimed 4x An NVIDIA K40 (on some applications)

Joe Curley (Director of Marketing in the Technical Computing Group at Intel Corporation) just completed his webinar on BrightTalk, “The Faster Path to Discovery: New Details on the Intel® Xeon Phi™ Product Family” that disclosed new details on the upcoming Knights Landing massively=parallel chip including a claimed 4x performance improvement over the NVIDIA K40 on some applications. The Intel Knights Landing co-processor will be used in the upcoming NERSC Cori leadership class supercomputer. Following is a short summary of the presentation. Click on the webinar link and register to view the full presentation and Q&A.

The Knights Landing processor utilizes modified Silvermont cores that also power Intel’s next generation Atom processors. The updated KNL cores are manufactured on a 14nm die process. It was hinted during the Q&A that additional features have been included in Knights Landing due to the revised manufacturing process.

This will be the first Intel Xeon Phi to feature stacked memory that provides a 5x improvement over DDR4 memory bandwidth and a nice increase in memory capacity.

More information about Micron”s MCDRAM is covered by Rich Brueckner at insideHPC, “Slidecast: Micron HMC Memory Technology to Enhance Knights Landing”

Claimed 4x the performance of an NVIDIA K40

Pricing should be aggressive.

Joe Curley gave an interesting answer during the Q&A session when asked if programmers should still consider using offload mode for programming. Since KNL is a bootable device (meaning it can be booted and used as an SMP system) – just like the current generation Knights Corner processors, Joe stated that one can favor utilizing the Xeon/KNL hybrid systems via a network (presumably MPI) model rather than by offload programming. This is consistent with Intel’s view that traditional programming models can enter the exascale era.