Placeholder Image

Subtitles section Play video

  • I decided to productize the Knights landing processor in our supercomputers

  • because it meets a an important customer need we have a diverse set of customers

  • we've got government customers people in all areas of science industrial

  • customers enterprise customers and with that diversity of customers comes a

  • diversity of applications needs some applications work better with multi-core

  • CPUs that have very large memories but others work better with the Xeon Phi

  • five many core processors that have very high memory bandwidth and i compute peak

  • new Knights landing xeon phi processors a significant advancement over the

  • previous generation Knights corner coprocessor first and foremost it's self

  • hosted in self booting

  • this provides an easier programming model and allows you to build more cost

  • effective systems

  • secondarily it has significantly higher performance in the previous generation

  • it has 16 gigabyte of on package memory with extremely high memory bandwidth and

  • we found that very high memory bandwidth is a key contributor to high compute

  • performance

  • it also has more cores the version that we're using uses 64 to 68 cores

  • they have wider vector units they've got more threads per core

  • this results in a single processor with over three teraflops a performance in

  • over 250 threads per socket which help you soak up that high memory band with

  • praise adaptive supercomputing strategy supports diverse processor technologies

  • in a single system

  • many of our customers historically have had systems with mixed processor times

  • and this is no exception with the new Knights landing systems that were

  • shipping this summer

  • those systems have large partitions of Xeon multi-core processors and Xeon Phi

  • many core processors a homogeneous system allows you to run a single

  • capability or hero job across the entire system can be good for getting great

  • benchmark numbers but a heterogeneous system is often better for running a

  • diverse set of applications you can match each application to the right

  • processor type and optimize the overall performance of your workflow

  • so let's take a look at how we've implemented the new Knights landing

  • processor and the Cray XE compute blades here you can see two different compute

  • blaze they actually have an identical form factor and they have an identical

  • base blade with the cray

  • aries high-performance network chip on it each aries network chip has four

  • integrated mix or network interface cards to connect to for local compute

  • notes on your left we can see the Xeon multi-core compute nodes each of those

  • is a dual socket xeon node each of those sockets has four memory dimms four memory

  • channels we have a total of eight memory dimms per Xion compute note on your right

  • we have the Knights landing blade again the same physical blade but we have a

  • different daughter card on this plate and this daughter card has single socket

  • knights landing compute notes

  • so there are a total of four compute nodes for sockets on this blade each

  • one of those sockets has six external memory dimms for the large capacity

  • memory and of course internal to the Knights landing processor

  • we've got a very high bandwidth on package memory system so we have this

  • hierarchical memory system each of those processors also is over three teraflops

  • of performance and so when you have four of those on a blade and we have 48

  • blades in a cabinet

  • you can have up to an excessive 575 teraflops of knights landing performance

  • in a single cry XE cabinet

  • the blades are the same form factor as i mentioned so they're interchangeable and

  • you have any number of zeon compute blades and and knights landing compute

  • blades in the same system

I decided to productize the Knights landing processor in our supercomputers

Subtitles and vocabulary

Operation of videos Adjust the video here to display the subtitles

B2 H-INT US compute memory landing processor phi blade

Insider Look: New Intel® Xeon Phi™ processor on the Cray® XC™ Supercomputer

  • 33 1
    alex posted on 2017/02/11
Video vocabulary