0 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Fujitsu Server-Infrastruktur für die 2nd Generation Intel Xeon Phi Product Family (codenamed Knights Landing)

Eric Schnepf

Fujitsu Technology Solutions GmbH

HPCN-Workshop 2016, Göttingen, 10. Mai 2016

1 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

More than 30 Years Fujitsu Supercomputing

AP3000

PRIMEPOWER HPC2500

Most Efficient Performance in Top500 (Nov. 2008)

PRIMEQUEST

FX1 K computer

FX10

Exascale

Japan’s First

Vector (Array)

Supercomputer(1977)

F230-75APU

VPP5000

VPP300/700

VPP500

VP Series

NWT* Developed with NAL

World’s Fastest

Vector Processor (1999)

ⒸJAXA

No.1 in Top500

(June and November 2011)

Japan’s Largest Cluster in Top500 (July 2004)

PRIMERGY BX900 / BX400 Cluster node

HX600 Cluster node

PRIMERGY RX200 Cluster node

PRIMERGY CX400

next x86 generation No.1 in Top500

(Nov. 1993)

Gordon Bell Prize

(1994, 95, 96)

AP1000

World’s Most

Scalable

Supercomputer

(2003)

SPARC Enterprise

*NWT: Numerical Wind Tunnel

FX100

new

2 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Select Your Preferred PRIMERGY Server Technology

Scalability,

Infrastructure

density

BX900

Capacity

Capability

Flexibility to address all kinds of customer requirements

PRIMERGY CX400 skinless server

Massive scale-out due ultra dense server

GPU coprocessor support

PRIMERGY blade server

Industry leading blade server density

PRIMERGY rack server

CELSIUS workstations

BX400

CELSIUS

CX400

RX2530 Scalability,

Compute density

RX2540

3 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

FUJITSU Server PRIMERGY Scale-out Systems

CX400 CX2550 CX2570 CX600 CX1640

Compact server nodes with

high power efficiency to realize

large scale-out solutions for

HPC, hosting, and hyper-

converged computing at lower

overall costs

Dual socket

server node in a

highly condensed

half-wide, 1U form

factor

Dual socket server

node for ambitious

HPC, analytics and

visualization solutions

HPC optimized scale-out server

platform based on 2nd Generation Intel

Xeon Phi (“Knights Landing”)

technology

Single socket Xeon Phi

server node for

significant performance

boost in parallel-

processing

Chassis Server Nodes Chassis Server Node

Platform for HPC, hosting, and hyper-converged stacks

More computing power in less space

Lower costs for energy due to shared power & cooling

Optional liquid cooling for lower cooling costs and higher density

4 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

2013: 0.25 TF

2014: 0.7 TF

2015: 1 TF

Intel Xeon Phi “Knights Landing” Ecosystem

2013: 1.2 TF

2016: 3 TF

Second Intel Xeon Phi generation

Single socket compute node

Significant performance boost versus conventional

Xeon CPU

Eliminates cost overhead by Xeon CPU and other

components

Fujitsu’s offering:

Platform: PRIMERGY system based on Xeon Phi

Rollout including

• Parallel application enabling program by

FSE Toulouse to support at application

adoption, porting and tuning

• Dedicated true HPC Software Stack

• Parallel file system optimization

Xeon Phi

Xeon CPU

TF = Teraflop

Attention:

Performance

information still tbc

5 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

2nd generation Intel® Xeon Phi™ Product Family

Achieve significant performance gains Source: https://software.intel.com/en-us/articles/what-

disclosures-has-intel-made-about-knights-landing

…

. . .

. . .

Integrated Fabric

up to 72 Cores

Processor Package

Compute

Intel® Xeon® Processor Binary-Compatible

3+ TFLOPS, 3X ST (single-thread) perf. vs KNC

2D Mesh Architecture

Out-of-Order Cores

On-Package Memory Up to 16 GB at launch

Over ~400 GB/s STREAM at launch

Platform Memory

Up to 384 GB DDR4

Fabric (optional) 1st Intel processor to integrate

Knights

Landing

6 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Fujitsu PRIMERGY CX600 M1

Your platform for highly

parallel computing

The FUJITSU Server PRIMERGY CX600 M1 is

the perfect choice for highly parallel applications in

the area of scientific research, product

development and business intelligence. Up to

eight server nodes per 2U and Intel® Xeon Phi™

Processors make for new levels of compute

density.

7 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

PRIMERGY CX600 M1 Chassis

Front side

Rear side

PSU (x4pcs)

CX1640 M1 Node (Max. x8 Node) See next page

FAN (x5pcs)

Front Panel

Chassis interface board

Mid-Plane

Power Distribution Board

Slot Numbering

Front view

Rear view

8 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

PRIMERGY CX1640 M1 Node

SATADOM

DIMM (Max. x6pcs)

PCIe Riser card

KNL CPU

HDD / SSD

- SATA HDD (Max. x1)

- SATA SSD (Max. x2)

- PCIe SSD (Max. x1)

HDD Backplane

PCIe card (x1pcs)

Front

view

Rear view

VGA LAN 1

LAN Port0

LAN 2

LAN Port1

USB 1

USB Port0

USB 2

USB Port1 Power-Button ID-Button

External Interface Connector From left side,

- ID LED

- Global Error LED

- HDD1 Status LED

- HDD2 Status LED

LED Description

9 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Cool-Central® Liquid Cooling Technology

Efficient cooling

The liquid cooling is a direct-to-chip hot water (40 °C /105 °F), based

cooling solution that captures up to 70 % of the PRIMERGY CX600

server heat

Removes heat from Xeon Phi CPU and voltage-regulator-module

(VRM) within the server, eliminating the need for chilling to cool these

components

Cool-Central® Liquid Cooling Technology is a reliable and effective

solution to relieve the burden on data center cooling

Helps to reduce data center cooling costs by over 50 %

Less air needs to be cooled and moved

Liquid cooled systems need less airflow

Allows for 2.5-5x higher data center server density

Low pressure in rack circuit eliminates leakage risks

Returning water is warm enough to enable waste heat recycling

10 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Cool-Central® LCT: How it Works

Heat exchanger

Liquid-to-liquid heat

exchangers transfer heat

between facilities and

servers liquid loop.

Facilities and server

liquids are kept separate

and never mix.

Warm water from facilities dry

cooler or cooling tower enters

RackCDU, hotter water returns

Tubes move cooling liquid to and from

RackCDU to servers

Pump/cold plate units

atop CPUs, GPUs and

RAM circulate liquid

through server and

RackCDU, collecting heat

and returning to

RackCDU for exchange

with facilities liquid

Leak Detection Panel

(not shown)

Coolant reservoir and control

Facility water requirement

Maximum water temperature

Facility supply: 40 °C

Facility Return: 59 °C

Water pressure

Max: 125 PSI / 8.6 bar

Min: 10 PSI / 0.7 bar

RackCDU (Coolant Distribution Unit)

11 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Cool-Central® LCT: How it Works

Server loops and tubes

Fully tested, filled and ready to install

IT staff never has to handle liquids

Low pressure, factory sealed design eliminates risk of leaks

Dual in-series pumps

Provide redundancy

Pumps / cold plate units replace air heat

sinks and circulate cooling liquid

Quick Connectors

100% dripless

From / to RackCDU

Low pressure

Memory Coolers

For 8 or 16 DIMMs

Wide range of compatible

components

Intel Xeon and Xeon Phi

Nvidia Tesla, Grid

Memory modules

Cooling set for server node

12 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

PRIMERGY CX600 M1 – Usage Scenarios

Positioning

Target market

University / Academic for scientific research

Market expansion

Business intelligence / Analytics (BA)

Graphic analysis (rendering)

HPC expertise leverage

Product development / optimization (ISVs like LS-DYNA,

ANSYS-FLUENT, etc.)

Usage

HPC Market

SMALL MID HIGH

General purpose HPC solution

Hybrid parallel / throughput

oriented solution

Xeon CPU

Xeon Phi

13 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS

Latest News – Order for KNL System for T2K

Tokyo, May 10, 2016

Fujitsu today announced that it has received an order for a many-core large-scale

supercomputer system from the University of Tokyo and the University of Tsukuba.

The system will be deployed to the Joint Center for Advanced High-Performance

Computing (JCAHPC), which the two universities jointly operate.

The new supercomputer will be an x86 cluster system consisting of 8,208 of the

latest FUJITSU Server PRIMERGY x86 servers. These will run on the next-

generation Intel® Xeon Phi™ processors (Intel development code name: Knights

Landing), and achieve a theoretical aggregate performance of 25 PFLOPS.

The system is due to be completely operational starting December 2016, when it is

expected to be Japan's highest-performance supercomputer.

14 © 2016 FUJITSU CONFIDENTIAL MATERIAL / RESTRICTED ACCESS