Aurora
System and Development Kit
Disclaimer
This presentation has been prepared by Eurotech S.p.A. (or “Eurotech”) and has to be read in conjunction with its oral
presentation.
The information contained in this presentation does nor purport to be comprehensive. Neither Eurotech nor any of its officers,
employees, advisers or agents accepts any responsibility for/or makes any representation or warranty, express or implied, as to the
truth, fullness, accuracy or completeness of the information in this presentation (or whether any information has been omitte d
from the presentation) or any other information relating to Eurotech, its subsidiaries or associated companies, whether written,
oral or in a visual or electric form, transmitted or made available.
This document is confidential and is being provided to you solely for your information and may not be reproduced, further
distributed to any other person or published, in whole or in part, for any purpose.
The distribution of this document in other jurisdictions may be restricted by law, and persons into whose possession this document
comes should inform themselves about, and observe, any such restrictions.
This document is directed only at relevant persons. Other persons should not act or rely on this document or any of its cont ents.
No reliance may be placed for any purposes whatsoever on the information contained in this document or any other material
discussed during this presentation, or on its completeness, accuracy or fairness.
The information in this document and any other material discussed at this presentation is subject to verification, completion and
change.
The information and opinions contained in this document are provided as at the date of the presentation and are subject to ch ange
without notice.
Some of the information is still in draft form and will only be finalized.
By attending the presentation you agree to be bound by the foregoing terms.
Trademarks or Registered Trademarks are the property of their respective owners.
Eurotech HPC Product Roadmap
Three product families: adapted to all HPC users & workloads
Aurora
Next Gen
Aurora
Aurora
“Hi√e” DK
Aurora
“Hi√e”
“Hi√e”
Next Gen
G-Station
Cube
Aurora
Tigon
Tigon
Next Gen
Scalar Performance
Accelerated Performance
General Purpose
(CPU)
Hybrid
(CPU + GPU)
Accelerated
(GPU)
Eurotech HPC Product Roadmap
Three product families: adapted to all HPC users & workloads
Aurora
Next Gen
Aurora
Aurora
“Hi√e” DK
Aurora
“Hi√e”
“Hi√e”
Next Gen
G-Station
Cube
Aurora
Tigon
Tigon
Next Gen
Scalar Performance
Accelerated Performance
General Purpose
(CPU)
Hybrid
(CPU + GPU)
Accelerated
(GPU)
Aurora Hi√e – High Velocity
.
• Most dense architecture in the market
• More than 1 Pflop/s DP per rack with
NVIDIA® Tesla K80
• Highest energy efficiency.
• At system level > 5 GFlops / Watt
• At datacenter level with PUE of 1.05
• Distinctive flexibility. Direct hot water cooled
with no constraints for the choice of components:
• Intel or ARM CPU
• NVIDIA Tesla accelerators
• Intel coprocessors
• Any other PCIe card
• Direct hot water cooling of all components
• First and only system with multiple direct hot
water cooled GPGPUs
Aurora Hi√e
.
Markets and applications
Aurora Hi√e is designed for heavy workloads
that need acceleration with no compromises in
space and energy
Examples of target application segments:
• High energy physics (QCD)
• Oil & Gas – Seismic migration
• Life Sciences – Molecular Dynamics and
Computational biology
• Machine learning
• Big data analytics
• Cyber security and OSINT
• CAE – Computational Fluid Dynamics
• Media and Entertainment – Rendering
• Defense – Signal processing
Aurora Hi√e
Unparalleled!
Performance Savings
Space needed for 50 Pflop/s
Energy needed for 50 Pflop/s
Thiane2
700m
2
Hi√e
100m
2
Thiane2
24 MW
Hi√e
11 MW
Flexibility
CPU
Modules
System
GPU,
coprocessor,
NVMe +a any
PCIe card
Performance
Efficiency
Hi√e
CPU Only
Effectively lower time
to solution with
configurations
optimized for specific
workloads
Standard
Hybrid
8, 16, 64, 128
nodes per
system rack
Multiple racks
per installation
Nodes
CPU
Modules
Network
PCIe switch
AURORA HI√E ARCHITECTURE
Aurora Hi√e architecture: highlights
• Optimized for accelerated workflows
• Designed to maximize energy efficiency
and density
• Leverage PCIe cards from leading
vendors
• Software stack – standard programming
environment
• Road Map – in line with state-of-the-art
technology introductions
• Cooled with Aurora Hot Direct Water
Cooling
• Immediately available with the Aurora
Hi√e Systems
Accelerate with no compromises
CPU card
PCIe Switch
Water cooling
High Speed
Interconnects
Software Stack
PCIe
Submodules
CONFIDENTIAL
10
The Hi√e building block is a highly
modular integration of 6 PCIe cards
on a mid-plane with a PCIe switch
The main components of the Hi√e
architecture are:
• Mid-plane with the PCIe switch (6
x PCIe gen3 x16)
• CPU card for management and
control
• Network card or Peripheral device
or Computational Module 0
• Computational Module 1
• Computational Module 2
• Computational Module 3
• Computational Module 4
The computational modules are
any ×1, ×2, ×4, ×8, or ×16 PCIe
• Low Profile Cards, 2.536” (64.4mm)
• Half Length Cards, 6.6” (167.65mm)
• Full Length Cards, 12.283” (312mm)
Aurora Hi√e architecture
Aurora Hi√e systems
.
PCIe Modules
IB card
CPU PCIe cards
PCIe Switch
Hi√e Node
1 x Hi√e Node
n x Hi√e Nodes
Hi√e data center
Hi√e Dev Kit
CONFIDENTIAL
12
Aurora Hi√e systems
.
• 128 Nodes configuration ( a full rack):
o more than 1 Pflop/s DP per rack with
NVIDIA® Tesla ® K80
o more than 5 GFlops / Watt sustained
o more than 5 Pflop/s SP per rack with NVIDIA
Tesla M60
• Datacenter level PUE: 1.05
• Aurora Direct Water Cooling on all
components (water out at >55 °C)
• Stripped and essential: no unnecessary
components
• Support for different configurations: GPUs like
NVIDIA Tesla K40, K80, coprocessors like
Intel Phi,
• Designed for NVIDIA Tesla M60, NVIDIA
GeForce GTX 980M, AMD Firepro and
storage NVMe cards + any other PCIe card
The Aurora Hi√e
The ultimate accelerated computing solution
Specifications:
CONFIGURATION
• 1 CPUs per node
• 4 Accelerators per node
• 4 nodes per rack row
• Up to 128 nodes per rack = 512 GPUs or Phi per rack
SUB-MODULES
5 slots for the computational submodules fitting any ×1, ×2,
×4, ×8, or ×16 PCIe card (bootable device or an accelerator):
• Standard Height Cards, 4.20” (106.7mm)
• Low Profile Cards, 2.536” (64.4mm)
• Half Length Cards, 6.6” (167.65mm)
• Full Length Cards, 12.283” (312mm)
PERFORMANCE
• More than 1PFlop/s DP per rack (with NVIDIA® Tesla®
K80)
• 5 PFlop/s SP per rack with NVIDIA M60
CPU PER NODE
• 1 x E3-12xx v3
• 1 x Applied Micro XGene ARM 64-bit processor
ACCELERATORS PER NODE
• 4 x NVIDIA® Tesla® K40, K80
• 4 x Intel® Xeon Phi™ 7120x
• Designed for AMD Firepro, NVIDIA® Tesla M60, NVIDIA
GeForce GTX M980
MEMORY
• Up to 32 GB DDR3 (8GB per core)
• Soldered high reliability memory
INTERFACES PER NODE
• 2 x 1GigE (1 x 10 GigE on ARM version)
• 2 x USB
• 1 x VGA
• 2 x FDR Infiniband
LOCAL STORAGE PER NODE
• 1 x 256 GB, 512 GB, 1TB or 2TB SATA SSD
SUB MODULES SWITCH
• PCIe3 switch: PLX PEX8796
OPERATING SYSTEM
• Cent OS, RedHat or Suse
RAS
• Soldered memory, no fans, no hot spots
• Monitoring of system and cooling loop
• Hot swap nodes
• Eurotech ESS safety software
POWER AND COOLING
• Aurora Direct Hot Water Cooling
• Max power consumption 166 kW per fully loaded rack
Aurora Hi√e Development Kit
.
• The Dev Kit is the fastest and easiest
way to obtain and install the Hi√e
technology for
o A trial and test
o A production environment, should
your computational needs be
satisfied by a single node system
• The Kit packages together one complete
Hi√e server plus all components
necessary to power and hot water cool it
• The Hi√e Development kit comes with
preinstalled drivers and operating system
plus all software necessary to test and try
out
MEMORY
• Up to 32 GB DDR3 (8GB per processor core)
• Soldered high reliability memory
INTERFACES
• 2 x 1GigE (1 x 10 GigE on ARM version)
• 2 x USB
• 1 x VGA
• 2 x FDR Infiniband
LOCAL STORAGE
• 1 x 256 GB, 512 GB, 1TB or 2TB SATA SSD
• Special configuration with NVMe card(s)
SUB MODULES SWITCH
• PCIe3 switch: PLX PEX8796
OPERATING SYSTEM
• Linux Cent OS
The Aurora Hi√e development kit
The ultimate accelerated computing solution
Specifications:
CONFIGURATION
• 1 x CPUs + 4 x Accelerators
SUB-MODULES
5 slots for the computational submodules fitting any ×1, ×2,
×4, ×8, or ×16 PCIe card (bootable device or an accelerator):
• Standard Height Cards, 4.20” (106.7mm)
• Low Profile Cards, 2.536” (64.4mm)
• Half Length Cards, 6.6” (167.65mm)
• Full Length Cards, 12.283” (312mm)
PERFORMANCE
• 8 Tflop/s DP with NVIDIA ® K80
• 40 TFlop/s SP with NVIDIA M60
CPU
• 1 x E3-12xx v3
• 1 x Applied Micro XGene ARM 64-bit processor
ACCELERATORS
• 4 x NVIDIA Tesla® K40, K80
• 4 x Intel® Xeon Phi™ 7120x
• Plus; AMD Firepro , NVIDIA Tesla M60, NVIDIA GeForce®
GTX M980
Development Kit
One Hi√e unit + power supply + Cooling unit
Koolance Cooling
Unit
Hi√e Node
Power supply
I/O
VGA, USB, GigE,
Infiniband, Debug
Development Kit Block Diagram
Hi√e Node Power Supply Cooling Unit
Drivers
OS – Linux CentOS 7
XWindows
Compilers and
libraries
GCC, OpenMPI, Intel compiler
15.0, PGI compiler 14.10,
NVIDIA CUDA, Mellanox OFED
Benchmarks
HPL, XHPL, STREAM2
Test Applications - GROMACS, HOMD, LAMMPS, AMBER + any other user installed software
Performance
Analysis
LIKWID, Open|SpeedShop
Hardware
OS and Tools
User Applications
Development Kit Configuration
Acceleration for a wide variety
of applications
Configuration
• CPU – Intel E3 12xx v3/v4:
• Memory: 8GB per core
• Infiniband FDR (2 port), EDR
• Standard: 4 x PCIe NVIDIA
Tesla K40
• Additional configurations:
NVIDIA Tesla K80 and Intel
Phi
• Plus: NVIDIA Tesla M60,
AMD Firepro, NVIDIA
GeForce GTX M980 and
storage NVMe cards
/
