heterogeneous exascale computing

© Copyright 2021 XilinxCopyright 2021 Xilinx©

Heterogeneous Exascale Computing

Ivo Bolsens, CTO

March 25, 2021

1


Digitalization

Towards Exascale Computing

Compute

P C E R A

2000 2010 2020

EXA

PETA

TERA

ASIC

RTL developers

Democratize hardware designCloud

Accelerator

Hardware savvy SW developers

HW/SW co-design

Virtex

Peer

Data scientists

Machine Learning

AI

Zynq

Alveo

Zen

Radeon

Versal

SCALAR VECTOR MATRIX SPATIAL

Mobile

Internet


3

Historic: CPU OnlyTop of Rack Switch (TOR)


4

Scale-Up: CPU Host With GPU and FPGA Accelerators


5

Scale-Out: CPU Host With GPU, FPGA and SmartNIC

HPC Ethernet

(Cray)


6

Scale-Up: Heterogeneous Fabric of Peers (CPU, GPU, FPGA, SSD)

SmartSSD

Communication

bottleneck through CPU

Communication

between peers


Communication

bottleneck through CPU

7

Scale-up: Heterogeneous fabric of Peers + Smart SSD

Accelerator

memory

Customizable

accelerator

5th gen

Samsung

VNAND


8

Scale-out : Compute On Data-In-Use, Data-In-Motion, Data-At-Rest

Data-In-Motion

Data-In-Use Data-At-Rest


Scale-Out : Disaggregation of Compute And Storage

1

1

2

9

1

2

FPGA# =Workload

CPU

GPU

Storage

NIC


10

Fabric

Data “In-Use” In Distributed Memory

Fabric


11

Communication using message passing

Software view : Local memory & message passing


12

Memory

Fabric

Future Programmer’s View: Unified Memory

Unified Distributed System Memory

Memory

Fabric

Any To Any Shared Memory Fabric


Memory

Fabric

Future: Shared Virtual Address Spaces

Memory

Fabric

13

Virtual Resource Group A

Virtual Resource Group B


14

Software Abstraction for Exascale Compute

Uniform System Memory

Any To Any Memory Fabric

Virtualized Address Spaces


OS

S h e l l

VITIS Software StackEnables Smart NIC|SSD + Heterogeneous Compute

SMART

NIC

F R A M E W OR K S

L I B R A R I E S

V++ dev ice code + XRT API

X R T

SMART

SSD

VIRTUALIZATION

ORCHESTRATION

COMPUTE

AIE

COMPUTE

PL


Conclusions

16


17

Future

Exascale computing and data - driven by new workloads such as AI

Increase compute power by

Scale-up of compute nodes

Scale-out of network

Enabled by Accelerators, SmartNIC and SmartSSD

Compute on data-in-use

Compute on data-in-motion

Compute on data-at-rest

Software stack :

Higher level of programming abstraction

Runtime enabling heterogeneous compute fabrics

Shared virtual address spaces

heterogeneous exascale computing

Documents