Upload
others
View
31
Download
0
Embed Size (px)
Citation preview
© Copyright 2019 Xilinx
How FPGAs Have Evolved to
Address Compute Acceleration
and Interconnects
Ken Chang
Xilinx at OFC Booth #1811
© Copyright 2019 Xilinx
Agenda
˃Why FPGA, Adaptable Compute?
˃FPGA to Adaptable Compute Evolution
˃FPGA Transceiver Evolution
˃Summary
>> 2
© Copyright 2019 Xilinx>> 3
Compute and Bandwidth Needs Across All Industries
>> 3
CLOUD EDGENETWORK
AI ADOPTION ACROSS MARKETS
EXPONENTIAL DATA BANDWIDTH GROWTH
© Copyright 2019 Xilinx
The Slowdown of Moore’s Law and Technology Scaling
Processing Architectures are Not Scaling
>> 4
10
100
1000
10,000
100,000 ?
2X / 3.5 Years
2X / 1.5 Years
2X / 6 Years
2X / 3.5 Years
1980 1985 1990 1995 2000 2005 2010 2015
Performancevs. VAX11-780
CISC
RISC
End of DennardScaling
AmdahlsLaw
40 YEARS OF PROCESSOR PERFORMANCE
Source: John Hennessy and David Patterson, Computer Architecture: A Quantitative Approach, 6/e 2018
© Copyright 2019 Xilinx
FPGA Evolution: From FPGA to SoC to ACAP
>> 5
First FPGA Introduced
First VirtexFPGA
Virtex-2 Pro
First 3D FPGA &
HW/SW
Programmable SoC
First MPSoC& RFSoC
ACAP
1985
2019
1998
2001
2011
2016
Moore’s Law Era SoC Era
© Copyright 2019 Xilinx>> 6
˃ Based on 7nm FinFET
˃ Platform Solution for Compute /
Storage / Network Acceleration
˃ IP Subsystems and a Network-on-Chip
˃ Programmable AI Engine
>> 6
Programmable Logic
GPIO & Mem I/F
SerDes HBMRFDAC/
ADC
Application Processors
Real-Time Processors
AI Engine
Next Generation Programmable
Logic
Adaptive Compute Acceleration Platform
© Copyright 2019 Xilinx
MLInference
5GWireless
Power
7nm ACAP 16nm FPGA
AI Engine
˃ Target Applications
ML Inference for Cloud DC
5G wireless: Radio, Baseband
ADAS/AD Embedded Vision
˃ Each Tile:
ISA-based Vector Processor
(SW programmable, C++, e.g.)
Local Memory
Adaptable, Non-blocking
Interconnect
>> 7
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Core
Me
mo
ry
Array Architecture
Source: Juanjo Noguera, “HW/SW Programmable Engine: Increased Compute Density Architecture for Project Everest”, Hot Chips 2018
10x
40%Less
Power
5x
© Copyright 2019 Xilinx
Industry’s First ACAP: Versal Architecture Overview
>> 8
Network On Chip
Versal
Adaptable
Hardware
Scalar Engines Adaptable Engines Intelligent Engines
Arm Dual-CoreCortex-R5Real-TimeProcessor
Arm Dual-CoreCortex-A72ApplicationProcessor
AI
Engines
DSP
Engines
28G
58G
112GPCIe &
CCIX
(w/DMA)
DDR HBMMultirate
Ethernet
600G
Cores
Direct
RF
MIPI
LVDS
GPIO
Block RAM
UltraRAM
Accelerator RAM
PlatformManagement
Controller
Processing System
7nm FinFET
© Copyright 2019 Xilinx
ACAP Meets Compute Demands for Cloud, Network, and Edge
>> 9
CLOUD EDGENETWORK
16nmFPGA
7nmACAP
Go
og
leN
et
V1
Im
g/S
ec
(<2m
s)
10X
16nmRFSoC
7nmACAP
Int
16
x1
6 D
SP
Co
mp
ute
(Tera
MA
C/
sec)
5X
16nmFPGA
7nmACAP
Sin
gle
-Chip
Encry
pte
d
Tra
ffic
(G
b/s
)
4X
16nmMPSoC
7nmACAP
ResN
et5
0 im
g/s
ec
(ba
tch
=1
)
15X
© Copyright 2019 Xilinx
1
10
100
0102030405060708090100110120
Dat
a Ra
te [G
bps]
Process Node [nm]
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
Backplane Transceiver Trend from Industry
>> 11Data source: ISSCC, VLSI Circuit Symposium
Xilinx
Backplane channel loss: > 25dB
© Copyright 2019 Xilinx>> 12
Backplane Transceiver Trend: Date Rate vs Year
>> 12
1
10
100
2006 2008 2010 2012 2014 2016 2018 2020
Data
Rate
[G
bp
s]
Year
56G PAM4
RX
100G FEC
56G PAM4
RX
First 28G LR paper (IBM)ADC based
© Copyright 2019 Xilinx>> 13
Backplane Roadmap (From 2012)
˃ Courtesy of E. Wu, “FPGA Optical Connectivity,” OFC Workshop 2012
© Copyright 2019 Xilinx>> 14
Backplane Transceiver Trend: Date Rate vs Year
>> 14
1
10
100
2006 2008 2010 2012 2014 2016 2018 2020
Dat
a R
ate
(Gb
/s)
Year
DR [Gb/s]
Roadmap
© Copyright 2019 Xilinx>> 15
Electrical vs Optical
>> 15
Ashok Krishnamoorthy, “Optical interconnects in computing and
switching systems: the anatomy of a 20Tbps switch card”, ISSCC 2018
Optical for 200G?
© Copyright 2019 Xilinx
On-Board Optics Solution
˃ Capable of error free (BER < 10-12) for compute application
˃ Electrical portion (16nm FinFET) based on CEI-56G-NRZ Published in ISSCC 2018
˃ Optical portion based on EAM silicon photonics with Driver/TIA in 16nm FinFET
Submitted to VLSI 2019
>> 16
XSR Example: 50mm, 8dB
Optical EngineASIC / FPGA Fiber1.5dB
PCB
Total Power
per Channel
mW pJ/bit
126 2.25
© Copyright 2019 Xilinx>> 17
Optical …
>> 17
1.00E-12
1.00E-11
1.00E-10
1.00E-09
1.00E-08
1.00E-07
1.00E-06
1.00E-05
1.00E-04
1.00E-03
4 6 8 10 12 14 16 18 20 22 24 26
BER
PI Code
Error free at 50Gb/s!
Measured Bathtub 50Gbps
Tx7 Rx4 No EDFA Ext Loop
Xilinx OFC Booth #1811
© Copyright 2019 Xilinx>> 18
Transceiver Evolution: the “On Package” Stage
˃ Front Panel Pluggable
˃Mid-Board Optics
˃On Package, Electrically Pluggable
˃On Package, Optically Pluggable
Slide courtesy of
© Copyright 2019 Xilinx>> 19
Summary
>> 19
Evolution of Compute BandwidthSurpassed Moore’s Law with New
Architecture
Time
Limits of Moore’s Law
Evolution of Serial I/O BandwidthSurpass Electrical Limits with New
Architecture?
Limits of Electrical Signaling
Compute
(TOPS)
Time
I/O
BW/Package
FPGA ACAP re-architecture
To meet expectations of growth
Optics?
Lo
g s
ca
le
Lo
g s
ca
le
© Copyright 2019 Xilinx
Acknowledgments
˃ The speaker would like to thank the following people for their valuable
contributions and discussions for the presentationAnthony Torza, Ehab Mohsen, Kees Vissers, Vamsi Boppana, Gaurav Singh, Yohan Frans, Geoff Zhang, Ephrem Wu, Liam Madden
>> 20