Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
| Applied Materials Confidential External Use
NAVIGATING THE PERFECT STORMENABLING THE A.I. ERA
Gary DickersonPresident and CEO, Applied Materials
| Applied Materials Confidential External Use
“Data is to this century what oil was to the last one: a driver of growth and change”– The Economist
| Applied Materials Confidential External Use
CO
NN
ECT
IVIT
Y
DATA GENERATION IoT and Industry 4.0 driving an explosion of data
DATA STORAGE
More data needs to be processed and stored –Storage alone is not sufficient or economical
COMPUTE
New compute models to turn data into value
New compute architectures to process data at edge and in cloud at right performance / watt
External Use
| Applied Materials Confidential External Use
2017 2018E 2022E
HOME and BUILDINGS
HUMANS
INFLECTION YEARData generated by
machines > humans
SMART FACTORIES
SMARTVEHICLES
HumansMachines
1.5ZB2ZB
>10ZB
53% 44% <10%
SAFETY and SECURITY
SOURCE: Applied Materials model based on forecasts published by Cisco, Intel, Western Digital
Explosion of Data Generation
| Applied Materials Confidential External Use
ACCELERATIONOF A.I. FUELED BY
MACHINE LEARNING relentless classification of data to make determinations or predictions
McCarthy & Minsky’s
INITIAL A.I. RESEARCH PAPER
1955 TODAY
Yann LeCun’s
CONCEPT OF CONVOLUTION
Quoc Le’s
IDEA OFBRUTE FORCE
1982 2012
1. Very large, accessible data sets
2. Affordable high performance computing to turn data to $$$
SOURCE: Historic references based on New Street Research, May 2018
| Applied Materials Confidential External Use
CPU TPU / GPU
DRAM
ALU
CONTROL
CACHE
ALU
DRAM
ALUALU
Sequential computing
General purpose
Good at many things
Parallel computing
Designed for throughput
Great at specialized tasks
A lot of memory(because there’s a lot of data)
Parallel computing(for throughput)
Extremely high logic memorybandwidth
A.I. WORKLOADS NEED
| Applied Materials Confidential External Use
A.I. – BIG DATA DRIVING A
RENAISSANCE OF HARDWAREDEVELOPMENT AND INVESTMENT
TPUGPUASICsFPGAs
DDRHBM (High bandwidth memory)
Flash
MRAMReRAM/CeRAMPRAMFeRAM
AnalogMemristorReRAM/PCM
QuantumSynaptic
ACCELERATORS
NEAR MEMORIES
NEW MEMORIES
IN MEMORY COMPUTE
NOVEL HPC
| Applied Materials Confidential External Use
IMPROVEMENT IN COMPUTE PERFORMANCE / WATT NEEDED
SOURCE: DARPA, Intel, NVIDIA, A.I. startups
| Applied Materials Confidential External Use
THE PERFECT STORM
THE PERFECT OPPORTUNITY?
A.I. NEEDS EDGE AND CLOUD INNOVATIONS…
HPCorders of magnitude
improvement in performance,energy efficiency and cost
STORAGEabundant low cost,high performance,
low power data storage
MOORE’S LAW CHALLENGEDas classic 2D feature shrink slows
WHILE AT THE SAME TIME…
OR
| Applied Materials Confidential External Use
SOURCE: Electronics, Volme 38, Number 8, April 19, 1965
1965: MOORE’S THESIS BASED ON FIVE DATA POINTS~1975: ESTIMATE UPDATED TO ‘DOUBLING EVERY 2 YEARS’
FIGURE 1 FIGURE 2
| Applied Materials Confidential External Use
CLASSIC 2D FEATURE SCALING SLOWING
10,000,000,000
1,000,000,000
100,000,000
10,000,000
1,000,000
100,000
10,000
1,0001970 1975 1980 1985 1990 1995 2000 2005 2010 2015
4004
80088080
8085
8086 8088
80286
83286
80486
PentiumPentium Pro
Pentium MMX
Pentium III
Pentium 4Pentium M
Nehalem
Sandy Bridge
Ivy
6800
68000
68020
68030
68040
601 603e
604e604
G3
G4
G5
G6
Projection Held For 40 Years…
SOURCE: University of Wisconsin
TRA
NSI
STO
RSHaswell
Broadwell
Skylake
Recent data points suggest~2x more every 5 years
2020
~2x more every 2 years(1.413)2 more every 2 years
1970 - 201040 1,000,000 1.413
| Applied Materials Confidential External Use
PERFORMANCE IMPROVEMENTS OVER TIME (VS. VAX-11/780)
100,000
10,000
1,000
100
10
1978
1986
200
3
2011
2015
2018
25%per year
52%per year
23%per year
12%per year
3.5%per year
End of Dennard Scaling
Limits of parallelism of Amdahl’s Law
End of Moore’s Law
VAX-11/780
TIME BETWEEN LOGICNODES IN YEARS
>4
2.5
2
1.8
14 to 10nm
22 to 14nm
32 to 22nm
45 to 32nm
PERFORMANCE IMPROVEMENTS SLOWING
SOURCE: Computer Architecture: A Quantitative Approach, Sixth Edition, John Hennessy and David Patterson, December 2017 SOURCE: Bernstein
| Applied Materials Confidential External Use
PPAC
105
104
103
102
10
1
In the Past…
COMPONENT PER IC
REL
AT
IVE
CO
ST/C
OM
PON
ENT
1 10 102 103 104 105
1962
1965
1970
“Classic” 2D feature shrinking
materials engineering to improve power and performance
ENABLED BY
PPAC
POWER
PERFORMANCE
AREA-COST
MOORE’S LAW
| Applied Materials Confidential External Use
FOUNDATION IS MATERIALS ENGINEERING
In the Future…
ENABLED BY
Advanced packaging
New structures / 3D
New ways to shrink
New architectures
New materialsPPAC
POWER
PERFORMANCE
AREA-COST
| Applied Materials Confidential External Use
NUMBER OF CRITICAL ELECTRONS IN NAND CELL
0
20
40
60
80
100
102030405060
1st
GEN2nd
GEN3RD
GEN
2D NAND
nm
2D NAND“Classic Scaling”
3D NANDNew architecture
enabled by materials
3D NANDshows power of architecture inflections
Extended NAND roadmap by >10 years + better device
Speed
Endurance
Power efficiency
2x10x
2x
3D NAND
cost
| Applied Materials Confidential External Use
Future 2D shrink is not only limited by resolution, but also
PLACEMENT ERRORS
State of the art A.I. chip can have up to
100B vias
Correctlylanded viaPlacement
error
Can be addressed by self-aligned structures
| Applied Materials Confidential External Use
MATERIALS-BASED APPROACHES CAN ELIMINATE PLACEMENT ERRORS
Example: ‘Multicolor’ = Fully self-aligned multi-material patterning
STEPS1 - 8
STEPS9 - 10
STEPS11 - 14
Material A
Material C
Material B
Material D
Material E
Self-aligned Spacer Patterning:Litho-spacer Etch
Materials Engineering:Gapfill-CMP
Deposition-Lithography-Etch:Large Overlay Tolerance
| Applied Materials Confidential External Use
ADVANCED PACKAGINGCan Optimize System Level Performance
DRAM ON PCB to STACKED DRAM IN PACKAGE HETEROGENEOUS INTEGRATION
System on Chip to System on Package
Integration of chiplets provides time, cost and yield benefits
GR
APH
ICS
/ IM
AG
ING
10nm
OT
HER
IP22
nm
CPU CORES10nm
COMMS14nm
I/O 14nm
Power savings per bit
3x
50%
Logic DRAM bandwidth performance
Substrate (PCB)
DRAM (stacked) GPU
Si Interposer
SOURCES: Intel, GLOBALFOUNDRIES
SOURCES: AMD, NVIDIA
STA
CK
ED
DR
AM
STA
CK
ED
DR
AM
STA
CK
ED
DR
AM
STA
CK
ED
DR
AM
GPU
| Applied Materials Confidential External Use
P PAC
POWER
PERFORMANCE
AREA-COST
New architectures
New ways to shrink
New materials
Advanced packaging
New structures / 3D NEW PLAYBOOKNEEDED FOR
CONNECTIVITY + SPEED
KEY ISSUES =Complexity
Integration challenges
Time to market
| Applied Materials Confidential External Use
CONNECTIVITY TO ACCELERATE INNOVATION
TODAY: Serial / compartmentalized interaction between key parts of eco-system
“Von Neumann” mindset “Neuromorphic” mindsetD
ESIG
N
INT
EGRA
TIO
N
EDA
MA
NU
FAC
TU
RIN
G
EQU
IPM
ENT
MA
TER
IALS
OPPORTUNITY: Parallel development to get powerful tools to designers faster
DESIGN
INTEGRATION
MANUFACTURINGEDA
EQUIPMENT
MATERIALS
vs.
| Applied Materials Confidential External Use
A.I. – Big Data Era = the biggest opportunity of our lifetimes
Hardware renaissance
Materials innovation to enable new architectures, structures, ways to shrink and packaging approaches
New eco-system playbook to drive connectivity and speed
UNLOCKED BY