fpga 2032 roadmap : a personal perspectivexilinx.eetrend.com/files-eetrend-xilinx/forum/... ·...
TRANSCRIPT
FPGA 2032 Roadmap :
A Personal Perspective
Page 1
Blackberry
2012 1992
Learn from Yesterday
Courtesy : Don McMillan
Cell Phone
2012 1972
Learn from Yesterday
Courtesy : Don McMillan
Silicon Roadmap
Courtesy : IMEC
Copyright 2012 Xilinx
Interconnect
NSN 5
15x drop in I/O-to-logic ratio
by 2020
Source: ITRS
• Growing gap between number of logic gates and I/O • Technology scaling favors logic density
Copyright 2012 Xilinx
What Does 2.5D and 3D Buy Us?
Page 6
Connectivity
Capacity
Crossovers
RAM
Logic
Package Substrate
Copyright 2012 Xilinx
3D versus 2.5D
3D 2.5D
Design Flow New Co-Design Evolutionary
Device impact Stress None
Cost High 65nm Interposer
Thermal Challenging Evolutionary
Testing New Methods Evolutionary
Reliability Challenging Evolutionary
Copyright 2012 Xilinx
2.5D and 3D Interconnect
Decoupling Capacitances
Integrated Passives
ESD
Memory
Photonics IO
IO devices
Interconnect
Time
Cost
Smart Interposer
Copyright 2012 Xilinx
Optical Interconnect
2012 2015 2020
ASIC Use Model
System Verify
System Integrate
System Design
Legacy Design Port
and Re-verify
3rd party IP re-use
(including Xilinx IP)
New IP Design
New IP Verify
Copyright 2012 Xilinx
Model of FPGA HW customer design effort
Design effort
(man-years)
Note: maximum 85% device utilization is assumed % design re-use (fabric only)
1.3 M LCs
2.6 M LCs
150 man-years ($30 million)
to complete the design of large
FPGA with 60% reuse
650K LCs
330K LCs
Copyright 2012 Xilinx
SW Use Model : FPGA and Multi-core ASSP
12
Applications
Quality Metric (Performance/
Power/Area)
ASSP
FPGA
Copyright 2012 Xilinx
Need for Design Tools
13
Hour Day Week Month
0.25
1
Year
4
16
64
256
Initial Design
Relative
Performance
Design-time
CPU
GPU
FPGA
Gap
Courtesy : David Thomas
Copyright 2012 Xilinx
The Programmable Processing Platform A heterogeneous multicore
Application processors
– Hard core and soft core
– External and embedded
– Caches and large memory space
– Unified shared memory
– Full OS support
Streaming micro-engines
– Configurable (soft) vector cores
– Tiny memory footprint
– Many, distributed, memories
– Compute kernels, no OS
Fixed function datapaths
– C to Gates generated
– HDL coded
– Library IP component
DDR3 MemCon
Interconnect A
SMP CPU
X86 CPU DSP
High speed I/O
FPGAs provide a rich set of mapping options for complex algorithms and communication patterns
Discrete GPU
Micro- Engine Array
HW Datapaths
Interconnect B
FPGA
Copyright 2012 Xilinx
Parallel Programming
Page 15
FPGA
Vid
eo
co
de
c
En
cry
ptio
n
Pa
cke
t
Pro
ce
ssin
g
FF
T
Se
arc
h
Application-Specific
Open CL Software Model
ARM Processor
A9
A9
Automated Hardware/software
partitioning & interfacing Unified
Compiler/Linker
Application Level Programming of Heterogeneous Multi-core
Copyright 2012 Xilinx
The Design Stack
Confidential
Design Definition
Design Implementation
Time
Design Definition
Focus R&D on Core Competencies
In-house In-house In-house
Design Definition
Design Implementation
Design Methodology
Design Manufacturing EDA
Design Methodology EDA
Design Methodology
Foundry
Design Manufacturing
FPGA Platform
Design Implementation
Foundry
Design Manufacturing
Copyright 2012 Xilinx
Conclusions : 2032 FPGA
Page 17
• Silicon Technology Roadmap continues
Challenges : Variability, Fault tolerance, Power limited, Cost
• 2.5 D and 3 D essential
Opportunity : Connectivity, Capacity, Heterogeneity
• HW Design Methods handle complexity
- Re-Use
- Modular Design Flow/Architecture + 10 min 1M LC P&R
- 33% utilization : dark silicon/fault tolerance
• Software Use Model
- Parallel programming
- Heterogeneous Multi-core
- Solve Timing Closure Copyright 2012 Xilinx
The Singularity
Page 18
10 Transistors = 10 Neurons
Parallel Architecture
Synapse : 10 denser interconnect
Fault tolerant
Asynchronous
Adaptable
Content Addressable Memory
10 more energy efficient
10 10
3
7
Copyright 2012 Xilinx