MAPLD 2005/254 C. Papachristou 1 Reconfigurable and Evolvable Hardware Fabric Chris Papachristou, Frank Wolff Robert Ewing Electrical Engineering & Computer Science AFRL/AFTA, Bldg 620 Case Western Reserve University Wright Patterson Cleveland, Ohio 44106 WPAFB, OH 45433 September 7, 2005
24
Embed
MAPLD 2005/254C. Papachristou 1 Reconfigurable and Evolvable Hardware Fabric Chris Papachristou, Frank Wolff Robert Ewing Electrical Engineering & Computer.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
• Communication Protocol adaptation: - adapt to changing communication protocols for each situation
Reconfigurable Hardware: enabling technology to meet these requirements.
MAPLD 2005/254C. Papachristou 7
Sensor Web Scenario
Comm Tradeoffs Bandwidth = Buffer/Latency
Data Rate, Protocol, ErrorBit Rate.
MAPLD 2005/254C. Papachristou 8
Approach
Architecture: reconfigurable at four Layers:
Layer 4: the Adaptation Manager.
Layer 3: the Real-Time Operating System RTOS.
Layer 2: the Embedded Processors and Memory.
Layer 1: the Reconfigurable Hardware Fabric.
MAPLD 2005/254C. Papachristou 9
Architecture: Non Traditional Reconfigurable
MAPLD 2005/254C. Papachristou 10
Occurs at several levels:
(a) Selection of application modules by the Adaptation Manager.
(b) Mapping of modules into the hardware fabric or the embedded processors, depending on performance requirements.
(c) Configuration of the hardware fabric and the embedded processor to meet performance and data delivery requirements.
The reconfigurable hardware is essential for mapping of wireless communications algorithms such as : IR filtering, multichannel CDMA, complex encoding, advanced imaging.
Reconfiguration Strategy
MAPLD 2005/254C. Papachristou 11
Self Adaptation - Dynamic Configuration
MAPLD 2005/254C. Papachristou 12
Reconfigurable Fabric
MAPLD 2005/254C. Papachristou 13
Reconfigurable Tile
MAPLD 2005/254C. Papachristou 14
Core Switch Matrix
MAPLD 2005/254C. Papachristou 15
Double Buffer Configuration Switch Cell
MAPLD 2005/254C. Papachristou 16
is capable of on-line adaptation by autonomously reconfiguring its architecture either through software or by directly morphing the hardware.
The key idea is to achieve evolution in the hardware by evolving configuration candidates via a neural network and testing them for fitness.
A best fit configuration will morph the hardware to best responding to a particular input stimulus.
Evolvable Hardware
MAPLD 2005/254C. Papachristou 17
Evolvable Platform
MAPLD 2005/254C. Papachristou 18
Operation mode: NN generates configuration code
Training mode: NN incrementally evolves configurations by training itself on input stimuli as well as configuration data that are recurrently applied after being improved by genetic operations.
Other evolution modes e.g. self-diagnosis and self repair are also feasible.
Evolution modes:
MAPLD 2005/254C. Papachristou 19
During training, candidate configurations are selected from a population via genetic operations.
Training continues until a candidate passes a fitness test depending on responses from the hardware fabric.
Training may start on command or autonomously, in new environment, new functions or upgrading for better performance.
A major aspect of this scheme is to design a robust training mechanism for configuration evolution of the dynamic hardware fabric.
Training
MAPLD 2005/254C. Papachristou 20
Evolvable Hardware Training
MAPLD 2005/254C. Papachristou 21
Configuration Tools
Binding Configurator
Algorithm Data Flow
Arch Resource Netlist
Connectivity Bindings
Architecture Mapper Synthesis tools
MAPLD 2005/254C. Papachristou 22
Configuration Tools (Cont.)
Synthesis: Data Flow transormation of the application into a resource graph.
Binding: allocation of resources into configurable modules, This involves functional, local memories and interconnect modules.
Configuration Core: compact description of the mapping -- in space and time
Loading the configuration matrix into Buffere FIFOs to employ the mapping.
MAPLD 2005/254C. Papachristou 23
Results on Some Benches
Application AllocatedMemories
OperatorUnits
Size of 2 CoreMatrices
Deq 5 5 (10X10) and (5X5)
Bandpass Filter 9 9 (18X18) and (9X9)
Cosine Filter 11 11 (22X22) and (11X11)
Elliptical Filter 8 8 (16X16) and (8X8)
Arfilter 9 9 (18X18) and (9X9)
Wave Filter 6 6 (12X12) and (6X6)
DCT 12 12 (24X24) and (12X12)
MAPLD 2005/254C. Papachristou 24
Proof of Concept
For proof of concept, we will employ advanced FPGA boards from Xilinx and Altera, as well as CAD tools that we have obtained from commercial vendors.
We will develop an advanced prototyping environment based on these tools and software.
We will implement by emulation our proposed reconfigurable hardware on these boards, without actual chip design. Emulation and prototyping is quite feasible.