A prototype parallel multi-FPGA accelerator for SPICE CMOS model evaluation
A prototype parallel multi-FPGA accelerator for SPICE CMOS model evaluation
Due to ever increasing complexity of circuits, EDA tools and algorithms are demanding more computational power. This made transistor-level simulation a growing bottleneck in the circuit development process. This thesis serves as a proof of concept to evaluate and quantify the cost of using multi-FPGA systems in SPICE-like simulations in terms of acceleration, throughput, area, and power. To this end, a multi-FPGA architecture is designed to exploit the inherent parallelism in the device model evaluation phase within the SPICE simulator. A code transformation flow which converts the high-level device model code to structural VHDL was also implemented. This flow showed that an automatic compiler system to design, map, and optimise SPICE-like simulations on FPGAs is feasible.
This thesis has two main contributions. The first contribution is the multi-FPGA accelerator of the device model evaluation which demonstrated speedup of 10 times over a conventional processor, while consuming six times less power. Results also showed that it is feasible to describe and optimise FPGA pipelined implementations to exploit other class of applications similar to the SPICE device model evaluation. The constant throughput of the pipelined architecture is one of the main factors for the FPGA accelerator to outperform conventional processors. The second contribution lies in the use of multi-FPGA synthesis to optimise the inter-FPGA connections through altering the process of mapping partitions to FPGA devices. A novel technique is introduced which reduces the inter-FPGA connections by an average of 18%.
The speedup and power efficiency results showed that the proposed multi-FPGA system can be used by the SPICE community to accelerate the transistor-level simulation. The experimental results also showed that it is worthwhile continuing this research further to explore the use of FPGAs to accelerate other EDA tools.
Maache, Ahmed
252935fa-408b-4ce5-ab26-720ced6bf7d7
January 2011
Maache, Ahmed
252935fa-408b-4ce5-ab26-720ced6bf7d7
Reeve, Jeffrey
dd909010-7d44-44ea-83fe-a09e4d492618
Zwolinski, Mark
adfcb8e7-877f-4bd7-9b55-7553b6cb3ea0
Maache, Ahmed
(2011)
A prototype parallel multi-FPGA accelerator for SPICE CMOS model evaluation.
University of Southampton, School of Electronics and Computer Science, Doctoral Thesis, 206pp.
Record type:
Thesis
(Doctoral)
Abstract
Due to ever increasing complexity of circuits, EDA tools and algorithms are demanding more computational power. This made transistor-level simulation a growing bottleneck in the circuit development process. This thesis serves as a proof of concept to evaluate and quantify the cost of using multi-FPGA systems in SPICE-like simulations in terms of acceleration, throughput, area, and power. To this end, a multi-FPGA architecture is designed to exploit the inherent parallelism in the device model evaluation phase within the SPICE simulator. A code transformation flow which converts the high-level device model code to structural VHDL was also implemented. This flow showed that an automatic compiler system to design, map, and optimise SPICE-like simulations on FPGAs is feasible.
This thesis has two main contributions. The first contribution is the multi-FPGA accelerator of the device model evaluation which demonstrated speedup of 10 times over a conventional processor, while consuming six times less power. Results also showed that it is feasible to describe and optimise FPGA pipelined implementations to exploit other class of applications similar to the SPICE device model evaluation. The constant throughput of the pipelined architecture is one of the main factors for the FPGA accelerator to outperform conventional processors. The second contribution lies in the use of multi-FPGA synthesis to optimise the inter-FPGA connections through altering the process of mapping partitions to FPGA devices. A novel technique is introduced which reduces the inter-FPGA connections by an average of 18%.
The speedup and power efficiency results showed that the proposed multi-FPGA system can be used by the SPICE community to accelerate the transistor-level simulation. The experimental results also showed that it is worthwhile continuing this research further to explore the use of FPGAs to accelerate other EDA tools.
Text
A.MAACHE.Thesis.pdf
- Other
More information
Published date: January 2011
Organisations:
University of Southampton
Identifiers
Local EPrints ID: 173435
URI: http://eprints.soton.ac.uk/id/eprint/173435
PURE UUID: 0de0d392-4e4f-479e-9b91-003f3c43035a
Catalogue record
Date deposited: 20 May 2011 10:20
Last modified: 14 Mar 2024 02:33
Export record
Contributors
Author:
Ahmed Maache
Thesis advisor:
Jeffrey Reeve
Thesis advisor:
Mark Zwolinski
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics