Transcript sangkyo_20071120.ppt

Leakage Power Analysis
of a 90nm FPGA
Published at IEEE
Custom Integrated Circuits Conference in 2003
Authors:
Presenter:
Tim Tuan (Xilinx),
Bocheng Lai (UCLA)
Sang-Kyo Han
(ECE, University of Maryland)
1
My Motivations for Paper Selection
I have an interest on FPGA power reduction.
I want to know the reference data about leakage power
of FPGA and the measurement method of it.
2
Contents
1.
2.
3.
4.
5.
Introduction
Background
Methodology
Result and Analysis
Conclusions
3
Introduction
FPGAs have become attractive implementation solutions.

Promising solutions for applications that require both performance
and flexibility
Motivations of the Research


To support reconfigurations, FPGAs use more transistors than
fixed-logic solutions, resulting in higher leakage power.
FPGA due to excessive leakage is unsuitable for mobile applications.
Methodology based on device-level simulations using
BSIM4 models

BSIM4 models include a number of leakage-modeling
improvements over prior versions.
Significant power reduction from analysis is possible.
4
Background – Existing Work
Leakage power has been extensively studied in ASICs
and microprocessors.
Focus on FPGA dynamic power consumption



E. Kusse, J.Rabaey [ISLPED,1998]
L. Shang [FPGA,2002]
K. Poon [FPL,2002]
Current data on FPGA leakage by commercial vendors


Limited to worst-case data points
Provide no details or understanding of the underlying issues
5
Background – Architectural Model (1)
1.2V, SRAM-based FPGAs built in a
90nm CMOS process
Modern FPGA architecture is composed
of CLBs, I/O blocks, clock managers,
block RAMs and multipliers.
This study focuses on the CLB array.
 CLB is the representative of smaller
or embedded FPGAs, which are more
suitable for mobile application.
CLB: Configurable Logic Block
6
Background – Architectural Model (2)
CLB consists of interconnect switch
matrix and four logic slices.
Switch matrix is composed of a set of
switches, each implemented with
configurable multiplexors.
OXBAR and IXBAR select CLB
outputs and inputs, respectively.
DOUBLE connect CLBs that are two
blocks apart.
Each logic slice consists of two 4-input
look-up tables (LUTs) and two flipflops (FFs).
7
Methodology – Overall
SPICE simulation using BSIM4 device
models

For sub-100nm devices, BSIM4
provides improvements in leakage
modeling.
Focus on a single CLB design since all
CLBs in an FPGA are identical


Each smaller block in CLB is
simulated individually to identify its
leakage power consumption.
Total leakage power of CLB array
= (∑ block leakages) * (# of CLBs)
8
Methodology – Input Data
Leakage power of a circuit depends on the values of its inputs.

Leakage varies by more than 4X depending on the values of inputs.
To model effects of input data variation, each block is simulated
under all possible input states.

For each block, its minimum, maximum and average leakage with
respect to input data under best-, worst- and average-case design data,
respectively, are stored.
9
Methodology – Resource Utilization
Designs implemented in FPGAs never use all available resources.



FPGA provides an excess routing resources to ensure the efficient
mapping and routing of most designs.
Typically, the utilization of interconnect resources is low, but for logic
blocks is relatively high.
TABLE 1: from survey of customer designs
Resource utilization affects total leakage power.



A block consumes less leakage power when unused.
Unused LUT stores all 0’s by default.
Each block is simulated in both used and unused configurations.
10
Result – FPGA Leakage
At 25̊C, the FPGA consumes 4.2uW per CLB, which means a 1000CLB FPGA would consume approximately 4.2mW.
At 85̊C, 18.9uW per CLB due to the exponential temperature
dependency of subthreshold leakage
GSM cell phone’s standby current budget for digital baseband
processing is 300uA.

A 300uA upper bound on leakage power would imply a limit of 86
CLBs at 25̊C , and 20 CLBs at 85̊C, which means too small sizes to
perform significant processing.
11
Result – Design Dependency
For best- and worst-case data conditions, the degree of variation from
the typical case is ±13% at 25̊C.
For CLB utilization factors of 50% and 100%, leakage power varies
by no more than ± 6%, which means non-trivial leakage in unused
resources.
For a design with worst-case data, 100% CLB utilization, at 85̊C is
26uW per CLB, 37.6% higher than average.
12
Result – Circuit-based Breakdown
In an FPGA, the three most common circuit types are configuration
SRAM cells, interconnect multiplexors and LUTs.
These three circuit types consume 88% of the total leakage power.
The configuration SRAM cells consume 38% of the total leakage.

Using high-Vth, thick-oxide transistors can eliminate most of the
SRAM leakage power, but will increase configuration time.
13
Result – Utilization-based Breakdown
The leakage consumption by the unused blocks is referred as
unused leakage.
For a small design that uses 50% of the CLBs, unused leakage is
56% of the total leakage power.
For a design of 100% CLB utilization, unused leakage is still a
significant portion (35%) of the total.


This amount is consumed entirely in unused interconnect switches.
FPGA CAD tools can help maximize the savings by grouping unused
resources such that they can be collectively powered down.
14
Conclusions
The results quantified the nominal leakage power of the FPGA and
the variation due to design-specific parameters.
The FPGA consumes 4.2uW per CLB nominally, and more than
26uW per CLB in the worst-case conditions.
To enable deployment of FPGAs in mobile applications, significant
leakage power reduction is needed.
The results indicated potential for substantial leakage reduction.
15
The End
Thank You
Q&A
16