Early power estimation is a critical requirement for both SoC (system-on-chip) architects and marketing, which must provide competitive, yet realistic, power consumption RFQs (request for quotation) to customers in order to achieve design wins. RFQs based on deterministic techniques may work reasonably well when there is enough information, and the available data closely match the target design. However, this approach can be strongly design dependent, and the same model, which provided a good estimate on one design/technology, may fail on another design/technology with different specs and operating conditions. In this talk, we will present an innovative approach for early power consumption estimation based on statistical learning. The proposed statistical model has been compared against existing RFQs, and validated on two taped-out products for networking in leading-edge CMOS technology.
Davide Pandini holds a Laurea Degree (Summa cum Laude) in Electronics Engineering from the University of Bologna, Italy, a Ph.D. in Electrical Engineering and Telecommunications from the University of Padova,Italy, and a Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University, Pittsburgh, U.S.He was a research intern at Philips Research Labs. in Eindhoven, the Netherlands, and at Digital EquipmentCorp., Western Research Labs. in Palo Alto, CA.He joined STMicroelectronics in Agrate Brianza, Italy, in 1995, where he is a R&D manager and a senior member of the technical staff. His current research interests include signal and power integrity, variation-tolerant design, design for manufacturing, statistical design, and EMC-aware design. Dr. Pandini has authored and coauthored more than fifty papers in international journals and conference proceedings, and during the academic years from 1998 to 2000, he was a visiting professor at the University of Brescia, Italy. He is a work-package manager in several funded European projects, and he is on the program committee of several international conferences. In 2008 Dr. Pandini was the recipient of the ST Corporate STAR 2008 Gold Award for leading the R&D excellence team on EMC-aware design. Since June 2015, Dr. Pandini is the Chairman of the Steering Committee of the ST Italy Technical Staff. In September 2015, Davide Pandini served as Volunteer at EXPO2015 "Feeding the Planet, Energy for Life", in Milano, Italy.
Power Consumption Prediction based on Statistical Learning Techniques - Davide Pandini, STMicroelectronics
1. Proposal Title:
Power consumption prediction based on
statistical learning technique
Abstract
Early power estimation is a critical requirement for both SoC (system-on-chip) architects and
marketing, which must provide competitive, yet realistic, power consumption RFQs (request for
quotation) to customers in order to achieve design wins. RFQs based on deterministic techniques
may work reasonably well when there is enough information, and the available data closely match
the target design. However, this approach can be strongly design dependent, and the same
model, which provided a good estimate on one design/technology, may fail on another
design/technology with different specs and operating conditions. In this talk, we will present an
innovative approach for early power consumption estimation based on statistical learning. The
proposed statistical model has been compared against existing RFQs, and validated on two
taped-out products for networking in leading-edge CMOS technology.
Proposal Description
Statistical learning (sometimes also referred as machine learning) is a set of powerful tools for
understanding and modeling data sets and for learning from data. Statistical learning builds a statistical
model for predicting a dependent (output) variable, like power consumption, from one or more
independent (input) variables, such as supply voltage, temperature, active area, working frequency, and
switching activity. Today, these techniques are successfully used in several fields such as stock markets,
artificial intelligence, pattern recognition, computer vision, bioinformatics, spam filtering, data mining,
etc. Although data interpolation and fitting techniques (i.e., least-squares) have been used in the past, a
unified theoretical framework that gives a probabilistic interpretation to data interpolation and exploits
powerful convex optimization algorithms to solve large-scale complex problems has been proposed since
only about twenty years. Only a formal probabilistic interpretation makes data interpolation and statistical
learning suitable to solve prediction and classification problems in many different domains.
To the best of our knowledge, statistical learning has never been used before for power consumption
prediction in the VLSI domain.
Previous attempts to provide realistic and consistent power consumption RFQ estimations across a wide
range of circuits and technologies have proven unreliable, with significant discrepancies with respect to
power analysis performed with standard CAD flows in the worst-power condition. Much more
importantly, the mismatch between RFQ estimations and CAD-based power analysis has put marketing
2. teams in a tough position with customers, since it was difficult for them to commit on the power
consumption targets set by the customers. Current RFQs are based on spreadsheets incorporating basic
formulas for power calculation that can hardly capture the increasing complexity and heterogeneity of
large SoC designs in advanced technologies. We believe that further attempts to refine these spreadsheets
are either doomed to failure or only some partial success, as they will never be able to realistically
modeling such complex designs.
The fundamental idea underlying this work was to change the perspective. Instead of tweaking by trial-
and-error basic formulas in order to capture the design complexity and advanced technological effects,
we decided to start from a sample space of power analysis results obtained with our CAD (computer-aided
design) flow, and to consider that sample space as the training set for statistical learning. Therefore, our
contribution was to exploit statistical learning techniques to develop a model to predict the power
consumption and the related scaling factors. In particular, our approach is based on multiple linear
regression and power transformations, as outlined in this flow chart shown below.
Of course, developing a model that satisfies all the statistical significance hypotheses is not a
straightforward procedure, and finding the optimal data transformations requires a careful data analysis
and inspection. The methodology outlined in the flowchart has been developed and tested using the most
popular statistical and numerical analysis frameworks, such as R, Mathematica, and Matlab, and the
resulting model and equations can be implemented into an Excel spreadsheet.
3. We trained our model on the set of power consumption data obtained from the sub-blocks of an ASIC for
networking in a leading edge CMOS technology, and we tested the model on the sign-off power
consumption data of two SoC designs for networking in the same silicon technology. Although this model
was obtained and tested on a specific technology/design architecture, we believe that the general
theoretical foundations of statistical learning make our approach suitable for different design
architectures and technology nodes, and it can effectively complement deterministic RFQ estimations.
Speaker's Bio:
Davide Pandini, Ph.D.,
STMicroelectronics
R&D Manager and Senior Member of the Technical Staff
Davide Pandini holds a Laurea Degree (Summa cum Laude) in Electronics Engineering from the University
of Bologna, Italy, a Ph.D. in Electrical Engineering and Telecommunications from the University of Padova,
Italy, and a Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University, Pittsburgh, U.S.
He was a research intern at Philips Research Labs. in Eindhoven, the Netherlands, and at Digital Equipment
Corp., Western Research Labs. in Palo Alto, CA.
He joined STMicroelectronics in Agrate Brianza, Italy, in 1995, where he is a R&D manager and a senior
member of the technical staff. His current research interests include signal and power integrity, variation-
tolerant design, design for manufacturing, statistical design, and EMC-aware design.
Dr. Pandini has authored and coauthored more than fifty papers in international journals and conference
proceedings, and during the academic years from 1998 to 2000, he was a visiting professor at the
4. University of Brescia, Italy. He is a work-package manager in several funded European projects, and he is
on the program committee of several international conferences.
In 2008 Dr. Pandini was the recipient of the ST Corporate STAR 2008 Gold Award for leading the R&D
excellence team on EMC-aware design.
Since June 2015, Dr. Pandini is the Chairman of the Steering Committee of the ST Italy Technical Staff.
In September 2015, Davide Pandini served as Volunteer at EXPO2015 "Feeding the Planet, Energy for Life",
in Milano, Italy.