Theses Doctoral

Machine Learning for AI-Augmented Design Space Exploration of Computer Systems

Kwon, Jihye

Advanced and emerging computer systems, ranging from supercomputers to embedded systems, feature high performance, energy efficiency, acceleration, and specialization. Design of such systems involves ever-increasing circuit complexity and architectural diversity. Commercial high-end processors, realized as very-large-scale integration circuits, have integrated exponentially increasing number of transistors on a chip over many decades. Along with the evolution of semiconductor manufacturing technology, another driving force behind the progress of processors has been the development of computer-aided design (CAD) software tools. Logic synthesis and physical design (LSPD) tool-chains allow designers to describe the computer system at the register-transfer level of abstraction and automatically convert the description into an integration circuit layout. The slowdown of technology scaling, on the other hand, has motivated the emergence of dark silicon and heterogeneous architectures with application-specific hardware accelerators. Design of various accelerators is facilitated by high-level synthesis (HLS) tools that translate a behavioral description of a computer system into a structural register-transfer level one. CAD approaches have evolved towards raising the level of design abstraction and providing more options to optimize the architecture.

For each system synthesized via advanced CAD tools, designers explore the design space in search of optimal configurations of the tool options and architectural choices, also called 𝘬𝘯𝘰𝘣𝘴. These knobs affect the execution of CAD algorithms and eventually impact the multi-dimensional 𝘲𝘶𝘢𝘭𝘪𝘵𝘺-𝘰𝘧-𝘳𝘦𝘴𝘶𝘭𝘵 (𝘘𝘰𝘙) of the final implementation. During design-space exploration (DSE), designers leverage their experience and expertise pertaining to determining the relationship between knobs and QoR. To further reduce the number of time and resource consuming CAD runs during DSE, a large number of heuristic and model-based approaches have been proposed. More recently, the rise of machine learning (ML) and artificial intelligence (AI) has prompted the possibility of AI-augmented DSE which exploits ML techniques to predict the knobs-QoR relationship. Yet, existing heuristic and ML-based approaches still require a sufficient number of CAD runs for each system because they do not accumulate and exploit experiential knowledge across the systems as designers would do.

To expand the potential of AI-augmented DSE and push the frontier forward, multiple challenges arise due to the characteristics of CAD flows. 1) Whereas many ML applications utilize data obtained from huge collections of users' input and public databases for a single problem, the QoR-prediction problem for each system suffers from limited availability of data obtained from expensive CAD runs. Especially, an industrial LSPD tool-chain specifies hundreds of separate knobs, resulting in an extreme curse of dimensionality. 2) Different systems exhibit different knobs-QoR relationship. Hence, learning from previously explored systems needs to be preceded by identifying distinct systems and relating them to one another. Often, it is difficult to obtain an efficient representation of a system. 3) Designers often apply different sets of knob configurations to different systems, which makes it harder to learn from previous DSE results. Especially in HLS, the heterogeneity of various systems leads to broad knob heterogeneity across them. To address these challenges and boost the ML performance, I propose to flexibly connect the elements of the many QoR-prediction problems with one another. My thesis is that 𝘵𝘩𝘦 𝘦𝘹𝘱𝘭𝘰𝘳𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘵𝘩𝘦 𝘥𝘦𝘴𝘪𝘨𝘯 𝘴𝘱𝘢𝘤𝘦 𝘰𝘧 𝘢 𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳 𝘴𝘺𝘴𝘵𝘦𝘮 𝘤𝘢𝘯 𝘣𝘦 𝘦𝘧𝘧𝘦𝘤𝘵𝘪𝘷𝘦𝘭𝘺 𝘢𝘶𝘨𝘮𝘦𝘯𝘵𝘦𝘥 𝘣𝘺 𝘢𝘳𝘵𝘪𝘧𝘪𝘤𝘪𝘢𝘭 𝘪𝘯𝘵𝘦𝘭𝘭𝘪𝘨𝘦𝘯𝘤𝘦 𝘷𝘪𝘢 𝘭𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘧𝘳𝘰𝘮 𝘵𝘩𝘦 𝘦𝘹𝘱𝘦𝘳𝘪𝘦𝘯𝘤𝘦 𝘸𝘪𝘵𝘩 𝘵𝘩𝘦 𝘥𝘦𝘴𝘪𝘨𝘯 𝘢𝘯𝘥 𝘰𝘱𝘵𝘪𝘮𝘪𝘻𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘰𝘵𝘩𝘦𝘳 𝘴𝘺𝘴𝘵𝘦𝘮𝘴.

For LSPD of industrial high-performance processors, I propose a novel collaborative recommender system approach that learns hidden features from the interactions (CAD runs) of many \textit{users} (systems) and \textit{items} (knob configurations). To cope with the curse of dimensionality, the item features are decomposed into features of item attributes (knobs). The combined model predicts QoR for each user-item pair. For HLS of application-specific accelerators, I present a series of neural network models in the order of evolution towards the proposed mixed-sharing \textit{transfer learning} model. Transfer learning aims at leveraging knowledge gained from previous problems; however, due to the system and knob heterogeneities, the model needs to distinguish which piece of that knowledge should be transferred. The proposed ML approaches aim to not only use experiential knowledge as designers do but also to ultimately assist designers by providing alternative insights and suggesting optimization possibilities for new systems. As an effort in this direction, I develop an AI-augmented DSE tool that exploits the aforementioned models and \textit{generates} recommended knob configurations for new target systems. Through this research, I investigate the potential of next-level AI-augmented DSE with the goal of promoting secure collaborative engineering in the CAD community without the need of sharing confidential information and intellectual properties.

Files

  • thumnail for Kwon_columbia_0054D_17226.pdf Kwon_columbia_0054D_17226.pdf application/pdf 2.21 MB Download File

More About This Work

Academic Units
Computer Science
Thesis Advisors
Carloni, Luca P.
Degree
Ph.D., Columbia University
Published Here
May 4, 2022