Mark夸我做得好
| Original | Correction | Reasoning |
|---|---|---|
| Central to my research interests are optimization and GPU-accelerated methods in VLSI design and test, as well as geometric deep learning and its applications in EDA. | My primary research interests lie in optimization, GPU-accelerated techniques in VLSI design and testing, alongside geometric deep learning and its applications in Electronic Design Automation (EDA). | Clarification and enhanced readability. |
| My research endeavors have culminated in over a dozen publications, including three best-paper awards. | My research efforts have resulted in over twelve publications, including three papers recognized as best in their respective categories. | Improved sentence structure for clarity. |
| Chiplets, exotic packages, 2.5D, 3D, mechanical and thermal concerns, gate all around, …, make the already-hard problem of IC design that much more challenging. | The incorporation of chiplets, exotic packages, 2.5D, 3D integration, as well as mechanical and thermal considerations, alongside emerging technologies like gate-all-around (GAA), exacerbate the already complex nature of IC design. | Enhanced clarity and precision. |
| Existing CPU-based approaches for design, analysis, and optimization are running out of steam, and simply migrating to GPU enhancements is not sufficient for keeping pace. | Traditional CPU-based approaches for design, analysis, and optimization are becoming inadequate, and a mere transition to GPU enhancements is insufficient to maintain pace. | Improved phrasing for academic style. |
| New ways of solving both existing and emerging problems are therefore desperately needed. I am very much attracted to these types of challenges and take pleasure in generating solutions that exceed old techniques by orders of magnitude. | There is an urgent need for innovative solutions to address both existing and emerging challenges. I am particularly drawn to these types of problems and find satisfaction in devising solutions that surpass previous methods by significant margins. | Strengthened expression of urgency and motivation. |
| In collaboration with Stanford University, my lab at CMU developed a new testing approach for eliminating defects that led to silent data errors in large compute enterprises \cite{li2022pepr}. Our approach involved analyzing both the physical layout and the logic netlist to identify single- or multi-output sub-circuits. The method is entirely infeasible without using GPUs, which allows us to extract more than 12 billion sub-circuits in less than an hour using an 8-GPU machine. In contrast, a CPU-based implementation required a runtime exceeding 150 hours. | In collaboration with Stanford University, my lab at CMU devised a novel testing methodology to rectify defects responsible for silent data errors in extensive compute infrastructures \cite{li2022pepr}. Our approach entailed scrutinizing both the physical layout and logic netlist to pinpoint single or multi-output sub-circuits. This method is entirely unattainable without the use of GPUs, enabling us to extract over 12 billion sub-circuits in under an hour using an 8-GPU system. In contrast, a CPU-based implementation demanded a runtime exceeding 150 hours. | Enhanced clarity and precision. |
| My summer intern project concerning global routing at NVIDIA\footnote{Submitted to IEEE/ACM Proceedings Design, Automation and Test in Europe, 2024} is another example to demonstrate. Specifically, traditional CPU-based global routing algorithms mostly route nets sequentially. However, with the support of GPUs, we proposed and demonstrated a novel differentiable global router that enables concurrent optimization of millions of nets. | My summer internship project focused on global routing at NVIDIA\footnote{Submitted to IEEE/ACM Proceedings Design, Automation and Test in Europe, 2024} serves as another illustration. Conventional CPU-based global routing algorithms predominantly route nets sequentially. Nonetheless, with the aid of GPUs, we introduced and demonstrated an innovative differentiable global router that facilitates simultaneous optimization of millions of nets. | Improved phrasing for academic style. |
| Motivated by my intern project at Apple in 2022. Unlike traditional floorplanning algorithms, which heavily relied on carefully designed data structure and heuristic cost function, I first proposed a Semi-definite programming-based method for initial floorplanning, which is a totally new method and outperforms previous methods significantly \cite{10247967}. Furthermore, I designed a novel differentiable floorplanning algorithm with the support of GPU, which is also the pioneering work that pixelized the floorplanning problem. | Inspired by my internship project at Apple in 2022, I introduced a fundamentally new approach to initial floorplanning. In contrast to conventional methods, which heavily depend on intricately crafted data structures and heuristic cost functions, I advocated for a Semi-definite programming-based approach that exhibits superior performance \cite{10247967}. Additionally, I devised an innovative differentiable floorplanning algorithm with GPU support, marking a pioneering effort in pixelizing the floorplanning problem. | Enhanced clarity and precision. |
| While Artificial Intelligence (AI) has witnessed resounding triumphs across diverse domains—from Convolutional Neural Networks revolutionizing Computer Vision to Transformers reshaping Natural Language Processing, culminating in Large Language Models propelling Artificial General Intelligence (AGI)—its impact on the IC domain has been somewhat less revolutionary than anticipated. | Although Artificial Intelligence (AI) has achieved remarkable success in various domains—from the revolutionizing effects of Convolutional Neural Networks on Computer Vision to the transformative impact of Transformers on Natural Language Processing, culminating in Large Language Models propelling Artificial General Intelligence (AGI)—its influence on the IC domain has been somewhat less groundbreaking than anticipated. | Strengthened expression and enhanced precision. |
| This can be attributed, in part, to the irregular nature of elements within the VLSI workflow. Notably, both logic netlists and Register Transfer Level (RTL) designs inherently lend themselves to representation as hyper-graphs. Moreover, the connectivity matrix among blocks, modules, and IP-cores is aptly described by a directed graph. Unlike the regularity found in images or textual constructs, the application of AI to glean insights from such irregular data remains an ongoing inquiry. | This can be attributed, at least in part, to the irregular nature of components within the VLSI workflow. Notably, both logic netlists and Register Transfer Level (RTL) designs inherently lend themselves to representation as hyper-graphs. Furthermore, the connectivity matrix among blocks, modules, and IP-cores is aptly described by a directed graph. In contrast to the regularity found in images or textual constructs, the application of AI to extract insights from such irregular data remains an ongoing area of exploration. | Enhanced clarity and precision. |
| My prior investigations into layout decomposition \cite{li2020adaptive} and routing tree construction \cite{li2021treenet} vividly underscore the immense potential and efficacy of geometric learning-based methodologies in tackling IC challenges. | My previous studies on layout decomposition \cite{li2020adaptive} and routing tree construction \cite{li2021treenet} strongly emphasize the significant potential and effectiveness of geometric learning-based approaches in addressing IC challenges. | Improved phrasing for academic style. |
| A recent contribution \cite{li2023char} of mine delves into the theoretical boundaries of Graph Neural Networks (GNNs) in representing logic netlists. Drawing upon these foundations and experiences, my overarching research ambition in my PhD trajectory is to develop the Large nelist model to capture the function information of logic netlist |
