Chi-Chun Liu
Chi-Chun Liu has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
Kafai Lai
IBM Research – TJ Watson, Yorktown Heights, NY, USA
2 shared publicationsAlice C. Lee
Finance Department, San Francisco State University, USA
1 shared publication- Andrea Fasoli
IBM Research USA
1 shared publication - Ankit Vora
IBM Research-Almaden , 650 Harry Rd, San Jose, California 95120, United States
1 shared publication Anthony Schepis
IBM Corp. (United States)
1 shared publicationAzalia A Krasnoperova
IBM Corp. (United States)
1 shared publication
Latest Publications
- Open Access
Is Finer Better? The Limits of Microscaling Formats in Large Language Models
Microscaling data formats leverage per-block tensor quantization to enable aggressive model compression with limited loss in accuracy. Unlocking their potential for efficient training and inference necessitates hardware-friendly implementations that handle matrix multiplications in a native format and adopt efficient error-mitigation strategies. Herein, we report the emergence of a surprising behavior associated with microscaling quantization, whereas the output of a quantized model degrades as block size is decreased below a given threshold. This behavior clashes with the expectation that a smaller block size should allow for a better representation of the tensor elements. We investigate this phenomenon both experimentally and theoretically, decoupling the sources of quantization error behind it. Experimentally, we analyze the distributions of several Large Language Models and identify the conditions driving the anomalous behavior. Theoretically, we lay down a framework showing remarkable agreement with experimental data from pretrained model distributions and ideal ones. Overall, we show that the anomaly is driven by the interplay between narrow tensor distributions and the limited dynamic range of the quantized scales. Based on these insights, we propose the use of FP8 unsigned E5M3 (UE5M3) as a novel hardware-friendly format for the scales in FP4 microscaling data types. We demonstrate that UE5M3 achieves comparable performance to the conventional FP8 unsigned E4M3 scales while obviating the need of global scaling operations on weights and activations.
Open AccessIs Finer Better? The Limits of Microscaling Formats in Large Language Models
Microscaling data formats leverage per-block tensor quantization to enable aggressive model compression with limited loss in accuracy. Unlocking their potential for efficient training and inference necessitates hardware-friendly implementations that handle matrix multiplications in a native format and adopt efficient error-mitigation strategies. Herein, we report the emergence of a surprising behavior associated with microscaling quantization, whereas the output of a quantized model degrades as block size is decreased below a given threshold. This behavior clashes with the expectation that a smaller block size should allow for a better representation of the tensor elements. We investigate this phenomenon both experimentally and theoretically, decoupling the sources of quantization error behind it. Experimentally, we analyze the distributions of several Large Language Models and identify the conditions driving the anomalous behavior. Theoretically, we lay down a framework showing remarkable agreement with experimental data from pretrained model distributions and ideal ones. Overall, we show that the anomaly is driven by the interplay between narrow tensor distributions and the limited dynamic range of the quantized scales. Based on these insights, we propose the use of FP8 unsigned E5M3 (UE5M3) as a novel hardware-friendly format for the scales in FP4 microscaling data types. We demonstrate that UE5M3 achieves comparable performance to the conventional FP8 unsigned E4M3 scales while obviating the need of global scaling operations on weights and activations.
- 12 October 2018
- Published by Springer Nature in Nature Electronics
- Vol. 1 (10) , 562-569
Directed self-assembly of block copolymers for 7 nanometre FinFET technology and beyond
- 12 October 2018
- Published by Springer Nature in Nature Electronics
- Vol. 1 (10) , 562-569
Directed self-assembly of block copolymers for 7 nanometre FinFET technology and beyond
- 29 July 2016
- Published by American Chemical Society (ACS) in ACS Nano
- Vol. 10 (8) , 7705-7720
Surface Assembly Configurations and Packing Preferences of Fibrinogen Mediated by the Periodicity and Alignment Control of Block Copolymer Nanodomains
The ability to control the specific adsorption and packing behaviors of biomedically important proteins by effectively guiding their preferred surface adsorption configuration and packing orientation on polymeric surfaces may have utility in many applications such as biomaterials, medical implants, and tissue engineering. Herein, we investigate the distinct adhesion configurations of fibrinogen (Fg) proteins and the different organization behaviors between single Fg molecules that are mediated by the changes in the periodicity and alignment of chemically alternating nanodomains in the thin films of polystyrene-block-polymethylmethacrylate (PS-b-PMMA) block copolymer (BCP). Specifically, the adsorption characteristics of individual Fg molecules were unambiguously resolved on four different PS-b-PMMA templates of dsa PS-b-PMMA, sm PS-b-PMMA, com PS-b-PMMA, and PS-r-PMMA. Via direct visualization by high resolution imaging, the distinct adsorption and packing configurations of both isolated and interacting Fg molecules were determined as a function of the BCP template-specific nanodomain periodicity, domain alignment (random versus fully aligned), and protein concentration. The three dominant Fg adsorption configurations, SP║, SP┴, and TP, were observed and their occurrence ratios were ascertained on each PS-b-PMMA template. During surface packing, the orientation of the protein backbone was largely governed by the periodicity and alignment of the underlying PS-b-PMMA nanodomains whose specific direction was explicitly resolved relative to the polymeric nanodomain axis. The use of PS-b-PMMA with a periodicity much smaller than (and comparable to) the length of Fg led to a Fg scaffold with the protein backbone aligned parallel (and perpendicular) to the nanodomain major axis. In addition, we have successfully created fully Fg-decorated BCP constructs analogous to two-dimensional Fg crystals in which aligned protein molecules are arranged either side-on or end-on, depending on the BCP template. Our results demonstrate that the geometry and orientation of the protein can be effectively guided during Fg self-assembly by controlling the physical dimensions and orientations of the underlying BCP templates. Finally, the biofunctionality of the BCP surface-bound Fg was assessed and the Fg/BCP construct was successfully used in the Ca-P nanoparticle nucleation/growth and microglia cell activation.
- 29 July 2016
- Published by American Chemical Society (ACS) in ACS Nano
- Vol. 10 (8) , 7705-7720
Surface Assembly Configurations and Packing Preferences of Fibrinogen Mediated by the Periodicity and Alignment Control of Block Copolymer Nanodomains
The ability to control the specific adsorption and packing behaviors of biomedically important proteins by effectively guiding their preferred surface adsorption configuration and packing orientation on polymeric surfaces may have utility in many applications such as biomaterials, medical implants, and tissue engineering. Herein, we investigate the distinct adhesion configurations of fibrinogen (Fg) proteins and the different organization behaviors between single Fg molecules that are mediated by the changes in the periodicity and alignment of chemically alternating nanodomains in the thin films of polystyrene-block-polymethylmethacrylate (PS-b-PMMA) block copolymer (BCP). Specifically, the adsorption characteristics of individual Fg molecules were unambiguously resolved on four different PS-b-PMMA templates of dsa PS-b-PMMA, sm PS-b-PMMA, com PS-b-PMMA, and PS-r-PMMA. Via direct visualization by high resolution imaging, the distinct adsorption and packing configurations of both isolated and interacting Fg molecules were determined as a function of the BCP template-specific nanodomain periodicity, domain alignment (random versus fully aligned), and protein concentration. The three dominant Fg adsorption configurations, SP║, SP┴, and TP, were observed and their occurrence ratios were ascertained on each PS-b-PMMA template. During surface packing, the orientation of the protein backbone was largely governed by the periodicity and alignment of the underlying PS-b-PMMA nanodomains whose specific direction was explicitly resolved relative to the polymeric nanodomain axis. The use of PS-b-PMMA with a periodicity much smaller than (and comparable to) the length of Fg led to a Fg scaffold with the protein backbone aligned parallel (and perpendicular) to the nanodomain major axis. In addition, we have successfully created fully Fg-decorated BCP constructs analogous to two-dimensional Fg crystals in which aligned protein molecules are arranged either side-on or end-on, depending on the BCP template. Our results demonstrate that the geometry and orientation of the protein can be effectively guided during Fg self-assembly by controlling the physical dimensions and orientations of the underlying BCP templates. Finally, the biofunctionality of the BCP surface-bound Fg was assessed and the Fg/BCP construct was successfully used in the Ca-P nanoparticle nucleation/growth and microglia cell activation.
- 31 March 2014
- Published by SPIE-Intl Soc Optical Eng
- Vol. 9052, 90521A-90521A-12
Computational lithography platform for 193i-guided directed self-assembly
- 31 March 2014
- Published by SPIE-Intl Soc Optical Eng
- Vol. 9052, 90521A-90521A-12
Computational lithography platform for 193i-guided directed self-assembly
- 1 March 2009
- Published by World Scientific Pub Co Pte Ltd in Review of Pacific Basin Financial Markets and Policies
- Vol. 12 (01) , 27-62
Balanced Performance Index and Its Implications: Evidence from Taiwan's Commercial Banks
Taking into account only financial factors does not provide complete information on performance. This paper takes into consideration of both financial and non-financial performances when evaluating 35 sampled publicly traded commercial banks in Taiwan. The performance of banks is measured using an indexing method consisting of financial and non-financial measures. Banks are classified into two categories according either to the year founded, or to the type of major stockholders of a bank when founded. The results show that privatized government-owned/old banks are larger than private/new banks, respectively. Moreover, privatized government-owned banks have significantly higher financial performance index than private banks but both types of banks are not significantly different from each other in non-financial performance index. New and old banks are not significantly different from each other in both financial and non-financial performance indexes. With relatively large scale, higher profitability and better management, banks will perform relatively better among competitors in the following year. Furthermore, non-financial factors are important predictors of future financial and total performance indexes, though individual factor may not be consistently significant. More branch offices, better capital structure and solvency, and higher rates of growth in deposits and loans all result in more profits, and lead to higher customer satisfaction and more efficient management. Providing better technology to customers is an efficient way in promoting customer services, which in turn produces more profits and results in efficient management. CEOs, on average, have plans for better management and more profits. Among the factors that have direct and positive impacts on profitability, increasing the efficiency of management is the most efficient way. On the contrary, adding more branch offices contributes the least profits. Therefore, to increase bank profits, CEOs should aim to improve bank management, capital structure and solvency, rather than to add more branch offices.
- 1 March 2009
- Published by World Scientific Pub Co Pte Ltd in Review of Pacific Basin Financial Markets and Policies
- Vol. 12 (01) , 27-62
Balanced Performance Index and Its Implications: Evidence from Taiwan's Commercial Banks
Taking into account only financial factors does not provide complete information on performance. This paper takes into consideration of both financial and non-financial performances when evaluating 35 sampled publicly traded commercial banks in Taiwan. The performance of banks is measured using an indexing method consisting of financial and non-financial measures. Banks are classified into two categories according either to the year founded, or to the type of major stockholders of a bank when founded. The results show that privatized government-owned/old banks are larger than private/new banks, respectively. Moreover, privatized government-owned banks have significantly higher financial performance index than private banks but both types of banks are not significantly different from each other in non-financial performance index. New and old banks are not significantly different from each other in both financial and non-financial performance indexes. With relatively large scale, higher profitability and better management, banks will perform relatively better among competitors in the following year. Furthermore, non-financial factors are important predictors of future financial and total performance indexes, though individual factor may not be consistently significant. More branch offices, better capital structure and solvency, and higher rates of growth in deposits and loans all result in more profits, and lead to higher customer satisfaction and more efficient management. Providing better technology to customers is an efficient way in promoting customer services, which in turn produces more profits and results in efficient management. CEOs, on average, have plans for better management and more profits. Among the factors that have direct and positive impacts on profitability, increasing the efficiency of management is the most efficient way. On the contrary, adding more branch offices contributes the least profits. Therefore, to increase bank profits, CEOs should aim to improve bank management, capital structure and solvency, rather than to add more branch offices.