An Zou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
An Zou has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
- Shalabh C. Maroo
Department of Mechanical and Aerospace Engineering, Syracuse University, Syracuse, New York 13244, United States
13 shared publications Manish Gupta
Department of Mechanical and Aerospace Engineering, Syracuse University, Syracuse, New York 13244, United States
3 shared publicationsSajag Poudel
Department of Mechanical and Aerospace Engineering, Syracuse University, 263 Link Hall, Syracuse, New York 13244, United States
3 shared publications- Sajag Poudel
Department of Mechanical & Aerospace Engineering, 263 Link Hall, Syracuse University, NY 13244, USA
3 shared publications - Yehan Ma
Shanghai Jiao Tong University
3 shared publications Qi Fan
Shanghai Jiao Tong University
2 shared publications
Latest Publications
- Open Access
TimeBill: Time-Budgeted Inference for Large Language Models
Large Language Models (LLMs) are increasingly deployed in time-critical systems, such as robotics, autonomous driving, embodied intelligence, and industrial automation, where generating accurate responses within a given time budget is crucial for decision-making, control, or safety-critical tasks. However, the auto-regressive generation process of LLMs makes it challenging to model and estimate the end-to-end execution time. Furthermore, existing efficient inference methods based on a fixed key-value (KV) cache eviction ratio struggle to adapt to varying tasks with diverse time budgets, where an improper eviction ratio may lead to incomplete inference or a drop in response performance. In this paper, we propose TimeBill, a novel time-budgeted inference framework for LLMs that balances the inference efficiency and response performance. To be more specific, we propose a fine-grained response length predictor (RLP) and an execution time estimator (ETE) to accurately predict the end-to-end execution time of LLMs. Following this, we develop a time-budgeted efficient inference approach that adaptively adjusts the KV cache eviction ratio based on execution time prediction and the given time budget. Finally, through extensive experiments, we demonstrate the advantages of TimeBill in improving task completion rate and maintaining response performance under various overrun strategies.
Open AccessTimeBill: Time-Budgeted Inference for Large Language Models
Large Language Models (LLMs) are increasingly deployed in time-critical systems, such as robotics, autonomous driving, embodied intelligence, and industrial automation, where generating accurate responses within a given time budget is crucial for decision-making, control, or safety-critical tasks. However, the auto-regressive generation process of LLMs makes it challenging to model and estimate the end-to-end execution time. Furthermore, existing efficient inference methods based on a fixed key-value (KV) cache eviction ratio struggle to adapt to varying tasks with diverse time budgets, where an improper eviction ratio may lead to incomplete inference or a drop in response performance. In this paper, we propose TimeBill, a novel time-budgeted inference framework for LLMs that balances the inference efficiency and response performance. To be more specific, we propose a fine-grained response length predictor (RLP) and an execution time estimator (ETE) to accurately predict the end-to-end execution time of LLMs. Following this, we develop a time-budgeted efficient inference approach that adaptively adjusts the KV cache eviction ratio based on execution time prediction and the given time budget. Finally, through extensive experiments, we demonstrate the advantages of TimeBill in improving task completion rate and maintaining response performance under various overrun strategies.
- Open Access
CUDA-LLM: LLMs Can Write Efficient CUDA Kernels
Large Language Models (LLMs) have demonstrated strong capabilities in general-purpose code generation. However, generating the code which is deeply hardware-specific, architecture-aware, and performance-critical, especially for massively parallel GPUs, remains a complex challenge. In this work, we explore the use of LLMs for the automated generation and optimization of CUDA programs, with the goal of producing high-performance GPU kernels that fully exploit the underlying hardware. To address this challenge, we propose a novel framework called \textbf{Feature Search and Reinforcement (FSR)}. FSR jointly optimizes compilation and functional correctness, as well as the runtime performance, which are validated through extensive and diverse test cases, and measured by actual kernel execution latency on the target GPU, respectively. This approach enables LLMs not only to generate syntactically and semantically correct CUDA code but also to iteratively refine it for efficiency, tailored to the characteristics of the GPU architecture. We evaluate FSR on representative CUDA kernels, covering AI workloads and computational intensive algorithms. Our results show that LLMs augmented with FSR consistently guarantee correctness rates. Meanwhile, the automatically generated kernels can outperform general human-written code by a factor of up to 179$\times$ in execution speeds. These findings highlight the potential of combining LLMs with performance reinforcement to automate GPU programming for hardware-specific, architecture-sensitive, and performance-critical applications.
Open AccessCUDA-LLM: LLMs Can Write Efficient CUDA Kernels
Large Language Models (LLMs) have demonstrated strong capabilities in general-purpose code generation. However, generating the code which is deeply hardware-specific, architecture-aware, and performance-critical, especially for massively parallel GPUs, remains a complex challenge. In this work, we explore the use of LLMs for the automated generation and optimization of CUDA programs, with the goal of producing high-performance GPU kernels that fully exploit the underlying hardware. To address this challenge, we propose a novel framework called \textbf{Feature Search and Reinforcement (FSR)}. FSR jointly optimizes compilation and functional correctness, as well as the runtime performance, which are validated through extensive and diverse test cases, and measured by actual kernel execution latency on the target GPU, respectively. This approach enables LLMs not only to generate syntactically and semantically correct CUDA code but also to iteratively refine it for efficiency, tailored to the characteristics of the GPU architecture. We evaluate FSR on representative CUDA kernels, covering AI workloads and computational intensive algorithms. Our results show that LLMs augmented with FSR consistently guarantee correctness rates. Meanwhile, the automatically generated kernels can outperform general human-written code by a factor of up to 179$\times$ in execution speeds. These findings highlight the potential of combining LLMs with performance reinforcement to automate GPU programming for hardware-specific, architecture-sensitive, and performance-critical applications.
- Open Access
- 26 June 2023
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 37 (7) , 8657-8665
Predictive Exit: Prediction of Fine-Grained Early Exits for Computation- and Energy-Efficient Inference
By adding exiting layers to the deep learning networks, early exit can terminate the inference earlier with accurate results. However, the passive decision-making of whether to exit or continue the next layer has to go through every pre-placed exiting layer until it exits. In addition, it is hard to adjust the configurations of the computing platforms alongside the inference proceeds. By incorporating a low-cost prediction engine, we propose a Predictive Exit framework for computation- and energy-efficient deep learning applications. Predictive Exit can forecast where the network will exit (i.e., establish the number of remaining layers to finish the inference), which effectively reduces the network computation cost by exiting on time without running every pre-placed exiting layer. Moreover, according to the number of remaining layers, proper computing configurations (i.e., frequency and voltage) are selected to execute the network to further save energy. Extensive experimental results demonstrate that Predictive Exit achieves up to 96.2% computation reduction and 72.9% energy-saving compared with classic deep learning networks; and 12.8% computation reduction and 37.6% energy-saving compared with the early exit under state-of-the-art exiting strategies, given the same inference accuracy and latency.
- 26 June 2023
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 37 (7) , 8657-8665
Open AccessPredictive Exit: Prediction of Fine-Grained Early Exits for Computation- and Energy-Efficient Inference
By adding exiting layers to the deep learning networks, early exit can terminate the inference earlier with accurate results. However, the passive decision-making of whether to exit or continue the next layer has to go through every pre-placed exiting layer until it exits. In addition, it is hard to adjust the configurations of the computing platforms alongside the inference proceeds. By incorporating a low-cost prediction engine, we propose a Predictive Exit framework for computation- and energy-efficient deep learning applications. Predictive Exit can forecast where the network will exit (i.e., establish the number of remaining layers to finish the inference), which effectively reduces the network computation cost by exiting on time without running every pre-placed exiting layer. Moreover, according to the number of remaining layers, proper computing configurations (i.e., frequency and voltage) are selected to execute the network to further save energy. Extensive experimental results demonstrate that Predictive Exit achieves up to 96.2% computation reduction and 72.9% energy-saving compared with classic deep learning networks; and 12.8% computation reduction and 37.6% energy-saving compared with the early exit under state-of-the-art exiting strategies, given the same inference accuracy and latency.
- Open Access
- 20 April 2023
- Published by American Chemical Society (ACS) in ACS Applied Materials & Interfaces
- Vol. 15 (17) , 21679-21689
Dropwise Condensation in Ambient on a Depleted Lubricant-Infused Surface
Durability of a lubricant-infused surface (LIS) is critical for heat transfer, especially in condensation-based applications. Although LIS promotes dropwise condensation, each departing droplet condensate acts as a lubricant-depleting agent due to the formation of wetting ridge and cloaking layer around the condensate, thus gradually leading to drop pinning on the underlying rough topography. Condensation heat transfer further deteriorates in the presence of non-condensable gases (NCGs) requiring special experimental arrangements to eliminate NCGs due to a decrease in the availability of nucleation sites. To address these issues while simultaneously improving heat-transfer performance of LIS in condensation-based systems, we report fabrication of both fresh LIS and a lubricant-depleted LIS using silicon porous nanochannel wicks as an underlying substrate. Strong capillarity in the nanochannels helps retain silicone oil (polydimethylsiloxane) on the surface even after it is severely depleted under tap water. The effect of oil viscosity was investigated for drop mobility and condensation heat transfer under ambient conditions, i.e., in the presence of NCGs. While fresh LIS prepared using 5 cSt silicone oil exhibited a low roll-off angle (∼1°) and excellent water drop (5 μL) sliding velocity ∼66 mm s-1, it underwent rapid depletion as compared to higher viscosity oils. Condensation performed on depleted nanochannel LIS with higher viscosity oil (50 cSt) resulted in a heat-transfer coefficient (HTC) of ∼2.33 kW m-2 K-1, which is a ∼162% improvement over flat Si-LIS (50 cSt). Such LIS promote fast drop shedding as is evident from the little change in the fraction of drops with diameter -2 K-1 was achieved over the last 2 days. The ability of reported LIS to maintain long-term hydrophobicity and dropwise condensation will aid in designing condensation-based systems with improved heat-transfer performance.
- 20 April 2023
- Published by American Chemical Society (ACS) in ACS Applied Materials & Interfaces
- Vol. 15 (17) , 21679-21689
Open AccessDropwise Condensation in Ambient on a Depleted Lubricant-Infused Surface
Durability of a lubricant-infused surface (LIS) is critical for heat transfer, especially in condensation-based applications. Although LIS promotes dropwise condensation, each departing droplet condensate acts as a lubricant-depleting agent due to the formation of wetting ridge and cloaking layer around the condensate, thus gradually leading to drop pinning on the underlying rough topography. Condensation heat transfer further deteriorates in the presence of non-condensable gases (NCGs) requiring special experimental arrangements to eliminate NCGs due to a decrease in the availability of nucleation sites. To address these issues while simultaneously improving heat-transfer performance of LIS in condensation-based systems, we report fabrication of both fresh LIS and a lubricant-depleted LIS using silicon porous nanochannel wicks as an underlying substrate. Strong capillarity in the nanochannels helps retain silicone oil (polydimethylsiloxane) on the surface even after it is severely depleted under tap water. The effect of oil viscosity was investigated for drop mobility and condensation heat transfer under ambient conditions, i.e., in the presence of NCGs. While fresh LIS prepared using 5 cSt silicone oil exhibited a low roll-off angle (∼1°) and excellent water drop (5 μL) sliding velocity ∼66 mm s-1, it underwent rapid depletion as compared to higher viscosity oils. Condensation performed on depleted nanochannel LIS with higher viscosity oil (50 cSt) resulted in a heat-transfer coefficient (HTC) of ∼2.33 kW m-2 K-1, which is a ∼162% improvement over flat Si-LIS (50 cSt). Such LIS promote fast drop shedding as is evident from the little change in the fraction of drops with diameter -2 K-1 was achieved over the last 2 days. The ability of reported LIS to maintain long-term hydrophobicity and dropwise condensation will aid in designing condensation-based systems with improved heat-transfer performance.
- Open Access
- 7 April 2022
- Published by American Chemical Society (ACS) in Energy & Fuels
- Vol. 36 (8) , 4549-4556
Thermal Management of Photovoltaics Using Porous Nanochannels
- 7 April 2022
- Published by American Chemical Society (ACS) in Energy & Fuels
- Vol. 36 (8) , 4549-4556
Open AccessThermal Management of Photovoltaics Using Porous Nanochannels
- Open Access
- 30 August 2021
- Published by American Chemical Society (ACS) in Nano Letters
- Vol. 21 (18) , 7769-7774
Disjoining Pressure of Water in Nanochannels
- 30 August 2021
- Published by American Chemical Society (ACS) in Nano Letters
- Vol. 21 (18) , 7769-7774
Open AccessDisjoining Pressure of Water in Nanochannels
- 1 April 2021
- Published by AIP Publishing in Physics of Fluids
- Vol. 33 (4) , 042007
Nano-confinement effects on liquid pressure
In this study, molecular dynamics simulations are performed to estimate the equilibrium pressure of liquid confined in nanopores. The simulations show that pressure is highly sensitive to the pore size and can significantly change from absolute positive to absolute negative values for a very small (0.1 nm) change in the pore size. The contribution from the solid–liquid interaction always dominates the pressure in the first liquid layer adjacent to the surface and the sensitiveness of pressure on the pore size is dependent on the atom distribution in the liquid layers. A surface influence number is introduced to quantitatively characterize the degree of the confinement. At constant system temperature, the number decreases with increasing pore size based on a power-law function. In nanopores with large S number, the pore liquid pressure is found to be independent of bulk liquid pressure, whereas in nanopores with small S number, the pore pressure is dependent and increases with bulk pressure.- 1 April 2021
- Published by AIP Publishing in Physics of Fluids
- Vol. 33 (4) , 042007
Nano-confinement effects on liquid pressure
In this study, molecular dynamics simulations are performed to estimate the equilibrium pressure of liquid confined in nanopores. The simulations show that pressure is highly sensitive to the pore size and can significantly change from absolute positive to absolute negative values for a very small (0.1 nm) change in the pore size. The contribution from the solid–liquid interaction always dominates the pressure in the first liquid layer adjacent to the surface and the sensitiveness of pressure on the pore size is dependent on the atom distribution in the liquid layers. A surface influence number is introduced to quantitatively characterize the degree of the confinement. At constant system temperature, the number decreases with increasing pore size based on a power-law function. In nanopores with large S number, the pore liquid pressure is found to be independent of bulk liquid pressure, whereas in nanopores with small S number, the pore pressure is dependent and increases with bulk pressure.- 28 December 2020
- Published by American Chemical Society (ACS) in ACS Applied Materials & Interfaces
- Vol. 13 (1) , 1853-1860
Droplet Evaporation on Porous Nanochannels for High Heat Flux Dissipation
Droplet wicking and evaporation in porous nanochannels is experimentally studied on a heated surface at temperatures ranging from 35 to 90 °C. The fabricated geometry consists of cross-connected nanochannels of height 728 nm with micropores of diameter 2 μm present at every channel intersection; the pores allow water from a droplet placed on the top surface to wick into the channels. Droplet volume is also varied, and a total of 16 experimental cases are conducted. Wicking characteristics such as wicked distance, capillary pressure, viscous resistance, and propagation coefficients are obtained at all surface temperatures. Evaporation flux from the nanochannels/micropores is estimated from the droplet experiments but is also independently confirmed via a new set of experiments where water is continuously fed to the sample through a microtube so that it matches the evaporation rate. Heat flux as high as ∼294 W/cm2 is achieved from channels and pores. The experimental findings are applied to evaluate the use of porous nanochannel geometry in spray cooling application and is found to be capable of passively dissipating high heat fluxes upto ∼77 W/cm2 at temperatures below nucleation, thus highlighting the thermal management potential of the fabricated geometry.
- 28 December 2020
- Published by American Chemical Society (ACS) in ACS Applied Materials & Interfaces
- Vol. 13 (1) , 1853-1860
Droplet Evaporation on Porous Nanochannels for High Heat Flux Dissipation
Droplet wicking and evaporation in porous nanochannels is experimentally studied on a heated surface at temperatures ranging from 35 to 90 °C. The fabricated geometry consists of cross-connected nanochannels of height 728 nm with micropores of diameter 2 μm present at every channel intersection; the pores allow water from a droplet placed on the top surface to wick into the channels. Droplet volume is also varied, and a total of 16 experimental cases are conducted. Wicking characteristics such as wicked distance, capillary pressure, viscous resistance, and propagation coefficients are obtained at all surface temperatures. Evaporation flux from the nanochannels/micropores is estimated from the droplet experiments but is also independently confirmed via a new set of experiments where water is continuously fed to the sample through a microtube so that it matches the evaporation rate. Heat flux as high as ∼294 W/cm2 is achieved from channels and pores. The experimental findings are applied to evaluate the use of porous nanochannel geometry in spray cooling application and is found to be capable of passively dissipating high heat fluxes upto ∼77 W/cm2 at temperatures below nucleation, thus highlighting the thermal management potential of the fabricated geometry.
- 12 June 2020
- Published by American Chemical Society (ACS) in Langmuir
- Vol. 36 (27) , 7801-7807
Evaporation Dynamics in Buried Nanochannels with Micropores
Cross-connected buried nanochannels of height ∼728 nm, with micropores of ∼2 μm diameter present at each intersection, are used in this work to numerically and experimentally study droplet-coupled evaporation dynamics at room temperature. The uniformly structured channels/pores, along with their well-defined porosity, allow for computational fluid dynamics simulations and experiments to be performed on the same geometry of samples. A water droplet is placed on top of the sample causing water to wick into the nanochannels through the micropores. After advancing, the meniscus front stabilizes when evaporation flux is balanced with the wicking flux, and it recedes once the water droplet is completely wicked in. Evaporation flux at the meniscus interface of channels/pores is estimated over time, while the flux at the water droplet interface is found to be negligible. When the meniscus recedes in the channels, local contact line regions are found to form underneath the pores, thus rapidly enhancing evaporation flux as a power-law function of time. Temporal variation of wicking flux velocity and pressure gradient in the nanochannels is also independently computed, from which the viscous resistance variation is estimated and compared to the theoretical prediction.
- 12 June 2020
- Published by American Chemical Society (ACS) in Langmuir
- Vol. 36 (27) , 7801-7807
Evaporation Dynamics in Buried Nanochannels with Micropores
Cross-connected buried nanochannels of height ∼728 nm, with micropores of ∼2 μm diameter present at each intersection, are used in this work to numerically and experimentally study droplet-coupled evaporation dynamics at room temperature. The uniformly structured channels/pores, along with their well-defined porosity, allow for computational fluid dynamics simulations and experiments to be performed on the same geometry of samples. A water droplet is placed on top of the sample causing water to wick into the nanochannels through the micropores. After advancing, the meniscus front stabilizes when evaporation flux is balanced with the wicking flux, and it recedes once the water droplet is completely wicked in. Evaporation flux at the meniscus interface of channels/pores is estimated over time, while the flux at the water droplet interface is found to be negligible. When the meniscus recedes in the channels, local contact line regions are found to form underneath the pores, thus rapidly enhancing evaporation flux as a power-law function of time. Temporal variation of wicking flux velocity and pressure gradient in the nanochannels is also independently computed, from which the viscous resistance variation is estimated and compared to the theoretical prediction.
- 19 April 2020
- Published by American Chemical Society (ACS) in The Journal of Physical Chemistry Letters
- Vol. 11 (9) , 3637-3641
Transpiration Mechanism in Confined Nanopores
In this work, molecular dynamics simulations show that liquid in a nanopore can be at thermodynamically-stable high pressure even when connected to conventional bulk liquid. Such high pressure is associated with strong surface-liquid interaction. Evaporation of liquid in the pore creates a flow from low pressure (bulk) region to the high pressure (nanopore) region. Such a counterintuitive flow occurs due to pressure being reduced in the pore from its thermodynamically-stable state. The transition from high pressures to negative pressures in thin liquid films is also studied. This work provides insight into the possible mechanism of passive liquid transport in tall trees such as redwoods.
- 19 April 2020
- Published by American Chemical Society (ACS) in The Journal of Physical Chemistry Letters
- Vol. 11 (9) , 3637-3641
Transpiration Mechanism in Confined Nanopores
In this work, molecular dynamics simulations show that liquid in a nanopore can be at thermodynamically-stable high pressure even when connected to conventional bulk liquid. Such high pressure is associated with strong surface-liquid interaction. Evaporation of liquid in the pore creates a flow from low pressure (bulk) region to the high pressure (nanopore) region. Such a counterintuitive flow occurs due to pressure being reduced in the pore from its thermodynamically-stable state. The transition from high pressures to negative pressures in thin liquid films is also studied. This work provides insight into the possible mechanism of passive liquid transport in tall trees such as redwoods.