Lieven Eeckhout
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Lieven Eeckhout has a total of 108 co-authors. Here are the top co-authors by number of shared publications:
- Magnus Jahre
Norwegian University of Science and Technology
10 shared publications - Ajeya Naithani
Ghent University, Ghent, Belgium
9 shared publications - Jaime Roelandts
Ghent University, Ghent, Belgium
4 shared publications - Josue Feliu
Department of Computer Engineering (DISCA), Universitat Politecnica de Valencia, Valencia, Valencia Spain
4 shared publications - Sam Ainsworth
University of Edinburgh, Edinburgh, U.K.
4 shared publications - Timothy M. Jones
University of Cambridge, Cambridge, United Kingdom
4 shared publications
Latest Publications
- Open Access
- 26 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (4) , 1-25
The GPU Resilience Lottery: Understanding and Taming Real-World Hardware Errors in Heterogeneous AI Infrastructure
The booming LLM demand is driving a rapid increase in the scale and heterogeneity of modern AI infrastructure. To achieve high availability for sustained LLM workloads, infrastructure providers must contend with frequent and severe hardware errors. Without a holistic resilience understanding, providers may inadvertently schedule AI jobs onto GPUs with high error risks and mitigate errors with significant costs, further degrading cluster availability. As a result, AI jobs are forced to ‘win’ a GPU resilience lottery for interruption-free execution. However, betting on a winning ticket, i.e., selecting resilient GPU hardware and adopting efficient mitigations, requires a cross-layer understanding of error heterogeneity and severity, which remains opaque to current fault injection and fleet distribution studies. In this article, we present a quantitative GPU resilience analysis in production clusters with 500+K GPUs, focusing on the multi-faceted heterogeneity and severity. We characterize distinct hardware errors via spatial patterns across distinct architectures and components, and temporal patterns correlated with deployment age. Furthermore, we quantify error severity across both hardware-layer outage, software-layer anomalies, and system-layer metrics. These insights lead to the design of LotteRig , a proactive availability enhancement framework that performs risk-aware GPU selection and cost-aware error mitigation. LotteRig achieves 9.9% fewer software interruptions and 12.1% higher hardware availability across diverse scenarios, substantially improving the odds of ‘winning’ the GPU resilience lottery.
- 26 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 23 (4) , 1-25
Open AccessThe GPU Resilience Lottery: Understanding and Taming Real-World Hardware Errors in Heterogeneous AI Infrastructure
The booming LLM demand is driving a rapid increase in the scale and heterogeneity of modern AI infrastructure. To achieve high availability for sustained LLM workloads, infrastructure providers must contend with frequent and severe hardware errors. Without a holistic resilience understanding, providers may inadvertently schedule AI jobs onto GPUs with high error risks and mitigate errors with significant costs, further degrading cluster availability. As a result, AI jobs are forced to ‘win’ a GPU resilience lottery for interruption-free execution. However, betting on a winning ticket, i.e., selecting resilient GPU hardware and adopting efficient mitigations, requires a cross-layer understanding of error heterogeneity and severity, which remains opaque to current fault injection and fleet distribution studies. In this article, we present a quantitative GPU resilience analysis in production clusters with 500+K GPUs, focusing on the multi-faceted heterogeneity and severity. We characterize distinct hardware errors via spatial patterns across distinct architectures and components, and temporal patterns correlated with deployment age. Furthermore, we quantify error severity across both hardware-layer outage, software-layer anomalies, and system-layer metrics. These insights lead to the design of LotteRig , a proactive availability enhancement framework that performs risk-aware GPU selection and cost-aware error mitigation. LotteRig achieves 9.9% fewer software interruptions and 12.1% higher hardware availability across diverse scenarios, substantially improving the odds of ‘winning’ the GPU resilience lottery.
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 46 (5) , 87-94
Accounting for Workload Churn in Design Space Exploration
Hundreds of billions may be spent on computing infrastructure in the coming decades. At such scale, it is critical that hardware maintains acceptable performance over its lifetime. Attainable performance depends on the degree of workload churn, i.e., how much the workload changes over time, and the hardware’s ability to adapt. Overly specialized hardware may experience performance degradation when used on a workload with high churn. Conversely, overly general hardware may represent a missed opportunity when used on a workload with little to no churn. Hence, specialization should be tailored to churn expectations. This requires being able to clearly reason about the specifics of workload churn. Our work furnishes developers with an abstract yet powerful framework for reasoning about workload churn in both qualitative and quantitative contexts. Additionally, we present the Neoscope design space exploration tool, which provides insight into how various forms of churn affect the optimal design of System-on-Chip (SoC) architectures.
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 46 (5) , 87-94
Accounting for Workload Churn in Design Space Exploration
Hundreds of billions may be spent on computing infrastructure in the coming decades. At such scale, it is critical that hardware maintains acceptable performance over its lifetime. Attainable performance depends on the degree of workload churn, i.e., how much the workload changes over time, and the hardware’s ability to adapt. Overly specialized hardware may experience performance degradation when used on a workload with high churn. Conversely, overly general hardware may represent a missed opportunity when used on a workload with little to no churn. Hence, specialization should be tailored to churn expectations. This requires being able to clearly reason about the specifics of workload churn. Our work furnishes developers with an abstract yet powerful framework for reasoning about workload churn in both qualitative and quantitative contexts. Additionally, we present the Neoscope design space exploration tool, which provides insight into how various forms of churn affect the optimal design of System-on-Chip (SoC) architectures.
- Open Access
- 20 July 2026
- Published by Association for Computing Machinery (ACM) in ACM SIGMETRICS Performance Evaluation Review
- Vol. 54 (1) , 126-128
cuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
Virtual memory is essential for modern GPUs, yet address translation remains a major bottleneck. In irregular workloads, frequent TLB misses and costly page table walks (PTWs) dominate memory latency. We propose C ompute U nit P age T able W alk (cuPTW), which repurposes idle GPU compute resources to accelerate PTWs. By offloading requests to idle functional units, cuPTW transforms address translation into a massively parallel task. We further optimize this with cuPTW-SW, which caches walks in local data store (LDS) memory, and cuPTW-MT, which parallelizes walks across SIMD lanes. Evaluation shows that cuPTW-Full achieves a 4.43× average speedup (up to 76.09×) and improves PTW throughput by 9.92×. Compared to state-of-the-art designs, Marching Page Walks and SnakeByte, cuPTW-Full delivers a 2.08× and 1.97× performance gain, respectively.
- 20 July 2026
- Published by Association for Computing Machinery (ACM) in ACM SIGMETRICS Performance Evaluation Review
- Vol. 54 (1) , 126-128
Open AccesscuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
Virtual memory is essential for modern GPUs, yet address translation remains a major bottleneck. In irregular workloads, frequent TLB misses and costly page table walks (PTWs) dominate memory latency. We propose C ompute U nit P age T able W alk (cuPTW), which repurposes idle GPU compute resources to accelerate PTWs. By offloading requests to idle functional units, cuPTW transforms address translation into a massively parallel task. We further optimize this with cuPTW-SW, which caches walks in local data store (LDS) memory, and cuPTW-MT, which parallelizes walks across SIMD lanes. Evaluation shows that cuPTW-Full achieves a 4.43× average speedup (up to 76.09×) and improves PTW throughput by 9.92×. Compared to state-of-the-art designs, Marching Page Walks and SnakeByte, cuPTW-Full delivers a 2.08× and 1.97× performance gain, respectively.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 363-366
Establishing Confidence in Average Metrics
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 363-366
Establishing Confidence in Average Metrics
- Open Access
- 8 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 126-128
cuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
- 8 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 126-128
Open AccesscuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
- Open Access
- 29 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Measurement and Analysis of Computing Systems
- Vol. 10 (2) , 1-26
cuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
Virtual memory has become a cornerstone of modern GPUs, enabling unified address spaces and advanced memory management techniques. However, the performance of address translation has emerged as a critical bottleneck, particularly under irregular workloads with massive memory footprints, where frequent TLB misses and costly page table walks dominate total memory access latency. Prior work has primarily focused on improving TLB effectiveness or optimizing page table walks through batching and coalescing, but these approaches remain limited by the lack of locality and memory bandwidth constraints. In this work, we propose Compute Unit Page Table Walk ( cuPTW ), a novel address translation architecture that repurposes idle GPU compute resources to accelerate page table walks. Specifically, we introduce (1) a single-threaded synchronous cuPTW , which offloads translation requests to idle functional units within compute units, and (2) two optimizations that further reduce latency and improve throughput by caching page table walks in local data store memory ( cuPTW-SW ) and parallelizing them across multiple SIMD lanes ( cuPTW-MT ). When combined, cuPTW-Full transforms low-parallelism page table walks into a massively parallel computation task. Our evaluation across 15 representative GPU workloads, including deep learning, graph analytics and scientific simulations, demonstrates that cuPTW-Full achieves a performance speedup of 4.43× on average (up to 76.09×) by improving page table walk throughput by 9.92× on average over our baseline GPU architecture. Compared to state-of-the-art GPU address translation proposals, cuPTW-Full achieves a 1.97× to 2.08× average speedup.
- 29 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Measurement and Analysis of Computing Systems
- Vol. 10 (2) , 1-26
Open AccesscuPTW: Leveraging Idle Compute Units for Massively Parallel GPU Page Table Walks
Virtual memory has become a cornerstone of modern GPUs, enabling unified address spaces and advanced memory management techniques. However, the performance of address translation has emerged as a critical bottleneck, particularly under irregular workloads with massive memory footprints, where frequent TLB misses and costly page table walks dominate total memory access latency. Prior work has primarily focused on improving TLB effectiveness or optimizing page table walks through batching and coalescing, but these approaches remain limited by the lack of locality and memory bandwidth constraints. In this work, we propose Compute Unit Page Table Walk ( cuPTW ), a novel address translation architecture that repurposes idle GPU compute resources to accelerate page table walks. Specifically, we introduce (1) a single-threaded synchronous cuPTW , which offloads translation requests to idle functional units within compute units, and (2) two optimizations that further reduce latency and improve throughput by caching page table walks in local data store memory ( cuPTW-SW ) and parallelizing them across multiple SIMD lanes ( cuPTW-MT ). When combined, cuPTW-Full transforms low-parallelism page table walks into a massively parallel computation task. Our evaluation across 15 representative GPU workloads, including deep learning, graph analytics and scientific simulations, demonstrates that cuPTW-Full achieves a performance speedup of 4.43× on average (up to 76.09×) by improving page table walk throughput by 9.92× on average over our baseline GPU architecture. Compared to state-of-the-art GPU address translation proposals, cuPTW-Full achieves a 1.97× to 2.08× average speedup.
- Open Access
- 22 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 361-376
Chips Need DIP: Time-Proportional Per-Instruction Cycle Stacks at Dispatch
- 22 March 2026
- Published by Association for Computing Machinery (ACM)
- p. 361-376
Open AccessChips Need DIP: Time-Proportional Per-Instruction Cycle Stacks at Dispatch
- 26 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (3) , 1499-1503
An Analytical Model for Performance-Carbon Co-Optimization of Edge AI Accelerators
Application-specified integrated circuit (ASIC) accelerators have emerged as vital solutions to deploy various AI workloads, especially in resource-constrained edge scenarios. Existing analytical models for guiding the architectural exploration of these accelerators focus solely on performance and energy efficiency metrics while overlooking the carbon cost. In contrast, existing carbon cost models focus on analyzing specific hardware configurations but lack the support for exploration of the optimal architecture choices and the tradeoff analysis between performance and carbon cost. To fill this gap, this brief aims to analyze the impact of different architecture configurations from both the performance and carbon perspectives, exploring the tradeoff between the performance and carbon cost for AI accelerator design. For this purpose, we first built an analytical model, named CarbonSpot, capable of modeling and estimating both performance and carbon cost for any accelerator architecture in the design space. Then, by benchmarking the overhead of these AI accelerators under MLPerf-Tiny and MLPerf-Mobile workloads, we show that architectures solely optimized for performance and energy efficiency produce $58\times $ more carbon emissions than designs designed for the highest carbon efficiency. Importantly, co-optimized architectural choices exist, with only <20% drop in performance and <6% overhead in carbon costs when compared to the respective best cases optimized for either maximum performance or minimal carbon impact. The model is open-sourced at: https://github.com/KULeuven-MICAS/carbonspot
- 26 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (3) , 1499-1503
An Analytical Model for Performance-Carbon Co-Optimization of Edge AI Accelerators
Application-specified integrated circuit (ASIC) accelerators have emerged as vital solutions to deploy various AI workloads, especially in resource-constrained edge scenarios. Existing analytical models for guiding the architectural exploration of these accelerators focus solely on performance and energy efficiency metrics while overlooking the carbon cost. In contrast, existing carbon cost models focus on analyzing specific hardware configurations but lack the support for exploration of the optimal architecture choices and the tradeoff analysis between performance and carbon cost. To fill this gap, this brief aims to analyze the impact of different architecture configurations from both the performance and carbon perspectives, exploring the tradeoff between the performance and carbon cost for AI accelerator design. For this purpose, we first built an analytical model, named CarbonSpot, capable of modeling and estimating both performance and carbon cost for any accelerator architecture in the design space. Then, by benchmarking the overhead of these AI accelerators under MLPerf-Tiny and MLPerf-Mobile workloads, we show that architectures solely optimized for performance and energy efficiency produce $58\times $ more carbon emissions than designs designed for the highest carbon efficiency. Importantly, co-optimized architectural choices exist, with only <20% drop in performance and <6% overhead in carbon costs when compared to the respective best cases optimized for either maximum performance or minimal carbon impact. The model is open-sourced at: https://github.com/KULeuven-MICAS/carbonspot
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1296-1310
Neoscope: How Resilient Is My SoC to Workload Churn?
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1296-1310
Open AccessNeoscope: How Resilient Is My SoC to Workload Churn?
- 9 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (4) , 72-78
Scalar Vector Runahead: Removing the Shackles of Indirect Memory Chains on In-Order Cores
Modern processors often face the memory wall as a significant bottleneck, a problem that becomes particularly severe when using stall-on-use in-order cores. Despite this limitation, there is growing demand for energy-efficient in-order cores due to privacy and sustainability concerns. Scalar Vector Runahead (SVR)1 provides an elegant solution by extracting high memory-level parallelism through piggybacking on existing instructions executed on the processor that lead to future irregular memory accesses. SVR speculatively executes multiple transient, independent, parallel instances of memory accesses and their instruction chains, by initiating memory accesses from many different values of a predicted induction variable. This approach moves mutually independent memory accesses next to each other to hide dependent stalls. With a hardware overhead of only 2 KiB and without the need for hardware vector extensions, SVR delivers 3.2× higher performance than a baseline 3-wide in-order core inspired by an Arm Cortex A510, and 1.3× higher performance than a full out-of-order core, while halving energy consumption.
- 9 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (4) , 72-78
Scalar Vector Runahead: Removing the Shackles of Indirect Memory Chains on In-Order Cores
Modern processors often face the memory wall as a significant bottleneck, a problem that becomes particularly severe when using stall-on-use in-order cores. Despite this limitation, there is growing demand for energy-efficient in-order cores due to privacy and sustainability concerns. Scalar Vector Runahead (SVR)1 provides an elegant solution by extracting high memory-level parallelism through piggybacking on existing instructions executed on the processor that lead to future irregular memory accesses. SVR speculatively executes multiple transient, independent, parallel instances of memory accesses and their instruction chains, by initiating memory accesses from many different values of a predicted induction variable. This approach moves mutually independent memory accesses next to each other to hide dependent stalls. With a hardware overhead of only 2 KiB and without the need for hardware vector extensions, SVR delivers 3.2× higher performance than a baseline 3-wide in-order core inspired by an Arm Cortex A510, and 1.3× higher performance than a full out-of-order core, while halving energy consumption.