The Quest for Reliable AI Accelerators: Cross-Layer Evaluation and Design Optimization

Meng Li, Tong Xie, Zuodong Zhang, Runsheng Wang · 2026

Read The Quest for Reliable AI Accelerators: Cross-Layer Evaluation and Design Optimization by Meng Li, Tong Xie, Zuodong Zhang, Runsheng Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14148.txt.

Introduction

Meng Li, Tong Xie, Zuodong Zhang, and Runsheng Wang tackle aging and variation in AI chips in 2026. As CMOS technology pushes to the nanoscale, aging effects grow more pronounced. Process variations also increase, posing significant reliability challenges for AI accelerators. Traditional guardband design uses pessimistic timing margins for safety.

Those margins sacrifice significant performance and computational efficiency. They render older approaches inadequate for high-performance AI computing demands. The team pursues cross-layer evaluation and design optimization for reliable accelerators. Chips that age with grace would serve data centers for longer. The full-stack view joins chip wear to code and task plans. That join is where most of the lost speed can be won back.

Catalog record

CREATORMeng Li, Tong Xie, Zuodong Zhang, Runsheng Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE16 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-14148

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The Quest for Reliable AI Accelerators: Cross-Layer Evaluation and Design Optimization Thanks: This work was supported in part by NSFC (62495102, 92464104, 62125401), National Key Research and Development Program (2024YFB4505004), Beijing Municipal Science and Technology Program (Z241100004224015), Beijing Outstanding Young Scientist Program (JWZQ20240101004), and the 111 Project (B18001). Thanks: *Corresponding author: {meng.li,r.wang}@pku.edu.cn Affiliation: Institute for Artificial Intelligence & School of Integrated Circuits, Peking University, Beijing, China Meng Li, Tong Xie, Zuodong Zhang, Runsheng Wang Abstract As the CMOS technology pushes to the nanoscale, aging effects and process variations have become increasingly pronounced, posing significant reliability challenges for AI accelerators. Traditional guardband-based design approaches, which rely on pessimistic timing margin, sacrifice significant performance and computational efficiency, rendering them inadequate for high-performance AI computing demands. Current reliability-aware AI accelerator design faces two core challenges: (1) the lack of systematic cross-layer analysis tools to capture coupling reliability effects across device, circuit, architecture, and application layers; and (2) the fundamental trade-off between conventional reliability optimization and computational efficiency. To address these challenges, this paper systematically presents a series of reliability-aware accelerator designs, encompassing (1) aging and variation-aware dynamic timing analyzer, (2) accelerator dataflow optimization using critical input pattern reduction, and (3) resilience characterization and novel architecture design for large language models (LLMs). By tightly integrating cross-layer reliability modeling and AI workload characteristics, these co-optimization approaches effectively achieve reliable and efficient AI acceleration. I Introduction Recently, AI has emerged as a game-changing technology and has revolutionized many different applications, including reliability-critical tasks such as autonomous driving etc. With high computational requirements, AI workloads, particularly large language models (LLMs), have been widely deployed on customized accelerators such as TPU-like systolic arrays. However, as Moore’s law pushes to the nanoscale, these accelerators are susceptible to hardware faults. These faults, including timing errors, arise from various reliability issues, such as process, voltage, temperature, and aging (PVTA) variations [ 5 , 7 , 8 , 6 ] . Therefore, it is crucial to enable both reliable and efficient AI acceleration. Fig. 1: Overview of our recent works. However, ensuring reliability in AI accelerators presents significant challenges. First, reliability issues exhibit strong cross-layer coupling and lack accurate modeling. For example, at the device level, voltage stress induces threshold-voltage shifts ( \Delta V_{\rm th} ); at the circuit level, these shifts manifest as specific timing-error patterns and timing error rates (TERs); and at the application level, such errors degrade AI model accuracy (Acc). These relationships are highly complex and difficult to model precisely. Second, existing mitigation methods fail to balance reliability and efficiency. Worst-case voltage margins eliminate timing errors but incur heavy overhead, while reduced margins increase bit error rates (BERs) and degrade model performance. Classical techniques such as Razor flip-flops [ 3 ] and algorithm-based fault tolerance (ABFT) [ 4 ] are either unscalable to large accelerators or incur frequent yet unnecessary recovery, making them impractical for compute-intensive AI workloads. To address these challenges, this paper summarize our recent efforts for cross layer evaluation and design optimization techniques [ 10 , 11 , 9 ] , as depicted in Fig. 1 . II Aging- and Variation-Aware DTA II-A Motivation As technology scaling amplifies timing guardbands, better-than-worst-case (BTWC) design seeks to improve efficiency by exploiting dynamic timing margins. Conventional dynamic timing analysis (DTA) tools add worst-case aging and variation guardbands, inflating margins and diminishing BTWC gains. AVATAR [ 10 ] addresses this by incorporating transistor aging and process variation directly into DTA, eliminating unnecessary guardbands and preserving performance and efficiency. II-B AVATAR Algorithm Fig. 2: Aging- and variation-aware DTA flow of AVATAR. The core idea of AVATAR is to integrate aging analysis into the DTA process, as illustrated in Fig. 2 . The algorithm consists of three main steps: (1) gate-level aging and variation model characterization, (2) workload analysis to estimate \Delta V_{\rm th} for each transistor, and (3) event-based DTA. In Step 1, the aging model takes as inputs the timing arc, input slew, output load, per-transistor \Delta V_{\rm th} , V_{\mathrm{DD}} , and temperature, and outputs aged cell delay and transition time using a first-order Taylor expansion. The variation model is based on POCV analysis with LVF data. In Step 2, zero-delay gate-level simulation is performed to obtain the toggle rate of each internal net, after which the aging model is applied to compute the \Delta V_{\rm th} of each transistor. In Step 3, events are defined as digital switching activities on pins. Event-based DTA begins by constructing a timing graph for event propagation, followed by cycle-by-cycle analysis that includes input event generation, event propagation, and path reporting, with the arrival times at all timing endpoints recorded. II-C Experiments TABLE I: Performance improvement from application-based DVFS based on the corner-based DTA and AVATAR Benchmark Corner-based DTA [ 1 , 2 ] AVATAR (this work) Max Freq. (MHz) Impro. (vs STA) Max Freq. (MHz) Impro. (vs STA) SHA 948 13.75% 1020 22.38% AES_CBC 883 5.99% 951 14.10% FIR 915 9.82% 986 18.35% BubbleSort 1290 55.38% 1380 65.36% Motion_Detection 958 15.00% 1030 23.97% CNN 868 4.18% 936 12.30% Convolution 868 4.19% 936 12.28% 2d_Filter 936 12.33% 1050 26.37% MatrixMult 916 9.89% 989 18.63% DCT 1170 40.77% 1270 52.15% To evaluate the benefits of AVATAR, we determine the application-specific V_{\min} / f_{\max} using two methods. (1) Corner-based DTA with extra guardbands [ 1 , 2 ] . Dynamic delay is computed as delay\times(1+total\_guardband) , where the aging guardband is assumed to be 15% and the random variation guardband 5% at nominal V_{\mathrm{DD}} . The trend of guardband versus V_{\mathrm{DD}} is characterized using FO4 delay as a representative cell. (2) AVATAR-based analysis. AVATAR inherently accounts for both aging and variation, eliminating the need for extra guardbands. The final delay is calculated as \mu(delay)+3\sigma(delay) . Table I shows the performance improvement by applying the maximum frequency at the nominal 0.8V. III Critical Input Pattern Reduction III-A Computing Sequence and Timing Error READ [ 11 ] proposes a reliability-enhanced accelerator dataflow optimization technique that can effectively reduce timing errors. For a given cycle, timing errors are primarily determined by two factors: (1) the input pattern, which dictates the activated paths, and (2) the operating conditions (e.g., temperature, voltage, aging) that influence the delay of these paths. While operating conditions are often uncontrollable, only a small subset of paths, i.e., critical paths, typically exceed the clock period. Thus, reducing the activation rate of these critical paths is key to lowering the timing error rate. Fig. 3: A 1 \times 4 convolution calculated in different orders. Reordering weights does not change the computing result, but avoids the critical input pattern. The basic processing element (PE) in our study is a multiplier-and-accumulator (MAC) unit. Using the 8-bit multiplier and 24-bit accumulator of a TPU MAC as an example, dynamic timing analysis of the synthesized unit reveals that the most frequent critical input patterns are those that flip the sign bit of the partial sum (PSUM). For instance, computing 3\times(-2)+2=-4 produces the 2’s complement encoding 111111111111111111111100 , where the sign bit flip triggers a long carry chain in the accumulator, activating critical paths. A similar effect occurs when PSUM transitions from negative to positive. Consequently, reducing the PSUM sign-flip rate is an effective strategy for minimizing timing error rates.

Read the complete text