Repository for Publications and Research Data
News from the ETH Library
Recently Added
Late Breaking Results: Structured Expert Routing for Efficient MoE Inference under GPU-CPU Orchestration
Item type: Conference Paper
Chen Y.; Akbari A.; Akbari A.; et al. (2026)
Hybrid GPU-CPU deployment of large Mixture-of-Experts (MoE) models is often latency-bound by CPU-side expert execution and cross-device orchestration overhead. To overcome this limitation, we propose a structured expert routing strategy, implemented via asymmetric expert skipping, which reduces expert activation in latency-critical layers while applying lightweight magnitude compensation and norm calibration to preserve model accuracy. Experiments on quantized Qwen3 (30B and 235B) models show that our method improves end-to-end generation throughput by up to 60% while achieving 3.6% higher accuracy than uniform expert reduction under the same compute budget.
MATCHA: Efficient Deployment of Deep Neural Networks on Multi-Accelerator Heterogeneous Edge SoCs
Item type: Conference Paper
Russo E.; Hamdi M.A.; Ottaviano A.; et al. (2026)
Deploying DNNs on System-on-Chips (SoC) with multiple heterogeneous acceleration engines is challenging, and the majority of deployment frameworks cannot fully exploit heterogeneity. We present MATCHA, a unified DNN deployment framework that generates highly concurrent schedules for parallel, heterogeneous accelerators and uses constraint programming to optimize L3/L2 memory allocation and scheduling. Using pattern matching, tiling, and mapping across individual HW units enables parallel execution and high accelerator utilization. On the MLPerf Tiny benchmark, using a SoC with two heterogeneous accelerators, MATCHA improves accelerator utilization and reduces inference latency by up to 35% with respect to the the state-of-the-art MATCH compiler.
Network Design for Wafer-Scale Systems with Wafer-on-Wafer Hybrid Bonding
Item type: Conference Paper
Iff P.; Bonato T.; Besta M.; et al. (2026)
Transformer-based large language models are increasingly constrained by data movement as communication bandwidth drops sharply beyond the chip boundary. Wafer-scale integration using wafer-on-wafer hybrid bonding alleviates this limitation by providing ultra-high bandwidth between reticles on bonded wafers. In this paper, we investigate how the physical placement of reticles on wafers influences the achievable network topology and the resulting communication performance. Starting from a 2D mesh-like baseline, we propose four reticle placements (Aligned, Interleaved, Rotated, and Contoured) that improve throughput by up to 250%, reduce latency by up to 36%, and decrease energy per transmitted byte by up to 38%.
Towards Always-on Interaction: A Sub-mW Visual Wakeup Acceleration Subsystem for Smart Glasses
Item type: Conference Paper
Prasad A.S.; Mazzola S.; Durrer C.; et al. (2026)
Always-on vision is essential for smart glasses, yet continuous visual processing under stringent power budgets remains a major challenge. This work presents an always-on visual wake-up acceleration engine for sub-milliwatt hand gesture recognition. A compact convolutional neural network (<13k parameters), trained on the HaGRID dataset, performs binary classification on 64×64 inputs, achieving over 92% accuracy at 3 b precision and requiring less than 9 kB of memory. When executed on our flexible accelerator engine implemented in TSMC 7 nm technology, at 100 MHz, it consumes only 64 nJ per frame, translating to an always-on power of 11 μW at 30 FPS, enabling energy-efficient, always-on interaction for next-generation smart-glasses.
Stitch: Assertion-Guided Patching of On-Chip Protocol Implementations using LLMs
Item type: Conference Paper
Zonta-Roudes M.; Hinderling N.; Shinde S. (2026)
Verification flows use Verification IPs (VIPs) to identify assertion violations. This is well-suited for on-chip protocols (e.g., AXI, AHB) with standard specifications (e.g., AMBA) for developers to write assertions that detect violations. Patching these violations, however, still requires manual effort. A good patch must not only fix the violation but also preserve functionality (e.g., pass the testsuite) without causing new violations. We propose Stitch, based on the intuition that given an implementation, a violated assertion, and a counterexample an LLM can try to synthesize patches. To evaluate Stitch, we develop a new dataset that comprises 100 violations in 11 implementations of 5 protocols (AXI, AHB, APB, Wishbone, and TileLink). Our first experiment reports 19% success rate across 4 LLMs (GPT4, GPT5, Gemini, and Claude). By analyzing the failed cases, we devise 3 improvement strategies: patch localization, violation specific context (cone of influence, counterexample), iterative feedback from model checker outputs over candidate patches. Our experiments show these strategies help Stitch achieve 61% patch success, with GPT5 dominating with 56%. We validate Stitch patches through simulation and on real hardware. We compare Stitch to 3 state-of-the-art non-LLM tools and existing patches.
