Repository for Publications and Research Data
News from the ETH Library
Recently Added
FLINT: Efficiently Exploiting High-Bandwidth Flash for Capacity-Scalable LLM Inference Acceleration
Item type: Journal Article
Oliveira G.F.; Zhu X.; Yuzuguler A.C.; et al. (2026)
Large language model (LLM) inference is increasingly constrained by accelerator memory capacity rather than compute throughput. High bandwidth flash (HBF) is an emerging 3D-stacked NAND flash technology that provides multi-terabyte near-accelerator capacity, making it a promising capacity tier for storing large language model (LLM) weights. However, existing high-bandwidth flash (HBF)-based proposals face three adoption challenges: they (i) rely on coarse-grained static prefetching, (ii) expose NAND flash refresh to the accelerator-visible critical path, and (iii) assume solid-state drive (SSD)-class flash-management mechanisms that are poorly matched to a read-only HBF weight tier. Our goal is to design an efficient HBF substrate that addresses these three challenges. To this end, we propose FLINT (flash inference tier), a workload-driven HBF substrate for capacity-scalable LLM inference. FLINT introduces three mechanisms: (i) a hardware burst-buffer controller, which pipelines HBF reads with accelerator-side memory requests using existing NAND flash page/cache buffers, (ii) a phantom-plane refresh mechanism, which removes refresh from the LLM decode critical path, and (iii) a read-only flash translation layer (FTL), which replaces SSD-class support for arbitrary writes with a compact table that translates logical weight bursts to physical HBF locations. We evaluate FLINT across dense and mixture-of-experts (MoE) LLM inference models. Our evaluation shows that FLINT improves LLM decode throughput by 355×, 3.0×, and 1.17× and reduces energy consumption by 294×, 2.5×, and 1.14× compared to an SSD-equipped graphics processing unit (GPU) system, an HBM-only GPU system, and a prior hybrid HBM+HBF GPU system, respectively.
Electromagnetic shower reconstruction and identification in FASER's emulsion detector for LHC forward neutrino measurements
Item type: Journal Article
Abraham R.M.; Ai X.; Monsalve S.A.; et al. (2026)
We present methods for electromagnetic shower reconstruction and identification in the FASERν emulsion detector using 100 GeV and 200 GeV electron test-beam data from the CERN SPS H4 beamline. The reconstruction employs a clustering-based algorithm without energy-dependent tuning to determine shower axes. A multi-level identification chain comprising track pre-selection, a cut-based selection, and a BDT classifier achieves combined background rejection rates of 99.99% (100 GeV) and 99.94% (200 GeV). The method reaches total reconstruction and identification efficiencies of 58.9% (100 GeV) and 70.8% (200 GeV) evaluated from simulated samples. Energy reconstruction using the total number of reconstructed segments as the calorimetric estimator yields relative biases of +0.6% (100 GeV) and -0.8% (200 GeV), with resolutions of 25.4% and 22.6%, respectively. Systematic uncertainties on the energy reconstruction are dominated by variations in emulsion film detection efficiency, with totals of +10.9-8.2% at 100 GeV and +10.3-6.9% at 200 GeV. The methodology provides a validated framework for electron neutrino identification with the FASERν detector at the LHC.
Whistling of turbulent cavity flows: self-consistent predictions with the linearised compressible Navier–Stokes equations
Item type: Journal Article
Bozikis N.; Hollweck T.; Özev D.; et al. (2026)
A self-consistent saturation model for the prediction of aeroacoustic limit cycles emerging in turbulent low-Mach cavity flows (Reynolds number (Formula presented), Mach number (Formula presented)) is proposed. It predicts the nonlinear interactions between the acoustic modes of a deep rectangular cavity and the hydrodynamic instabilities of the turbulent shear layer that forms over its opening due to the presence of a grazing flow. The model is based on the triple decomposition of the flow variables and the compressible Navier–Stokes equations. At each step of the iterative process, the nonlinear eigenvalue problem associated with perturbations around the mean flow is updated with the steady component of the forcing from the unstable eigenmode’s Reynolds stress. The iterations are performed until the dominant eigenmode becomes marginally stable, i.e. its growth rate vanishes. The evolution of the coherent velocity fluctuations as a function of the oscillation amplitude is in good qualitative agreement with previously published compressible large eddy simulations. Furthermore, the predictions of the frequency and amplitude of the aeroacoustic limit-cycle oscillations are validated against the ones obtained from a low-order model, whose parameters were adjusted to reproduce the experimental measurements of the deep cavity whistling.
MINDS: Intertwined evolution of dust and gas in large planet-forming disks: Diversity driven by halted pebble drift?
Item type: Journal Article
Tabone B.; Temmink M.; Waters L.B.F.M.; et al. (2026)
Context. The JWST gives unique access to the chemical and physical conditions in inner disks (<10 au), where the majority of detected exoplanets are thought to form. Aims. Our goal is to investigate the diversity of inner disks with a specific focus on large and massive disks around Sun-like stars. These are thought to be the progenitors of planetary systems with wide-orbit planets and potential instances of disks with quenched pebble drift. Methods. We analyzed the MIRI-MRS spectra of three disks from the MINDS program that orbit young Sun-like stars (≃0.8 M⊙): V1094 Sco, DL Tau, and IM Lup. They were placed in context with the MINDS sample and millimeter observations. Results. The JWST spectra reveal a striking diversity that agrees with previous observations. V1094 Sco and DL Tau exhibit the highest C2H2/H2O flux ratio of the MINDS sample of full T Tauri disks (M∗ > 0.4M⊙). In V1094 Sco, even cold C4H2 is detected. In contrast, the IM Lup spectrum is dominated by O-bearing species. No one-to-one correspondence is found between the gas in the outer disk, as traced by the C2H (3–2)/C18O (2–1) flux ratios, and that of the inner disk, as traced by the C2H2/H2O flux ratios. To account for these observational results, we propose a scenario based on a toy model of halted pebble drift relevant to the pebble-rich disks. We show that a volatile C/O ratio close to unity and low elemental C and O abundance can be achieved in inner disks only if (1) about 95% of the icy grains are blocked in the outer disk, (2) the outer disk is chemically evolved to reach elevated gas-phase C/O and low (C+O)/H abundance ratio, and (3) the gas in the outer disk had time to be advected to the inner disk. In this scenario, DL Tau and perhaps V1094 Sco would be the rare examples for which all these conditions are met. Therefore, a high C2H2/H2O flux ratio in pebble-rich disks would have a different origin than that proposed for very low-mass stars, for which fast drift of oxygen-rich pebbles would eventually leave a carbon-rich inner disk. Destruction of carbonaceous grains appears to be a less compelling scenario to account for the high C2H2/H2O flux ratio in DL Tau and V1094 Sco because existing models require strong pebble drift and low accretion rates to enrich the inner disk in volatile carbon. Alternatively, the mid-IR molecular features might be related to the low optical depth of the dust, which would reveal a normally hidden reservoir of C2H2 deeper down to the midplane. Finally, we show for the first time that the disks with a high C2H2/H2O flux ratio exhibit a prominent silica dust component, a result found in four disks published so far (V1094 Sco, DL Tau, CY Tau, and DoAr 33). We propose that the re-formation of dust at the sublimation front of silicates in a gas with super-solar (but below unity) C/O ratio leads to a silica stoichiometry (SiO2). In turn, silica constitutes a promising diagnostic of the C/O ratio in the inner disks. Conclusions. The diversity of inner disks as revealed by mid-IR spectroscopy might be driven by the joint evolution of the outer versus inner disk via chemical evolution of the cold outer disk and radial redistribution of gas and ice-coated dust. Studies gathering large samples and forward modeling are required to confirm this proposal and obtain a global picture of disk evolution.
Hop-and-sink locomotion and escape in benthic copepods under localized suction flow
Item type: Journal Article
Kastinger M.; Brizzolara S.; Redaelli T.; et al. (2026)
Benthic copepods inhabit complex boundary layers where flow varies rapidly, yet their sensory-motor responses to localized hydrodynamic signals remain poorly understood. We quantified the behavior of the harpacticoid Tigriopus californicus under no-flow and localized suction flow using 3D particle tracking velocimetry of both copepod trajectories and the surrounding flow. Copepods frequently entered the water column and maintained a hop-and-sink gait, spending ~40% of time jumping with peak speeds up to 68 mm s−1. In the far field (>2 mm from the source), kinematic statistics were indistinguishable from no-flow controls and orientations were consistent with random swimming. In the near field, individuals performed directed escape jumps (mean 80.6 ± 17.9 mm s−1; max. 105.7 mm s−1). Escape onset occurred at relative speeds (≈ 6 mm s−1) and strain rates (≈ 5 s−1) characteristic of localized deformation, at distances from the source on the order of the organism size. These results reveal a context-dependent modulation of locomotion that preserves routine hop-and-sink gait under weak flow while enabling rapid evasion near localized disturbances. The findings are consistent with gradient-based mechanosensation, although direct contact with the suction inlet may have contributed to some escapes. Broader generalization will require replication across taxa, life stages and flow regimes.