IBM Research - Zurich

Abstract Analog in-memory computing (AIMC) offers a promising solution to the von Neumann bottleneck. However, deploying transformer models on AIMC remains challenging due to their inherent need for flexibility and adaptability across diverse tasks. For the benefits of AIMC to be fully realized, weights of static vector-matrix multiplications must be mapped and programmed to analog devices in a weight-stationary manner. This poses two challenges for adapting a base network to hardware and downstream tasks: (i) conventional analog hardware-aware (AHWA) training requires retraining the entire model, and (ii) reprogramming analog devices is both time- and energy-intensive. To address these issues, we propose AHWA low-rank adaptation (AHWA-LoRA) training, a novel approach for efficiently adapting transformers to AIMC hardware. AHWA-LoRA training keeps the analog weights fixed as meta-weights and introduces lightweight external LoRA modules for both hardware and task adaptation. We validate AHWA-LoRA training on SQuAD v1.1 and the GLUE benchmark, demonstrate its scalability to larger models, and show its effectiveness in instruction tuning and reinforcement learning. We further evaluate a practical deployment scenario that balances AIMC tile latency with digital LoRA processing using optimized pipeline strategies, with RISC-V-based programmable multi-core accelerators. This hybrid architecture achieves efficient transformer inference with only a 4% per-layer overhead compared to a fully AIMC implementation.

Modern AI accelerators adopt dataflow architectures to achieve both high peak throughput (TOPS) and energy efficiency (TOPS/W). These designs feature wide datapaths and hierarchical scratchpad memories that supply dense compute arrays with high-bandwidth data access and extensive operand reuse. Complementing the compute–memory subsystem is a lightweight control path that orchestrates data movement, program loading, and register initialization. To reduce energy and area overheads, conventional processor features—such as instruction caches, execution stacks, and branch speculation—are deliberately omitted. While this streamlined design maximizes efficiency, it shifts a critical responsibility onto the compiler: transforming complex kernels into highly compact instruction streams that must fit entirely within the limited instruction buffers (IBUFFs) of the accelerator’s programmable units.In this paper, we introduce two novel compiler transformations—Loop Absorption (LA) and Loop Index Set Merging (LISM) for ultra compact code generation. Loop Absorption merges isomorphic sibling operations into a single loop body, while LISM unifies adjacent loops with similar bodies into a unified iteration space. Together, these complementary techniques eliminate redundant code patterns and produce compact hierarchical loop nests. We implement LA and LISM in the IBM Spyre compiler and evaluate them on diverse deep learning workloads including ResNet-50, Inception-v3, SSD, and BERT-Large. Across these models, our combined approach achieves a geometric mean compression of 1.48× over the baseline, enabling layers that previously exceeded IBUFF capacity to compile successfully.

AI Accelerators employ dataflow architectures to achieve impressive peak compute performance (TOPS) and processing efficiencies (TOPS/W). Typically, dataflow architectures use wide data-paths to connect off-chip memory to dense compute arrays (via hierarchy of on-chip memories/vector register files) for efficient data movement with reuse, as well as compute. Such architectures often possess an independent lightweight control-path for loading programs and initializing registers, and lack traditional architectural features like instruction cache and execution stacks. This poses a unique challenge to compiler requiring program generation of complex compute kernels to fit within an instruction buffer and allocating a limited set of scalar registers without support to spill to memory.This paper contributes a significant step towards spill-free compilation and proposes a Live range reduction optimization based on Affine expression propagation analysis. Our solution performs a global, compiler-directed analysis to model variable values as affine expressions of in-scope variables, enabling safe symbolic re-materialization of values at their use-sites leveraging near- by variables without introducing new operations. This shortens variable lifetimes, while significantly reducing register pressure without incurring program binary and execution overhead. The static nature and regular memory access patterns of AI applications make them well-suited for the proposed optimization. We demonstrate the effectiveness of the technique in the context of IBM Spyre accelerator and its compiler. Our results over a range of AI workloads spanning transformer and CNN models demonstrate spill-free code generation, with most of the workloads requiring less than 50% of the available registers.

A spiking neural network (SNN) in 5 nm FinFET CMOS operated at a spike rate of 2.6 GHz is presented. It uses a hybrid spiking scheme in which the data transmission between neurons is coded jointly in time and code domain, enabling a substantial latency reduction. Compared to a spiking scheme where a single spike is transmitted using a range of 64 time steps (6b resolution), the proposed hybrid scheme is designed to achieve a 16-fold latency reduction, resulting in only 4 time steps applied to a 4b binary-encoded spike bus <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\left(4 \times 2^{4}=64\right)$</tex>. This extension of the classical, purely time-coded spiking to the proposed hybrid spiking in the time and code domain is motivated by improving the energy efficiency (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{E}=\mathrm{P} \Delta \mathrm{t}$</tex>) where pure time coding is disadvantageous because the latency <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\Delta \mathrm{t}$</tex> is proportional to the code range of the input features. The proposed hybrid spiking scheme breaks this relationship and thus contributes to improving the energy efficiency. Key performance metrics of the experimental SNN chip include an energy efficiency of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{1.15} \text{~ p J} / \text{S O P}$</tex> and a minimum response time of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{1. 5 ~ n s}$</tex>.

Abstract Quantum computing is a new computational paradigm with the prospect of outperforming classical computing on many tasks. Optimization is frequently mentioned as a promising area for achieving a quantum advantage, i.e., solving a problem faster, more accurately, or more efficiently than classical computers. This chapter gives a brief summary of the potential benefits the state of the art and recommends possible directions to further progress.

When arranged in a crossbar configuration, resistive memory devices can be used to execute Matrix-Vector Multiplications (MVMs), the most dominant operation of many Machine Learning (ML) algorithms, in constant time complexity. Nonetheless, when performing computations in the analog domain, novel challenges are introduced in terms of arithmetic precision and stochasticity, due to non-ideal circuit and device behaviour. Moreover, these non-idealities have a temporal dimension, resulting in a degrading application accuracy over time. Facing these challenges, we propose a novel framework, named <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LionHeart</i>, to obtain hybrid analog-digital mappings to execute Deep Learning (DL) inference workloads using heterogeneous accelerators. The accuracy-constrained mappings derived by <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LionHeart</i> showcase, across different Convolutional Neural Networks (CNNs) and one transformer-based network, high accuracy and potential for speedup. The results of the full system simulations highlight run-time reductions and energy efficiency gains that exceed 6×, with a user-defined accuracy threshold for a fully digital floating point implementation.

Circuit knitting is the process of partitioning large quantum circuits into smaller subcircuits such that the result of the original circuits can be deduced by only running the subcircuits. Such techniques will be crucial for near-term and early fault-tolerant quantum computers, as the limited number of qubits is likely to be a major bottleneck for demonstrating quantum advantage. One typically distinguishes between gate cuts and wire cuts when partitioning a circuit. The cost for any circuit knitting approach scales exponentially in the number of cuts. One possibility to realize a cut is via the quasiprobability simulation technique. In fact, we argue that all existing rigorous circuit knitting techniques can be understood in this framework. Furthermore, we characterize the optimal overhead for wire cuts where the subcircuits can exchange classical information or not. We show that the optimal cost for cutting <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</i> wires without and with classical communication between the subcircuits scales as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">O</i>(16<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ⁿ</i>) and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">O</i>(4<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ⁿ</i>), respectively.

Superconducting quantum interference devices (SQUIDs) are exceptionally sensitive magnetometers capable of detecting weak magnetic fields. Miniaturizing these devices and integrating them onto scanning probes enables high-resolution imaging at low-temperature. Here, we fabricate nanometer-scale niobium SQUIDs with inner-loop sizes down to 10 nm at the apex of individual planar silicon cantilevers via a combination of wafer-scale optical lithography and focused-ion-beam (FIB) milling. These robust SQUID-on-lever probes overcome many of the limitations of existing devices, achieving spatial resolution better than 100 nm, magnetic flux sensitivity of $0.3~\mu\Phi_0/\sqrt{\rm{Hz}}$, and operation in magnetic fields up to about 0.5 T at 4.2 K. Nanopatterning via Ne- or He-FIB allows for the incorporation of a modulation line for coupling magnetic flux into the SQUID or a third Josephson junction for shifting its phase. Such advanced functionality, combined with high spatial resolution, large magnetic field range, and the ease of use of a cantilever-based scanning probe, extends the applicability of scanning SQUID microscopy to a wide range of magnetic, normal conducting, superconducting, and quantum Hall systems. We demonstrate magnetic imaging of skyrmions at the surface of bulk Cu$_2$OSeO$_3$. Analysis of the point spread function determined from imaging a single skyrmion yields a full-width-half-maximum of 87 nm. Moreover, we image modulated magnetization patterns with a period of 65 nm.

Planetary population synthesis is a tool that is used to better understand the key processes of planet formation at the statistical level. We seek to quantify the fidelity with which modern population syntheses reproduce observations in view of their use as predictive tools. We compared synthetic populations from the Generation 3 Bern Model of Planet Formation and Evolution (core accretion, solar-type host stars) and the HARPS/Coralie radial velocity sample. We biased the synthetic planet population according to the completeness of the observed data. We then performed quantitative statistical comparisons and systematically identified agreements and differences. Our nominal population reproduces many of the main features of the HARPS planets, such as two main groups of planets in the mass-distance diagram (close-in sub-Neptunes and distant giants), a bimodal mass function with a less populated `desert', an observed mean multiplicity of about 1.6, and several key correlations regarding the stellar metallicity dependency, the period ratio distribution, and the eccentricity distribution. Considering that the model was not optimised beforehand to reproduce any particular survey, this indicates that some of the important physical processes governing planetary formation could be captured. The remaining discrepancies that can be quantified thanks to the population synthesis approach point to areas that are not fully captured in the model. For instance, we find that the synthetic population has 1) in absolute terms too many planets by ≈70 2) a `desert' that is too deep by ≈60 3) a relative excess of giant planets by ≈40 4) planet eccentricities that are on average too low by a factor of about two (median of versus ), and 5) a metallicity effect that is too weak. Finally, the synthetic planets are overall too close to the star compared to the HARPS sample. The differences allowed us to find model parameters that better reproduce the observed planet masses, for which we computed additional synthetic populations. We find that decreasing the planet formation efficiency by increasing the planetesimal size re-balances the number of sub-Neptunes versus giant planets. Changing the efficiency of gas-driven migration also affects the sub-Neptune to giant planet ratio, with lower migration rates resulting in more giant planets and fewer sub-Neptunes. However, only modifying the model parameters seems to be insufficient for the model to fully reproduce both the observed mass and distance distributions at the same time. Instead, physical processes appear to be missing. Planets may originate on wider orbits than our model predicts. Mechanisms leading to higher eccentricities and slower disc-limited gas accretion also seem necessary. We also advocate that theoretical models should make a quantitative, rather than merely a qualitative, comparison between the many current and future large surveys and theoretical results to better understand the origins of planetary systems.