Sachin S. Sapatnekar’s research while affiliated with University of Minnesota and other places

Many combinatorial problems can be mapped to Ising machines, i.e., networks of coupled oscillators that settle to a minimum-energy ground state, from which the problem solution is inferred. This work proposes DROID, a novel event-driven method for simulating the evolution of a CMOS Ising machine to its ground state. The approach is accurate under general delay-phase relations that include the effects of the transistor nonlinearities and is computationally efficient. On a realistic-size all-to-all coupled ring oscillator array, DROID is nearly four orders of magnitude faster than a traditional HSPICE simulation in predicting the evolution of a coupled oscillator system and is demonstrated to attain a similar distribution of solutions as the hardware.

Device sizing is crucial for meeting performance specifications in operational transconductance amplifiers (OTAs), and this work proposes an automated sizing framework based on a transformer model. The approach first leverages the driving-point signal flow graph (DP-SFG) to map an OTA circuit and its specifications into transformer-friendly sequential data. A specialized tokenization approach is applied to the sequential data to expedite the training of the transformer on a diverse range of OTA topologies, under multiple specifications. Under specific performance constraints, the trained transformer model is used to accurately predict DP-SFG parameters in the inference phase. The predicted DP-SFG parameters are then translated to transistor sizes using a precomputed look-up table-based approach inspired by the gm/Id methodology. In contrast to previous conventional or machine-learning-based methods, the proposed framework achieves significant improvements in both speed and computational efficiency by reducing the need for expensive SPICE simulations within the optimization loop; instead, almost all SPICE simulations are confined to the one-time training phase. The method is validated on a variety of unseen specifications, and the sizing solution demonstrates over 90% success in meeting specifications with just one SPICE simulation for validation, and 100% success with 3-5 additional SPICE simulations.

As circuit designs become more intricate, obtaining accurate performance estimation in early stages, for effective design space exploration, becomes more time-consuming. Traditional logic optimization approaches often rely on proxy metrics to approximate post-mapping performance and area. However, these proxies do not always correlate well with actual post-mapping delay and area, resulting in suboptimal designs. To address this issue, we explore a ground-truth-based optimization flow that directly incorporates the exact post-mapping delay and area during optimization. While this approach improves design quality, it also significantly increases computational costs, particularly for large-scale designs. To overcome the runtime challenge, we apply machine learning models to predict post-mapping delay and area using the features extracted from AIGs. Our experimental results show that the model has high prediction accuracy with good generalization to unseen designs. Furthermore, the ML-enhanced logic optimization flow significantly reduces runtime while maintaining comparable performance and area outcomes.

Performance modeling is a key bottleneck for analog design automation. Although machine learning-based models have advanced the state-of-the-art, they have so far suffered from huge data preparation cost, very limited reusability, and inadequate accuracy for large circuits. We introduce ML-based macro-modeling techniques to mitigate these problems for linear analog ICs and ADC/DACs. The modeling techniques are based on macro-models, which can be assembled to evaluate circuit system performance, and more appealingly can be reused across different circuit topologies. On representative testcases, our method achieves more than $1700\times$ speedup for data preparation and remarkably smaller model errors compared to recent ML approaches. It also attains $3600\times$ acceleration over SPICE simulation with very small errors and reduces data preparation time for an ADC design from 40 days to 9.6 h.

The design of active array structures in analog circuits requires careful matching to minimize the impact of variations. This work presents a constructive approach for building these arrays to directly incorporate shifts due to process variations, considering systematic first-order and second order gradients; to account for systematic layout effects, including parasitic mismatch and layout-dependent effects due to stress; and to ensure that the resulting layout delivers high performance. The proposed algorithms are targeted to FinFET technologies and are validated for multiple analog blocks in a commercial 12nm FinFET process. The layouts generated by the proposed method are demonstrated to provide better matching and performance than prior methods.

Ising machines are effective solvers for complex combinatorial optimization problems. The idea is mapping the optimal solution(s) to a combinatorial optimization problem to the minimum energy state(s) of a physical system, which naturally converges to and stabilizes at a minimum energy state upon perturbance. The underlying mathematical abstraction, the Ising model, was originally developed to explain dynamic behavior of ferromagnetic materials and was shown to generalize to numerous other physical systems. In a generic optimization problem, each variable can interact with another in different ways. At the same time, problem sizes of practical importance are growing very fast. Unfortunately, both the number and connectivity of spins in hardware are subject to fundamental physical limits. Different problems feature different interaction patterns between variables which may not always directly match the network topology supported by a specific Ising machine. In the presence of a mismatch, emulating generic interactions using the machine topology is usually possible, however, comes at the cost of additional physical spins to facilitate the mapping. Furthermore, mismatches in the problem vs. hardware connectivity render even more physical spins necessary. Combinatorial optimization problems of practical importance come with diverse connectivity patterns, which a rigid network topology in hardware cannot efficiently cover. To bridge the gap between application demand and hardware resources, in analogy to classical heterogeneous chip multiprocessors, in this paper we make the case for heterogeneous Ising multiprocessors, where each Ising core features a different connectivity. We provide a detailed design space exploration and quantify the efficiency of different design options in terms of time or energy to solution along with solution accuracy compared to homogeneous alternatives.

Today’s performance analysis frameworks for deep learning accelerators suffer from two significant limitations. First, although modern convolutional neural networks (CNNs) consist of many types of layers other than convolution, especially during training, these frameworks largely focus on convolution layers only. Second, these frameworks are generally targeted towards inference, and lack support for training operations. This work proposes a novel open-source performance analysis framework, SimDIT, for general ASIC-based systolic hardware accelerator platforms. The modeling effort of SimDIT comprehensively covers convolution and non-convolution operations of both CNN inference and training on a highly parameterizable hardware substrate. SimDIT is integrated with a backend silicon implementation flow and provides detailed end-to-end performance statistics (i.e., data access cost, cycle counts, energy, and power) for executing CNN inference and training workloads. SimDIT-enabled performance analysis reveals that on a 64 × 64 processing array, non-convolution operations constitute 59.5% of total runtime for ResNet-50 training workload. In addition, by optimally distributing available off-chip DRAM bandwidth and on-chip SRAM resources, SimDIT achieves 18 × performance improvement over a generic static resource allocation for ResNet-50 inference.