arXiv subjects

Shubham Negi

Publications and source records attributed to Shubham Negi.

11 recordsLinked to original sources

arXiv2026

TraceNAS: Zero-shot LLM Pruning via Gradient Trace Correlation

Structured pruning is essential for efficient deployment of Large Language Models (LLMs). The varying sensitivity of LLM sub-blocks to pruning necessitates the identification of optimal non-uniformly pruned models. Existing methods evaluate the importance of layers, attention heads, or weight channels in isolation. Such localized focus ignores the complex global structural dependencies that exist across the model. Training-aware structured pruning addresses global dependencies, but its computational cost can be just as expensive as post-pruning training. To alleviate the computational burden of training-aware pruning and capture global structural dependencies, we propose TraceNAS, a training-free Neural Architecture Search (NAS) framework that jointly explores structured pruning of LLM depth and width. TraceNAS identifies pruned models that maintain a high degree of loss landscape alignment with the pretrained model using a scale-invariant zero-shot proxy, effectively selecting models that exhibit maximal performance potential during post-pruning training. TraceNAS is highly efficient, enabling high-fidelity discovery of pruned models on a single GPU in 8.5 hours, yielding a 10$\times$ reduction in GPU-hours compared to training-aware methods. Evaluations on the Llama and Qwen families demonstrate that TraceNAS is competitive with training-aware baselines across commonsense and reasoning benchmarks.

cs.LG↗

arXiv2025

HALO: Memory-Centric Heterogeneous Accelerator with 2.5D Integration for Low-Batch LLM Inference

The rapid adoption of Large Language Models (LLMs) has driven a growing demand for efficient inference, particularly in latency-sensitive applications such as chatbots and personalized assistants. Unlike traditional deep neural networks, LLM inference proceeds in two distinct phases: the prefill phase, which processes the full input sequence in parallel, and the decode phase, which generates tokens sequentially. These phases exhibit highly diverse compute and memory requirements, which makes accelerator design particularly challenging. Prior works have primarily been optimized for high-batch inference or evaluated only short input context lengths, leaving the low-batch and long context regime, which is critical for interactive applications, largely underexplored. We propose HALO, a heterogeneous memory centric accelerator designed for these unique challenges of prefill and decode phases in low-batch LLM inference. HALO integrates HBM based Compute-in-DRAM (CiD) with an on-chip analog Compute-in-Memory (CiM), co-packaged using 2.5D integration. To further improve the hardware utilization, we introduce a phase-aware mapping strategy that adapts to the distinct demands of the prefill and decode phases. Compute bound operations in the prefill phase are mapped to CiM to exploit its high throughput matrix multiplication capability, while memory-bound operations in the decode phase are executed on CiD to benefit from reduced data movement within DRAM. Additionally, we present an analysis of the performance tradeoffs of LLMs under two architectural extremes: a fully CiD and a fully on-chip analog CiM design to highlight the need for a heterogeneous design. We evaluate HALO on LLaMA-2 7B and Qwen3 8B models. Our experimental results show that LLMs mapped to HALO achieve up to 18x geometric mean speedup over AttAcc, an attention-optimized mapping and 2.5x over CENT, a fully CiD based mapping.

cs.AR↗

arXiv2025

COMET: A Framework for Modeling Compound Operation Dataflows with Explicit Collectives

Modern machine learning accelerators are designed to efficiently execute deep neural networks (DNNs) by optimizing data movement, memory hierarchy, and compute throughput. However, emerging DNN models such as large language models, state space models increasingly rely on compound operations-structured compositions of multiple basic operations-which introduce new challenges for dataflow optimization and minimizing off-chip memory traffic. Moreover, as model size continues to grow, deployment across spatially distributed compute clusters becomes essential, requiring frequent and complex collective communication. Existing dataflow optimization frameworks and performance models either focus on single operations or lack explicit modeling of collective communication cost, limiting their applicability to modern workloads. To address these limitations, we propose, a framework for modeling and optimizing dataflow for compound operations on machine learning accelerators. COMET introduces a novel representation that explicitly models collective communication across spatial clusters, along with latency and energy cost models that account for both GEMM and non-GEMM operation level dependencies within compound operations. We demonstrate COMET's capabilities to analyze and optimize dataflows for compound operations such as GEMM--Softmax, GEMM--LayerNorm, and self-attention, across both edge and cloud accelerator configurations. Our collective-aware modeling enables exploration of a broader mapping space, leading to improved performance and energy efficiency. Specifically, our optimized dataflows achieve up to 1.42$\times$ speedup for GEMM-Softmax, 3.46$\times$ for GEMM-LayerNorm and 1.82$\times$ for self-attention compared to unfused baselines.

cs.AR↗

arXiv2024

TSkips: Efficiency Through Explicit Temporal Delay Connections in Spiking Neural Networks

Spiking Neural Networks (SNNs) with their bio-inspired Leaky Integrate-and-Fire (LIF) neurons inherently capture temporal information. This makes them well-suited for sequential tasks like processing event-based data from Dynamic Vision Sensors (DVS) and event-based speech tasks. Harnessing the temporal capabilities of SNNs requires mitigating vanishing spikes during training, capturing spatio-temporal patterns and enhancing precise spike timing. To address these challenges, we propose TSkips, augmenting SNN architectures with forward and backward skip connections that incorporate explicit temporal delays. These connections capture long-term spatio-temporal dependencies and facilitate better spike flow over long sequences. The introduction of TSkips creates a vast search space of possible configurations, encompassing skip positions and time delay values. To efficiently navigate this search space, this work leverages training-free Neural Architecture Search (NAS) to identify optimal network structures and corresponding delays. We demonstrate the effectiveness of our approach on four event-based datasets: DSEC-flow for optical flow estimation, DVS128 Gesture for hand gesture recognition and Spiking Heidelberg Digits (SHD) and Spiking Speech Commands (SSC) for speech recognition. Our method achieves significant improvements across these datasets: up to 18% reduction in Average Endpoint Error (AEE) on DSEC-flow, 8% increase in classification accuracy on DVS128 Gesture, and up to 8% and 16% higher classification accuracy on SHD and SSC, respectively.

cs.NE↗

arXiv2024

SpiDR: A Reconfigurable Digital Compute-in-Memory Spiking Neural Network Accelerator for Event-based Perception

Spiking Neural Networks (SNNs), with their inherent recurrence, offer an efficient method for processing the asynchronous temporal data generated by Dynamic Vision Sensors (DVS), making them well-suited for event-based vision applications. However, existing SNN accelerators suffer from limitations in adaptability to diverse neuron models, bit precisions and network sizes, inefficient membrane potential (Vmem) handling, and limited sparse optimizations. In response to these challenges, we propose a scalable and reconfigurable digital compute-in-memory (CIM) SNN accelerator \chipname with a set of key features: 1) It uses in-memory computations and reconfigurable operating modes to minimize data movement associated with weight and Vmem data structures while efficiently adapting to different workloads. 2) It supports multiple weight/Vmem bit precision values, enabling a trade-off between accuracy and energy efficiency and enhancing adaptability to diverse application demands. 3) A zero-skipping mechanism for sparse inputs significantly reduces energy usage by leveraging the inherent sparsity of spikes without introducing high overheads for low sparsity. 4) Finally, the asynchronous handshaking mechanism maintains the computational efficiency of the pipeline for variable execution times of different computation units. We fabricated \chipname in 65 nm Taiwan Semiconductor Manufacturing Company (TSMC) low-power (LP) technology. It demonstrates competitive performance (scaled to the same technology node) to other digital SNN accelerators proposed in the recent literature and supports advanced reconfigurability. It achieves up to 5 TOPS/W energy efficiency at 95% input sparsity with 4-bit weights and 7-bit Vmem precision.

cs.AR↗

arXiv2024

Approximate ADCs for In-Memory Computing

In memory computing (IMC) architectures for deep learning (DL) accelerators leverage energy-efficient and highly parallel matrix vector multiplication (MVM) operations, implemented directly in memory arrays. Such IMC designs have been explored based on CMOS as well as emerging non-volatile memory (NVM) technologies like RRAM. IMC architectures generally involve a large number of cores consisting of memory arrays, storing the trained weights of the DL model. Peripheral units like DACs and ADCs are also used for applying inputs and reading out the output values. Recently reported designs reveal that the ADCs required for reading out the MVM results, consume more than 85% of the total compute power and also dominate the area, thereby eschewing the benefits of the IMC scheme. Mitigation of imperfections in the ADCs, namely, non-linearity and variations, incur significant design overheads, due to dedicated calibration units. In this work we present peripheral aware design of IMC cores, to mitigate such overheads. It involves incorporating the non-idealities of ADCs in the training of the DL models, along with that of the memory units. The proposed approach applies equally well to both current mode as well as charge mode MVM operations demonstrated in recent years., and can significantly simplify the design of mixed-signal IMC units.

cs.ET↗

arXiv2024

HCiM: ADC-Less Hybrid Analog-Digital Compute in Memory Accelerator for Deep Learning Workloads

Analog Compute-in-Memory (CiM) accelerators are increasingly recognized for their efficiency in accelerating Deep Neural Networks (DNN). However, their dependence on Analog-to-Digital Converters (ADCs) for accumulating partial sums from crossbars leads to substantial power and area overhead. Moreover, the high area overhead of ADCs constrains the throughput due to the limited number of ADCs that can be integrated per crossbar. An approach to mitigate this issue involves the adoption of extreme low-precision quantization (binary or ternary) for partial sums. Training based on such an approach eliminates the need for ADCs. While this strategy effectively reduces ADC costs, it introduces the challenge of managing numerous floating-point scale factors, which are trainable parameters like DNN weights. These scale factors must be multiplied with the binary or ternary outputs at the columns of the crossbar to ensure system accuracy. To that effect, we propose an algorithm-hardware co-design approach, where DNNs are first trained with quantization-aware training. Subsequently, we introduce HCiM, an ADC-Less Hybrid Analog-Digital CiM accelerator. HCiM uses analog CiM crossbars for performing Matrix-Vector Multiplication operations coupled with a digital CiM array dedicated to processing scale factors. This digital CiM array can execute both addition and subtraction operations within the memory array, thus enhancing processing speed. Additionally, it exploits the inherent sparsity in ternary quantization to achieve further energy savings. Compared to an analog CiM baseline architecture using 7 and 4-bit ADC, HCiM achieves energy reductions up to 28% and 12%, respectively

cs.AR↗

arXiv2023

Best of Both Worlds: Hybrid SNN-ANN Architecture for Event-based Optical Flow Estimation

In the field of robotics, event-based cameras are emerging as a promising low-power alternative to traditional frame-based cameras for capturing high-speed motion and high dynamic range scenes. This is due to their sparse and asynchronous event outputs. Spiking Neural Networks (SNNs) with their asynchronous event-driven compute, show great potential for extracting the spatio-temporal features from these event streams. In contrast, the standard Analog Neural Networks (ANNs) fail to process event data effectively. However, training SNNs is difficult due to additional trainable parameters (thresholds and leaks), vanishing spikes at deeper layers, and a non-differentiable binary activation function. Furthermore, an additional data structure, membrane potential, responsible for keeping track of temporal information, must be fetched and updated at every timestep in SNNs. To overcome these challenges, we propose a novel SNN-ANN hybrid architecture that combines the strengths of both. Specifically, we leverage the asynchronous compute capabilities of SNN layers to effectively extract the input temporal information. Concurrently, the ANN layers facilitate training and efficient hardware deployment on traditional machine learning hardware such as GPUs. We provide extensive experimental analysis for assigning each layer to be spiking or analog, leading to a network configuration optimized for performance and ease of training. We evaluate our hybrid architecture for optical flow estimation on DSEC-flow and Multi-Vehicle Stereo Event-Camera (MVSEC) datasets. On the DSEC-flow dataset, the hybrid SNN-ANN architecture achieves a 40% reduction in average endpoint error (AEE) with 22% lower energy consumption compared to Full-SNN, and 48% lower AEE compared to Full-ANN, while maintaining comparable energy usage.

cs.CV↗

arXiv2021

NAX: Co-Designing Neural Network and Hardware Architecture for Memristive Xbar based Computing Systems

In-Memory Computing (IMC) hardware using Memristive Crossbar Arrays (MCAs) are gaining popularity to accelerate Deep Neural Networks (DNNs) since it alleviates the "memory wall" problem associated with von-Neumann architecture. The hardware efficiency (energy, latency and area) as well as application accuracy (considering device and circuit non-idealities) of DNNs mapped to such hardware are co-dependent on network parameters, such as kernel size, depth etc. and hardware architecture parameters such as crossbar size. However, co-optimization of both network and hardware parameters presents a challenging search space comprising of different kernel sizes mapped to varying crossbar sizes. To that effect, we propose NAX -- an efficient neural architecture search engine that co-designs neural network and IMC based hardware architecture. NAX explores the aforementioned search space to determine kernel and corresponding crossbar sizes for each DNN layer to achieve optimal tradeoffs between hardware efficiency and application accuracy. Our results from NAX show that the networks have heterogeneous crossbar sizes across different network layers, and achieves optimal hardware efficiency and accuracy considering the non-idealities in crossbars. On CIFAR-10 and Tiny ImageNet, our models achieve 0.8%, 0.2% higher accuracy, and 17%, 4% lower EDAP (energy-delay-area product) compared to a baseline ResNet-20 and ResNet-18 models, respectively.

cs.ET↗

arXiv2020

Exploration of Optimized Semantic Segmentation Architectures for edge-Deployment on Drones

In this paper, we present an analysis on the impact of network parameters for semantic segmentation architectures in context of UAV data processing. We present the analysis on the DroneDeploy Segmentation benchmark. Based on the comparative analysis we identify the optimal network architecture to be FPN-EfficientNetB3 with pretrained encoder backbones based on Imagenet Dataset. The network achieves IoU score of 0.65 and F1-score of 0.71 over the validation dataset. We also compare the various architectures in terms of their memory footprint and inference latency with further exploration of the impact of TensorRT based optimizations. We achieve memory savings of ~4.1x and latency improvement of 10% compared to Model: FPN and Backbone: InceptionResnetV2.

cs.CV↗

arXiv2020

Methodology for Realizing VMM with Binary RRAM Arrays: Experimental Demonstration of Binarized-ADALINE Using OxRAM Crossbar

In this paper, we present an efficient hardware mapping methodology for realizing vector matrix multiplication (VMM) on resistive memory (RRAM) arrays. Using the proposed VMM computation technique, we experimentally demonstrate a binarized-ADALINE (Adaptive Linear) classifier on an OxRAM crossbar. An 8x8 OxRAM crossbar with Ni/3-nm HfO2/7 nm Al-doped-TiO2/TiN device stack is used. Weight training for the binarized-ADALINE classifier is performed ex-situ on UCI cancer dataset. Post weight generation the OxRAM array is carefully programmed to binary weight-states using the proposed weight mapping technique on a custom-built testbench. Our VMM powered binarized-ADALINE network achieves a classification accuracy of 78% in simulation and 67% in experiments. Experimental accuracy was found to drop mainly due to crossbar inherent sneak-path issues and RRAM device programming variability.

cs.ET↗