arXiv subjects

Umit Y. Ogras

Publications and source records attributed to Umit Y. Ogras.

At least 19 recordsLinked to original sources

arXiv2026

Uni-SFU: Algorithm-HW Co-Design for Universal SFUs via Mixed-Degree Piecewise Approximation

Nonlinear activation functions are essential to modern deep neural networks (DNNs), but their hardware evaluation places significant pressure on the special-function units (SFUs) of GPUs and custom accelerators. Therefore, piecewise polynomial approximations are commonly used within allowed error bounds to improve computational efficiency. However, existing techniques often approximate each activation function in isolation using fixed-degree polynomials and uniform segments, leading to hardware redundancy and sub-optimal precision. To address these limitations, we present Uni-SFU, an algorithm-hardware co-design framework that jointly optimizes approximation accuracy and silicon area for a diverse set of activation functions. Uni-SFU leverages a joint search across all target functions to assign mixed-degree polynomials to nonuniform segments, guided by an RTL-derived area cost model. This approach identifies a unified hardware configuration to implement the target activation functions under given accuracy constraints. Validated across over 700 neural network variants and three Natural Language Processing (NLP) models, Uni-SFU achieves a superior Mean Squared Error (MSE) below 8.22x10^-8, limiting top-1 accuracy degradation to within 1.02% compared to floating-point baselines. The proposed design occupies only 6,800 um2 in GF 22nm CMOS technology, achieving a superior trade-off between silicon area and system-level accuracy compared to SOTA counterparts.

cs.AR↗

arXiv2026

APEX: Adaptive Expert Prefetching for Memory-Efficient Edge MoE Inference

Mixture-of-Experts (MoE) models are attractive for edge deployment because they provide high model capacity while activating only a small subset of parameters per token, improving compute efficiency. However, MoE inference at the edge is fundamentally limited by memory. Expert parameters are large and often reside in off-chip memory due to capacity, cost, and power constraints, putting expert loading to the critical path. We present APEX: Adaptive Expert Prefetching, a predictive resource management framework that overlaps expert loading with useful computation. APEX introduces a lightweight prefetch router that predicts candidate experts before the attention block to dynamically fetch additional experts using a learned confidence model. This adaptive strategy achieves over 99% overlap accuracy, significantly outperforming fixed top-k prefetching techniques. APEX supports two execution modes: a correctness-preserving mode that guarantees exact routing semantics, and a stall-free mode that eliminates residual stalls by operating on available experts with negligible impact on application accuracy. Across multiple MoE models, the correctness-preserving mode reduces per-token latency by up to 26% and improves energy-delay product (EDP) by up to 41% over state-of-the-art baselines, while the stall-free mode provides additional efficiency gains with negligible impact on application accuracy. These results establish adaptive, confidence-driven expert prefetching as an effective approach for efficient MoE inference on edge systems.

cs.AR↗

arXiv2026

ReVolt: Power Delivery Network-Aware Voltage Droop Control for 2.5D PIM Chiplet Architectures

Processing-in-memory (PIM)-based 2.5D multi-chiplet platforms are enablers for machine learning (ML) workloads. However, their performance is affected by the power delivery network (PDN), where varying chiplet-level current demand induces spatially and temporally varying voltage droop. These droop events lead to voltage violations, degrades system performance, and impact inference accuracy for ML workloads. In this work, we propose ReVolt, a dynamic operation unit (OU)-based framework for mitigating voltage droop in PIM-based multi-chiplet systems. ReVolt leverages an LSTM-based PDN surrogate to predict per-chiplet supply voltage trajectories at runtime, enabling proactive adjustment of OU size to mitigate droop events. By treating OU size as a control knob, ReVolt regulates chiplet-level current demand while maintaining computational accuracy. This approach prevents voltage droop violations and improves energy-delay product (EDP) while preserving ML model inference accuracy. Experimental results demonstrate that ReVolt prevents voltage droop violations while achieving an average 76x reduction in EDP compared to existing fixed and dynamic OU-based baselines, without compromising inference accuracy of ML models.

cs.AR↗

arXiv2026

DUET: Disaggregated Hybrid Mamba-Transformer LLMs with Prefill and Decode-Specific Packages

Large language models operate in distinct compute-bound prefill followed by memory bandwidth-bound decode phases. Hybrid Mamba-Transformer models inherit this asymmetry while adding state space model (SSM) recurrences and element-wise operations that map poorly to matmul-centric accelerators. This mismatch causes performance bottlenecks, showing that a homogeneous architecture cannot satisfy all requirements. We introduce DUET, a disaggregated accelerator that assigns prefill and decode phases to specialized packages. The Prefill package utilizes systolic array chiplets with off-package memory for efficient large matrix multiplications and long-sequence SSMs. The Decode package utilizes vector-unit arrays with high-bandwidth in-package memory to accelerate token-by-token SSM and vector-matrix multiplications. Both architectures are runtime-configurable to support hybrid models with mixed Mamba and attention layers. Evaluations on Nemotron-H-56B, Zamba2-7B, and Llama3-8B across four workloads show that DUET achieves 4x faster time to first token, 1.4x higher throughput, and 1.5x lower time between tokens over the B200 GPU.

cs.AR↗

arXiv2025

CHIPSIM: A Co-Simulation Framework for Deep Learning on Chiplet-Based Systems

Due to reduced manufacturing yields, traditional monolithic chips cannot keep up with the compute, memory, and communication demands of data-intensive applications, such as rapidly growing deep neural network (DNN) models. Chiplet-based architectures offer a cost-effective and scalable solution by integrating smaller chiplets via a network-on-interposer (NoI). Fast and accurate simulation approaches are critical to unlocking this potential, but existing methods lack the required accuracy, speed, and flexibility. To address this need, this work presents CHIPSIM, a comprehensive co-simulation framework designed for parallel DNN execution on chiplet-based systems. CHIPSIM concurrently models computation and communication, accurately capturing network contention and pipelining effects that conventional simulators overlook. Furthermore, it profiles the chiplet and NoI power consumptions at microsecond granularity for precise transient thermal analysis. Extensive evaluations with homogeneous/heterogeneous chiplets and different NoI architectures demonstrate the framework's versatility, up to 340% accuracy improvement, and power/thermal analysis capability.

cs.AR↗

arXiv2025

eMamba: Efficient Acceleration Framework for Mamba Models in Edge Computing

State Space Model (SSM)-based machine learning architectures have recently gained significant attention for processing sequential data. Mamba, a recent sequence-to-sequence SSM, offers competitive accuracy with superior computational efficiency compared to state-of-the-art transformer models. While this advantage makes Mamba particularly promising for resource-constrained edge devices, no hardware acceleration frameworks are currently optimized for deploying it in such environments. This paper presents eMamba, a comprehensive end-to-end hardware acceleration framework explicitly designed for deploying Mamba models on edge platforms. eMamba maximizes computational efficiency by replacing complex normalization layers with lightweight hardware-aware alternatives and approximating expensive operations, such as SiLU activation and exponentiation, considering the target applications. Then, it performs an approximation-aware neural architecture search (NAS) to tune the learnable parameters used during approximation. Evaluations with Fashion-MNIST, CIFAR-10, and MARS, an open-source human pose estimation dataset, show eMamba achieves comparable accuracy to state-of-the-art techniques using 1.63-19.9$\times$ fewer parameters. In addition, it generalizes well to large-scale natural language tasks, demonstrating stable perplexity across varying sequence lengths on the WikiText2 dataset. We also quantize and implement the entire eMamba pipeline on an AMD ZCU102 FPGA and ASIC using GlobalFoundries (GF) 22 nm technology. Experimental results show 4.95-5.62$\times$ lower latency and 2.22-9.95$\times$ higher throughput, with 4.77$\times$ smaller area, 9.84$\times$ lower power, and 48.6$\times$ lower energy consumption than baseline solutions while maintaining competitive accuracy.

cs.LG↗

arXiv2025

THERMOS: Thermally-Aware Multi-Objective Scheduling of AI Workloads on Heterogeneous Multi-Chiplet PIM Architectures

Chiplet-based integration enables large-scale systems that combine diverse technologies, enabling higher yield, lower costs, and scalability, making them well-suited to AI workloads. Processing-in-Memory (PIM) has emerged as a promising solution for AI inference, leveraging technologies such as ReRAM, SRAM, and FeFET, each offering unique advantages and trade-offs. A heterogeneous chiplet-based PIM architecture can harness the complementary strengths of these technologies to enable higher performance and energy efficiency. However, scheduling AI workloads across such a heterogeneous system is challenging due to competing performance objectives, dynamic workload characteristics, and power and thermal constraints. To address this need, we propose THERMOS, a thermally-aware, multi-objective scheduling framework for AI workloads on heterogeneous multi-chiplet PIM architectures. THERMOS trains a single multi-objective reinforcement learning (MORL) policy that is capable of achieving Pareto-optimal execution time, energy, or a balanced objective at runtime, depending on the target preferences. Comprehensive evaluations show that THERMOS achieves up to 89% faster average execution time and 57% lower average energy consumption than baseline AI workload scheduling algorithms with only 0.14% runtime and 0.022% energy overhead.

cs.AR↗

arXiv2025

Designing High-Performance and Thermally Feasible Multi-Chiplet Architectures enabled by Non-bendable Glass Interposer

Multi-chiplet architectures enabled by glass interposer offer superior electrical performance, enable higher bus widths due to reduced crosstalk, and have lower capacitance in the redistribution layer than current silicon interposer-based systems. These advantages result in lower energy per bit, higher communication frequencies, and extended interconnect range. However, deformation of the package (warpage) in glass interposer-based systems becomes a critical challenge as system size increases, leading to severe mechanical stress and reliability concerns. Beyond a certain size, conventional packaging techniques fail to manage warpage effectively, necessitating new approaches to mitigate warpage induced bending with scalable performance for glass interposer based multi-chiplet systems. To address these inter-twined challenges, we propose a thermal-, warpage-, and performance-aware design framework that employs architecture and packaging co-optimization. The proposed framework disintegrates the surface and embedded chiplets to balance conflicting design objectives, ensuring optimal trade-offs between performance, power, and structural reliability. Our experiments demonstrate that optimized multi-chiplet architectures from our design framework achieve up to 64.7% performance improvement and 40% power reduction compared to traditional 2.5D systems to execute deep neural network workloads with lower fabrication costs.

cs.AR↗

arXiv2025

MFIT: Multi-Fidelity Thermal Modeling for 2.5D and 3D Multi-Chiplet Architectures

Rapidly evolving artificial intelligence and machine learning applications require ever-increasing computational capabilities, while monolithic 2D design technologies approach their limits. Heterogeneous integration of smaller chiplets using a 2.5D silicon interposer and 3D packaging has emerged as a promising paradigm to address this limit and meet performance demands. These approaches offer a significant cost reduction and higher manufacturing yield than monolithic 2D integrated circuits. However, the compact arrangement and high compute density exacerbate the thermal management challenges, potentially compromising performance. Addressing these thermal modeling challenges is critical, especially as system sizes grow and different design stages require varying levels of accuracy and speed. Since no single thermal modeling technique meets all these needs, this paper introduces MFIT, a range of multi-fidelity thermal models that effectively balance accuracy and speed. These multi-fidelity models can enable efficient design space exploration and runtime thermal management. Our extensive testing on systems with 16, 36, and 64 2.5D integrated chiplets and 16x3 3D integrated chiplets demonstrates that these models can reduce execution times from days to mere seconds and milliseconds with negligible loss in accuracy.

cs.AR↗

arXiv2023

Theoretical Analysis and Evaluation of NoCs with Weighted Round-Robin Arbitration

Fast and accurate performance analysis techniques are essential in early design space exploration and pre-silicon evaluations, including software eco-system development. In particular, on-chip communication continues to play an increasingly important role as the many-core processors scale up. This paper presents the first performance analysis technique that targets networks-on-chip (NoCs) that employ weighted round-robin (WRR) arbitration. Besides fairness, WRR arbitration provides flexibility in allocating bandwidth proportionally to the importance of the traffic classes, unlike basic round-robin and priority-based arbitration. The proposed approach first estimates the effective service time of the packets in the queue due to WRR arbitration. Then, it uses the effective service time to compute the average waiting time of the packets. Next, we incorporate a decomposition technique to extend the analytical model to handle NoC of any size. The proposed approach achieves less than 5% error while executing real applications and 10% error under challenging synthetic traffic with different burstiness levels.

cs.PF↗

arXiv2023

PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm

Multi-objective reinforcement learning (MORL) approaches have emerged to tackle many real-world problems with multiple conflicting objectives by maximizing a joint objective function weighted by a preference vector. These approaches find fixed customized policies corresponding to preference vectors specified during training. However, the design constraints and objectives typically change dynamically in real-life scenarios. Furthermore, storing a policy for each potential preference is not scalable. Hence, obtaining a set of Pareto front solutions for the entire preference space in a given domain with a single training is critical. To this end, we propose a novel MORL algorithm that trains a single universal network to cover the entire preference space scalable to continuous robotic tasks. The proposed approach, Preference-Driven MORL (PD-MORL), utilizes the preferences as guidance to update the network parameters. It also employs a novel parallelization approach to increase sample efficiency. We show that PD-MORL achieves up to 25% larger hypervolume for challenging continuous control tasks and uses an order of magnitude fewer trainable parameters compared to prior approaches.

cs.LG↗

arXiv2023

Machine Learning-based Low Overhead Congestion Control Algorithm for Industrial NoCs

Network-on-Chip (NoC) congestion builds up during heavy traffic load and cripples the system performance by stalling the cores. Moreover, congestion leads to wasted link bandwidth due to blocked buffers and bouncing packets. Existing approaches throttle the cores after congestion is detected, reducing efficiency and wasting line bandwidth unnecessarily. In contrast, we propose a lightweight machine learning-based technique that helps predict congestion in the network. Specifically, our proposed technique collects the features related to traffic at each destination. Then, it labels the features using a novel time reversal approach. The labeled data is used to design a low overhead and an explainable decision tree model used at runtime congestion control. Experimental evaluations with synthetic and real traffic on industrial 6$\times$6 NoC show that the proposed approach increases fairness and memory read bandwidth by up to 114\% with respect to existing congestion control technique while incurring less than 0.01\% of overhead.

cs.AR↗

arXiv2022

COIN: Communication-Aware In-Memory Acceleration for Graph Convolutional Networks

Graph convolutional networks (GCNs) have shown remarkable learning capabilities when processing graph-structured data found inherently in many application areas. GCNs distribute the outputs of neural networks embedded in each vertex over multiple iterations to take advantage of the relations captured by the underlying graphs. Consequently, they incur a significant amount of computation and irregular communication overheads, which call for GCN-specific hardware accelerators. To this end, this paper presents a communication-aware in-memory computing architecture (COIN) for GCN hardware acceleration. Besides accelerating the computation using custom compute elements (CE) and in-memory computing, COIN aims at minimizing the intra- and inter-CE communication in GCN operations to optimize the performance and energy efficiency. Experimental evaluations with widely used datasets show up to 105x improvement in energy consumption compared to state-of-the-art GCN accelerator.

cs.AR↗

arXiv2022

Fast and Scalable Human Pose Estimation using mmWave Point Cloud

Millimeter-Wave (mmWave) radar can enable high-resolution human pose estimation with low cost and computational requirements. However, mmWave data point cloud, the primary input to processing algorithms, is highly sparse and carries significantly less information than other alternatives such as video frames. Furthermore, the scarce labeled mmWave data impedes the development of machine learning (ML) models that can generalize to unseen scenarios. We propose a fast and scalable human pose estimation (FUSE) framework that combines multi-frame representation and meta-learning to address these challenges. Experimental evaluations show that FUSE adapts to the unseen scenarios 4$\times$ faster than current supervised learning approaches and estimates human joint coordinates with about 7 cm mean absolute error.

eess.IV↗

arXiv2022

tinyMAN: Lightweight Energy Manager using Reinforcement Learning for Energy Harvesting Wearable IoT Devices

Advances in low-power electronics and machine learning techniques lead to many novel wearable IoT devices. These devices have limited battery capacity and computational power. Thus, energy harvesting from ambient sources is a promising solution to power these low-energy wearable devices. They need to manage the harvested energy optimally to achieve energy-neutral operation, which eliminates recharging requirements. Optimal energy management is a challenging task due to the dynamic nature of the harvested energy and the battery energy constraints of the target device. To address this challenge, we present a reinforcement learning-based energy management framework, tinyMAN, for resource-constrained wearable IoT devices. The framework maximizes the utilization of the target device under dynamic energy harvesting patterns and battery constraints. Moreover, tinyMAN does not rely on forecasts of the harvested energy which makes it a prediction-free approach. We deployed tinyMAN on a wearable device prototype using TensorFlow Lite for Micro thanks to its small memory footprint of less than 100 KB. Our evaluations show that tinyMAN achieves less than 2.36 ms and 27.75 $μ$J while maintaining up to 45% higher utility compared to prior approaches.

eess.SP↗

arXiv2021

Performant, Multi-objective Scheduling of Highly Interleaved Task Graphs on Heterogeneous System on Chip Devices

Performance-, power-, and energy-aware scheduling techniques play an essential role in optimally utilizing processing elements (PEs) of heterogeneous systems. List schedulers, a class of low-complexity static schedulers, have commonly been used in static execution scenarios. However, list schedulers are not suitable for runtime decision making, particularly when multiple concurrent applications are interleaved dynamically. For such cases, the static task execution times and expectation of idle PEs assumed by list schedulers lead to inefficient system utilization and poor performance. To address this problem, we present techniques for optimizing execution of list scheduling algorithms in dynamic runtime scenarios via a family of algorithms inspired by the well-known heterogeneous earliest finish time (HEFT) list scheduler. Through dynamically arriving, realistic workload scenarios that are simulated in an open-source discrete event heterogeneous SoC simulator, we exhaustively evaluate each of the proposed algorithms across two SoCs modeled after the Xilinx Zynq Ultrascale+ ZCU102 and O-Droid XU3 development boards. Altogether, depending on the chosen variant in this family of algorithms, we are able to achieve an up to 39% execution time improvement, up to 7.24x algorithmic speedup, or up to 30% energy consumption improvement compared to the baseline HEFT implementation.

cs.DC↗

arXiv2021

DAS: Dynamic Adaptive Scheduling for Energy-Efficient Heterogeneous SoCs

Domain-specific systems-on-chip (DSSoCs) aim at bridging the gap between application-specific integrated circuits (ASICs) and general-purpose processors. Traditional operating system (OS) schedulers can undermine the potential of DSSoCs since their execution times can be orders of magnitude larger than the execution time of the task itself. To address this problem, we propose a dynamic adaptive scheduling (DAS) framework that combines the benefits of a fast (low-overhead) scheduler and a slow (sophisticated, high-performance but high-overhead) scheduler. Experiments with five real-world streaming applications show that DAS consistently outperforms both the fast and slow schedulers. For 40 different workloads, DAS achieves on average 1.29x speedup and 45% lower EDP compared to the sophisticated scheduler at low data rates and 1.28x speedup and 37% lower EDP than the fast scheduler when the workload complexity increases.

cs.AR↗

arXiv2021

SIAM: Chiplet-based Scalable In-Memory Acceleration with Mesh for Deep Neural Networks

In-memory computing (IMC) on a monolithic chip for deep learning faces dramatic challenges on area, yield, and on-chip interconnection cost due to the ever-increasing model sizes. 2.5D integration or chiplet-based architectures interconnect multiple small chips (i.e., chiplets) to form a large computing system, presenting a feasible solution beyond a monolithic IMC architecture to accelerate large deep learning models. This paper presents a new benchmarking simulator, SIAM, to evaluate the performance of chiplet-based IMC architectures and explore the potential of such a paradigm shift in IMC architecture design. SIAM integrates device, circuit, architecture, network-on-chip (NoC), network-on-package (NoP), and DRAM access models to realize an end-to-end system. SIAM is scalable in its support of a wide range of deep neural networks (DNNs), customizable to various network structures and configurations, and capable of efficient design space exploration. We demonstrate the flexibility, scalability, and simulation speed of SIAM by benchmarking different state-of-the-art DNNs with CIFAR-10, CIFAR-100, and ImageNet datasets. We further calibrate the simulation results with a published silicon result, SIMBA. The chiplet-based IMC architecture obtained through SIAM shows 130$\times$ and 72$\times$ improvement in energy-efficiency for ResNet-50 on the ImageNet dataset compared to Nvidia V100 and T4 GPUs.

cs.LG↗