Geonhwa Jeong

Geonhwa Jeong, Sana Damani, Abhimanyu Rajeshkumar Bambhaniya et al.

Deep Learning (DL) acceleration support in CPUs has recently gained a lot of traction, with several companies (Arm, Intel, IBM) announcing products with specialized matrix engines accessible via GEMM instructions. CPUs are pervasive and need to handle diverse requirements across DL workloads running in edge/HPC/cloud platforms. Therefore, as DL workloads embrace sparsity to reduce the computations and memory size of models, it is also imperative for CPUs to add support for sparsity to avoid under-utilization of the dense matrix engine and inefficient usage of the caches and registers. This work presents VEGETA, a set of ISA and microarchitecture extensions over dense matrix engines to support flexible structured sparsity for CPUs, enabling programmable support for diverse DL models with varying degrees of sparsity. Compared to the state-of-the-art (SOTA) dense matrix engine in CPUs, a VEGETA engine provides 1.09x, 2.20x, 3.74x, and 3.28x speed-ups when running 4:4 (dense), 2:4, 1:4, and unstructured (95%) sparse DNN layers.

Hao Kang, Qingru Zhang, Souvik Kundu et al. · gatech

Key-value (KV) caching has become the de-facto to accelerate generation speed for large language models (LLMs) inference. However, the growing cache demand with increasing sequence length has transformed LLM inference to be a memory bound problem, significantly constraining the system throughput. Existing methods rely on dropping unimportant tokens or quantizing all entries uniformly. Such methods, however, often incur high approximation errors to represent the compressed matrices. The autoregressive decoding process further compounds the error of each step, resulting in critical deviation in model generation and deterioration of performance. To tackle this challenge, we propose GEAR, an efficient KV cache compression framework that achieves near-lossless high-ratio compression. GEAR first applies quantization to majority of entries of similar magnitudes to ultra-low precision. It then employs a low rank matrix to approximate the quantization error, and a sparse matrix to remedy individual errors from outlier entries. By adeptly integrating three techniques, GEAR is able to fully exploit their synergistic potentials. Our experiments demonstrate that compared to alternatives, GEAR achieves near-lossless 4-bit KV cache compression with up to 2.38x throughput improvement, while reducing peak-memory size up to 2.29x. Our code is publicly available at https://github.com/HaoKang-Timmy/GEAR.

Abhimanyu Bambhaniya, Geonhwa Jeong, Jason Park et al.

Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling higher-quality outputs at manageable serving costs. However, MoE inference at scale is fundamentally bottlenecked by expert load imbalance and inefficient token routing, especially in multi-node deployments where tokens are not guaranteed to be routed to local experts, resulting in significant inter-node all-to-all communication overhead. To systematically characterize these challenges, we profile SOTA open-source MoE models, including Llama 4 Maverick, DeepSeek V3-671B, and Qwen3-230B-A22B, on various datasets and collected over 100k real expert activation traces. Upon studying the expert activation patterns, we uncover various persistent properties across all the frontier MoE models: variable expert load imbalance, domain-specific expert activation where expert popularity shifts across task families (code, math, chat, general), and a strong correlation between prefill and decode expert activations. Motivated by these findings, we propose workload-aware micro-batch grouping and an expert placement strategy to maximize token locality to the destination expert, thereby reducing inter-node communication. Across models and datasets, these optimizations help reduce all2all communication data up to 20, resulting in lower MoE decode latency and better accelerator utilization.

Abhimanyu Bambhaniya, Ritik Raj, Geonhwa Jeong et al.

Large language models (LLMs) have shown remarkable performance across a wide range of applications, often outperforming human experts. However, deploying these gigantic models efficiently for diverse inference use cases requires carefully designed hardware platforms with ample computing, memory, and network resources. With constant innovation in LLM serving optimizations and model architecture evolving at breakneck speed, the hardware requirements to meet Service Level Objectives (SLOs) remain an open research question. To answer the question, we present an analytical tool, GenZ, to efficiently navigate the relationship between diverse LLM model architectures(Dense, GQA, MoE, Mamba), LLM serving optimizations(Chunking, Speculative decoding, quanitization), and AI platform design parameters. Our tool estimates LLM inference performance metrics for the given scenario. We have validated against real hardware platforms running various different LLM models, achieving a max geomean error of 5.82.We use GenZ to identify compute, memory capacity, memory bandwidth, network latency, and network bandwidth requirements across diverse LLM inference use cases. We also study diverse architectural choices in use today (inspired by LLM serving platforms from several vendors) to help inform computer architects designing next-generation AI hardware accelerators and platforms. The trends and insights derived from GenZ can guide AI engineers deploying LLMs as well as computer architects designing next-generation hardware accelerators and platforms. Ultimately, this work sheds light on the platform design considerations for unlocking the full potential of large language models across a spectrum of applications. The source code is available at https://github.com/abhibambhaniya/GenZ-LLM-Analyzer . Users can also be tried it on at https://genz-llm-analyzer.streamlit.app/ without any setup on your web browser.

Akshat Ramachandran, Zishen Wan, Geonhwa Jeong et al.

Traditional Deep Neural Network (DNN) quantization methods using integer, fixed-point, or floating-point data types struggle to capture diverse DNN parameter distributions at low precision, and often require large silicon overhead and intensive quantization-aware training. In this study, we introduce Logarithmic Posits (LP), an adaptive, hardware-friendly data type inspired by posits that dynamically adapts to DNN weight/activation distributions by parameterizing LP bit fields. We also develop a novel genetic-algorithm based framework, LP Quantization (LPQ), to find optimal layer-wise LP parameters while reducing representational divergence between quantized and full-precision models through a novel global-local contrastive objective. Additionally, we design a unified mixed-precision LP accelerator (LPA) architecture comprising of processing elements (PEs) incorporating LP in the computational datapath. Our algorithm-hardware co-design demonstrates on average <1% drop in top-1 accuracy across various CNN and ViT models. It also achieves ~ 2x improvements in performance per unit area and 2.2x gains in energy efficiency compared to state-of-the-art quantization accelerators using different data types.

Geonhwa Jeong, Po-An Tsai, Abhimanyu R. Bambhaniya et al.

Exploiting sparsity in deep neural networks (DNNs) has been a promising area for meeting the growing computation requirements. To minimize the overhead of sparse acceleration, hardware designers have proposed structured sparsity support, but it provides limited flexibility and requires extra model fine-tuning. Moreover, any sparse model fine-tuned for certain structured sparse HW cannot be accelerated by other structured hardware. To enable acceleration using unstructured sparsity of DNNs on structured sparse hardware, we propose an approximation method leveraging the distributive property in linear algebra to turn any sparse tensor into a series of structured sparse tensors. We also develop a software framework, TASDER, to apply high-quality structured approximation on weights and activations of DNNs. Our method accelerates dense and sparse DNNs without fine-tuning and improves energy-delay-product (EDP) by up to 83% and 74%. It achieves up to 39% speed-up on a real system.

Daniel Haziza, Timothy Chou, Dhruv Choudhary et al.

In this paper, we demonstrate how to leverage 2:4 sparsity, a popular hardware-accelerated GPU sparsity pattern, to activations to accelerate large language model training and inference. Crucially we exploit the intrinsic sparsity found in Squared-ReLU activations to provide this acceleration with no accuracy loss. Our approach achieves up to 1.3x faster Feed Forward Network (FFNs) in both the forwards and backwards pass. This work highlights the potential for sparsity to play a key role in accelerating large language model training and inference.

Geonhwa Jeong, Po-An Tsai, Stephen W. Keckler et al.

Recently, large language models (LLMs) have shown surprising performance in task-specific workloads as well as general tasks with the given prompts. However, to achieve unprecedented performance, recent LLMs use billions to trillions of parameters, which hinder the wide adaptation of those models due to their extremely large compute and memory requirements. To resolve the issue, various model compression methods are being actively investigated. In this work, we propose SDQ (Sparse Decomposed Quantization) to exploit both structured sparsity and quantization to achieve both high compute and memory efficiency. From our evaluations, we observe that SDQ can achieve 4x effective compute throughput with <1% quality drop.

Geonhwa Jeong, Eric Qin, Ananda Samajdar et al.

As AI-based applications become pervasive, CPU vendors are starting to incorporate matrix engines within the datapath to boost efficiency. Systolic arrays have been the premier architectural choice as matrix engines in offload accelerators. However, we demonstrate that incorporating them inside CPUs can introduce under-utilization and stalls due to limited register storage to amortize the fill and drain times of the array. To address this, we propose RASA, Register-Aware Systolic Array. We develop techniques to divide an execution stage into several sub-stages and overlap instructions to hide overheads and run them concurrently. RASA-based designs improve performance significantly with negligible area and power overhead.

Geonhwa Jeong, Gokcen Kestor, Prasanth Chatarasi et al.

To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these "domain-specific" accelerators are not fully programmable like CPUs and GPUs, they retain varying levels of flexibility with respect to data orchestration, i.e., dataflow and tiling optimizations to enhance efficiency. There are several challenges when designing new algorithms and mapping approaches to execute the algorithms for a target problem on new hardware. Previous works have addressed these challenges individually. To address this challenge as a whole, in this work, we present a HW-SW co-design ecosystem for spatial accelerators called Union within the popular MLIR compiler infrastructure. Our framework allows exploring different algorithms and their mappings on several accelerator cost models. Union also includes a plug-and-play library of accelerator cost models and mappers which can easily be extended. The algorithms and accelerator cost models are connected via a novel mapping abstraction that captures the map space of spatial accelerators which can be systematically pruned based on constraints from the hardware, workload, and mapper. We demonstrate the value of Union for the community with several case studies which examine offloading different tensor operations(CONV/GEMM/Tensor Contraction) on diverse accelerator architectures using different mapping schemes.

Gordon E. Moon, Hyoukjun Kwon, Geonhwa Jeong et al.

There is a growing interest in custom spatial accelerators for machine learning applications. These accelerators employ a spatial array of processing elements (PEs) interacting via custom buffer hierarchies and networks-on-chip. The efficiency of these accelerators comes from employing optimized dataflow (i.e., spatial/temporal partitioning of data across the PEs and fine-grained scheduling) strategies to optimize data reuse. The focus of this work is to evaluate these accelerator architectures using a tiled general matrix-matrix multiplication (GEMM) kernel. To do so, we develop a framework that finds optimized mappings (dataflow and tile sizes) for a tiled GEMM for a given spatial accelerator and workload combination, leveraging an analytical cost model for runtime and energy. Our evaluations over five spatial accelerators demonstrate that the tiled GEMM mappings systematically generated by our framework achieve high performance on various GEMM workloads and accelerators.

Sheng-Chun Kao, Geonhwa Jeong, Tushar Krishna

DNN accelerators provide efficiency by leveraging reuse of activations/weights/outputs during the DNN computations to reduce data movement from DRAM to the chip. The reuse is captured by the accelerator's dataflow. While there has been significant prior work in exploring and comparing various dataflows, the strategy for assigning on-chip hardware resources (i.e., compute and memory) given a dataflow that can optimize for performance/energy while meeting platform constraints of area/power for DNN(s) of interest is still relatively unexplored. The design-space of choices for balancing compute and memory explodes combinatorially, as we show in this work (e.g., as large as O(10^(72)) choices for running \mobilenet), making it infeasible to do manual-tuning via exhaustive searches. It is also difficult to come up with a specific heuristic given that different DNNs and layer types exhibit different amounts of reuse. In this paper, we propose an autonomous strategy called ConfuciuX to find optimized HW resource assignments for a given model and dataflow style. ConfuciuX leverages a reinforcement learning method, REINFORCE, to guide the search process, leveraging a detailed HW performance cost model within the training loop to estimate rewards. We also augment the RL approach with a genetic algorithm for further fine-tuning. ConfuciuX demonstrates the highest sample-efficiency for training compared to other techniques such as Bayesian optimization, genetic algorithm, simulated annealing, and other RL methods. It converges to the optimized hardware configuration 4.7 to 24 times faster than alternate techniques.