arXiv subjects

George A. Constantinides

Publications and source records attributed to George A. Constantinides.

At least 19 recordsLinked to original sources

arXiv2026

When Does Disaggregation Pay? Simulating Prefill--Decode--Attention--FFN Specialization for Agentic LLM Inference

Agentic inference now dominates the LLM inference landscape, requiring LLMs to actively engage in multi-turn interactions with tool-calling capabilities. This introduces a more complex workload for the underlying inference system: serving stages such as prefill and decode exhibit substantially different behaviors and demand distinct compute and memory-bandwidth capabilities. As a result, a single homogeneous GPU system now struggles to support agentic inference, motivating an industry shift toward heterogeneous systems with disaggregated serving capabilities, such as the emerging Vera-Rubin platform with GPUs and Groq LPUs. However, the question of what the optimal hardware should look like for each component in a heterogeneous system remains underexplored. To this end, we propose a novel simulation framework for disaggregated serving, termed \textbf{HeteroPanacea}, that enables system-level simulation across three dimensions: 1) disaggregated quantization, 2) automated intra- and inter-device parallelization scheduling, and 3) PDAF (prefill-decode-attention-FFN) NPU architectural heterogeneity. By combining these three axes, we provide a cross-stack simulation framework for future heterogeneous agentic serving systems. We confirm the benefit of Prefill Decode disaggregation, simulating increased serving throughput by up to 75\% compared to traditional serving with current GPUs and demonstrate 4 way Prefill Decode Attention FFN disaggregation is the most consistent for increasing throughput across different models, assuming custom NPUs. We also investigate the relationship between model architecture and gain from disaggregation by running a set of ablation studies.

cs.DC↗

arXiv2026

CascadeLUT: Information-Ordered Streaming Inference for Bandwidth-Constrained FPGAs

Mapping neural networks to FPGAs enables low-latency, energy-efficient inference, particularly for lookup table (LUT)-based models that eliminate multipliers and map directly to reconfigurable fabric. While prior work achieves high compute efficiency, it typically assumes full-sample availability, causing pipeline stalls in bandwidth-limited streaming scenarios. Here, the bottleneck shifts from computation to data movement, as large input transfers limit throughput and energy efficiency. We present CascadeLUT, an information-structured inference framework organized around bandwidth constraints. Instead of buffering the full input, features are partitioned into ordered subsets and predictions are progressively refined as subsets arrive. The cascade statically controls which layers consume incoming features, enabling deterministic streaming inference without runtime branching. By co-designing feature scheduling with hardware dataflow, CascadeLUT reduces data movement while maintaining accuracy. Across datasets, it achieves 4.0 to 12.5 times lower latency, 3.0 to 5.0 times higher throughput and up to 13.8 times lower energy/sample than prior LUT baselines, using 1.2 to 4.4 times the LUTs of the smallest DWN baseline per task. We also demonstrate on-device input quantization integrated with LUT-based inference and present end-to-end FPGA results on real-world workloads, with 5 times reductions in quantization overhead.

cs.AR↗

arXiv2026

A3 : an Analytical Low-Rank Approximation Framework for Attention

Large language models have demonstrated remarkable performance; however, their massive parameter counts make deployment highly expensive. Low-rank approximation offers a promising compression solution, yet existing approaches have two main limitations: (1) They focus on minimizing the output error of individual linear layers, without considering the architectural characteristics of Transformers, and (2) they decompose a large weight matrix into two small low-rank matrices. Consequently, these methods often fall short compared to other compression techniques like pruning and quantization, and introduce runtime overhead such as the extra GEMM kernel launches and memory operations for decomposed small matrices. To address these limitations, we propose $A^3$, a post-training low-rank approximation framework. $A^3$ splits a Transformer layer into three functional components, namely $\texttt{QK}$, $\texttt{OV}$, and $\texttt{MLP}$ and provides analytical solutions that reduces the hidden dimension size inside each component while minimizing the component's functional loss. This approach directly reduces model sizes, KV cache sizes, and FLOPs without introducing any runtime overheads. Through extensive experiments, we show that $A^3$ maintains superior performance compared to SoTAs. For example, under the same reduction budget in computation and memory, our low-rank approximated LLaMA 3.1-70B achieves a perplexity of 4.69 on WikiText-2, outperforming the previous SoTA's 7.87 by 3.18. We also show versatile applications of $A^3$ in KV cache compression, integration with quantization, fine-tuning and mixed-rank assignments. We open-sourced our framework and code at https://github.com/DeepWok/a3.

cs.CL↗

arXiv2026

Direction-Preserving Number Representations

Low-precision number formats are widely used in modern machine learning systems due to their efficiency. Accurate direction representation is key to the accuracy of vector operations. This work precisely explores the extent to which the direction of a vector can be represented by selecting its scalar elements from a common finite alphabet of a given size. This is standard practice in machine learning, where low-precision significands may be narrow-width floating-point or integer values. A geometric framework is introduced for analyzing the directional coverage of such product-structured codes. This work analytically quantifies the suboptimality gap between such product-structured codes and spherical codes for the vector as a whole, in both low and asymptotically high dimensions. Furthermore, within the product code class, it is proven that the standard formats of two's complement, fixed-point, and floating-point are suboptimal, again with quantified gap, pointing to the potential to develop new scalar number formats. Such scalar alphabets are numerically optimized across multiple block dimensions for directional coverage, including the dimension used in NVIDIA's NVFP4 format. Experimental results are presented comparing the performance of standard formats and the optimized alphabet. We find that for four bits, NVIDIA's choice of E2M1 closely approximates the optimized alphabet, providing a geometric explanation for its strong performance in low-precision machine learning workloads and an analytical understanding of the link between that superiority and block size. We provide open-source formal proofs in Lean for the theorems in this work, along with the experimental code and the optimized alphabets obtained.

cs.LG↗

arXiv2025

Training with Fewer Bits: Unlocking Edge LLMs Training with Stochastic Rounding

LLM training is resource-intensive. Quantized training improves computational and memory efficiency but introduces quantization noise, which can hinder convergence and degrade model accuracy. Stochastic Rounding (SR) has emerged as a theoretically attractive alternative to deterministic rounding, offering unbiased gradient estimates. However, its interaction with other training factors -- especially batch size -- remains under explored. In this paper, we present a theoretical and empirical study of mini-batch stochastic gradient descent (SGD) with SR, showing that increased batch sizes can compensate for reduced precision during back-propagation. Furthermore, we show that quantizing weights and activations impacts gradient variance in distinct ways. Our experiments validate these theoretical insights.

cs.LG↗

arXiv2025

An Improved Template for Approximate Computing

Deploying neural networks on edge devices entails a careful balance between the energy required for inference and the accuracy of the resulting classification. One technique for navigating this tradeoff is approximate computing: the process of reducing energy consumption by slightly reducing the accuracy of arithmetic operators. In this context, we propose a methodology to reduce the area of the small arithmetic operators used in neural networks - i.e., adders and multipliers - via a small loss in accuracy, and show that we improve area savings for the same accuracy loss w.r.t. the state of the art. To achieve our goal, we improve on a boolean rewriting technique recently proposed, called XPAT, where the use of a parametrisable template to rewrite circuits has proved to be highly beneficial. In particular, XPAT was able to produce smaller circuits than comparable approaches while utilising a naive sum of products template structure. In this work, we show that template parameters can act as proxies for chosen metrics and we propose a novel template based on parametrisable product sharing that acts as a close proxy to synthesised area. We demonstrate experimentally that our methodology converges better to low-area solutions and that it can find better approximations than both the original XPAT and two other state-of-the-art approaches.

cs.LG↗

arXiv2025

BitMoD: Bit-serial Mixture-of-Datatype LLM Acceleration

Large language models (LLMs) have demonstrated remarkable performance across various machine learning tasks. Yet the substantial memory footprint of LLMs significantly hinders their deployment. In this paper, we improve the accessibility of LLMs through BitMoD, an algorithm-hardware co-design solution that enables efficient LLM acceleration at low weight precision. On the algorithm side, BitMoD introduces fine-grained data type adaptation that uses a different numerical data type to quantize a group of (e.g., 128) weights. Through the careful design of these new data types, BitMoD is able to quantize LLM weights to very low precision (e.g., 4 bits and 3 bits) while maintaining high accuracy. On the hardware side, BitMoD employs a bit-serial processing element to easily support multiple numerical precisions and data types; our hardware design includes two key innovations: First, it employs a unified representation to process different weight data types, thus reducing the hardware cost. Second, it adopts a bit-serial dequantization unit to rescale the per-group partial sum with minimal hardware overhead. Our evaluation on six representative LLMs demonstrates that BitMoD significantly outperforms state-of-the-art LLM quantization and acceleration methods. For discriminative tasks, BitMoD can quantize LLM weights to 4-bit with $<\!0.5\%$ accuracy loss on average. For generative tasks, BitMoD is able to quantize LLM weights to 3-bit while achieving better perplexity than prior LLM quantization scheme. Combining the superior model performance with an efficient accelerator design, BitMoD achieves an average of $1.69\times$ and $1.48\times$ speedups compared to prior LLM accelerators ANT and OliVe, respectively.

cs.LG↗

arXiv2025

ATHEENA: A Toolflow for Hardware Early-Exit Network Automation

The continued need for improvements in accuracy, throughput, and efficiency of Deep Neural Networks has resulted in a multitude of methods that make the most of custom architectures on FPGAs. These include the creation of hand-crafted networks and the use of quantization and pruning to reduce extraneous network parameters. However, with the potential of static solutions already well exploited, we propose to shift the focus to using the varying difficulty of individual data samples to further improve efficiency and reduce average compute for classification. Input-dependent computation allows for the network to make runtime decisions to finish a task early if the result meets a confidence threshold. Early-Exit network architectures have become an increasingly popular way to implement such behaviour in software. We create: A Toolflow for Hardware Early-Exit Network Automation (ATHEENA), an automated FPGA toolflow that leverages the probability of samples exiting early from such networks to scale the resources allocated to different sections of the network. The toolflow uses the data-flow model of fpgaConvNet, extended to support Early-Exit networks as well as Design Space Exploration to optimize the generated streaming architecture hardware with the goal of increasing throughput/reducing area while maintaining accuracy. Experimental results on three different networks demonstrate a throughput increase of $2.00\times$ to $2.78\times$ compared to an optimized baseline network implementation with no early exits. Additionally, the toolflow can achieve a throughput matching the same baseline with as low as $46\%$ of the resources the baseline requires.

cs.AR↗

arXiv2025

NeuraLUT-Assemble: Hardware-aware Assembling of Sub-Neural Networks for Efficient LUT Inference

Efficient neural networks (NNs) leveraging lookup tables (LUTs) have demonstrated significant potential for emerging AI applications, particularly when deployed on field-programmable gate arrays (FPGAs) for edge computing. These architectures promise ultra-low latency and reduced resource utilization, broadening neural network adoption in fields such as particle physics. However, existing LUT-based designs suffer from accuracy degradation due to the large fan-in required by neurons being limited by the exponential scaling of LUT resources with input width. In practice, in prior work this tension has resulted in the reliance on extremely sparse models. We present NeuraLUT-Assemble, a novel framework that addresses these limitations by combining mixed-precision techniques with the assembly of larger neurons from smaller units, thereby increasing connectivity while keeping the number of inputs of any given LUT manageable. Additionally, we introduce skip-connections across entire LUT structures to improve gradient flow. NeuraLUT-Assemble closes the accuracy gap between LUT-based methods and (fully-connected) MLP-based models, achieving competitive accuracy on tasks such as network intrusion detection, digit classification, and jet classification, demonstrating up to $8.42\times$ reduction in the area-delay product compared to the state-of-the-art at the time of the publication.

cs.LG↗

arXiv2025

Banked Memories for Soft SIMT Processors

Recent advances in soft GPGPU architectures have shown that a small (<10K LUT), high performance (770 MHz) processor is possible in modern FPGAs. In this paper we architect and evaluate soft SIMT processor banked memories, which can support high bandwidth (up to 16 ports) while maintaining high speed (over 770 MHz). We compare 9 different memory architectures, including simpler multi-port memories, and run a total of 51 benchmarks (different combinations of algorithms, data sizes and processor memories) to develop a comprehensive set of data which will guide the reader in making an informed memory architecture decision for their application. Our benchmarks are comprised of matrix transpositions (memory intensive) and FFTs (split between memory accesses, floating point, and integer computations) to provide a balanced evaluation. We show that the simpler (but more memory block intensive) multi-port memories offer higher performance than the more architecturally complex banked memories for many applications, especially for smaller memories, but the effective footprint cost of the multi-port memories quickly becomes prohibitive as dataset sizes increase. Our banked memory implementation results - high bandwidth, high Fmax, and high density - can be used for other FPGA applications as well, such as HLS (High Level Synthesis).

cs.AR↗

arXiv2025

QERA: an Analytical Framework for Quantization Error Reconstruction

The growing number of parameters and computational demands of large language models (LLMs) present significant challenges for their efficient deployment. Recently, there is an increasing interest in quantizing weights to extremely low precision while offsetting the resulting error with low-rank, high-precision error reconstruction terms. The combination of quantization and low-rank approximation is now popular in both adapter-based, parameter-efficient fine-tuning methods such as LoftQ and low-precision inference techniques including ZeroQuant-V2. Usually, the low-rank terms are calculated via the singular value decomposition (SVD) of the weight quantization error, minimizing the Frobenius and spectral norms of the weight approximation error. Recent methods like LQ-LoRA and LQER introduced hand-crafted heuristics to minimize errors in layer outputs (activations) rather than weights, resulting improved quantization results. However, these heuristic methods lack an analytical solution to guide the design of quantization error reconstruction terms. In this paper, we revisit this problem and formulate an analytical framework, named Quantization Error Reconstruction Analysis (QERA), and offer a closed-form solution to the problem. We show QERA benefits both existing low-precision fine-tuning and inference methods -- QERA achieves a fine-tuned accuracy gain of $Δ_{\text{acc}}$ = 6.05% of 2-bit RoBERTa-base on GLUE compared to LoftQ; and obtains $Δ_{\text{acc}}$ = 2.97% higher post-training quantization accuracy of 4-bit Llama-3.1-70B on average than ZeroQuant-V2 and $Δ_{\text{ppl}}$ = - 0.28 lower perplexity on WikiText2 than LQER.

cs.LG↗

arXiv2025

PolyLUT: Ultra-low Latency Polynomial Inference with Hardware-Aware Structured Pruning

Standard deep neural network inference involves the computation of interleaved linear maps and nonlinear activation functions. Prior work for ultra-low latency implementations has hardcoded these operations inside FPGA lookup tables (LUTs). However, FPGA LUTs can implement a much greater variety of functions. In this paper, we propose a novel approach to training DNNs for FPGA deployment using multivariate polynomials as the basic building block. Our method takes advantage of the flexibility offered by the soft logic, hiding the polynomial evaluation inside the LUTs with minimal overhead. By using polynomial building blocks, we achieve the same accuracy using considerably fewer layers of soft logic than by using linear functions, leading to significant latency and area improvements. LUT-based implementations also face a significant challenge: the LUT size grows exponentially with the number of inputs. Prior work relies on a priori fixed sparsity, with results heavily dependent on seed selection. To address this, we propose a structured pruning strategy using a bespoke hardware-aware group regularizer that encourages a particular sparsity pattern that leads to a small number of inputs per neuron. We demonstrate the effectiveness of PolyLUT on three tasks: network intrusion detection, jet identification at the CERN Large Hadron Collider, and MNIST.

cs.LG↗

arXiv2024

ReducedLUT: Table Decomposition with "Don't Care" Conditions

Lookup tables (LUTs) are frequently used to efficiently store arrays of precomputed values for complex mathematical computations. When used in the context of neural networks, these functions exhibit a lack of recognizable patterns which presents an unusual challenge for conventional logic synthesis techniques. Several approaches are known to break down a single large lookup table into multiple smaller ones that can be recombined. Traditional methods, such as plain tabulation, piecewise linear approximation, and multipartite table methods, often yield inefficient hardware solutions when applied to LUT-based NNs. This paper introduces ReducedLUT, a novel method to reduce the footprint of the LUTs by injecting don't cares into the compression process. This additional freedom introduces more self-similarities which can be exploited using known decomposition techniques. We then demonstrate a particular application to machine learning; by replacing unobserved patterns within the training data of neural network models with don't cares, we enable greater compression with minimal model accuracy degradation. In practice, we achieve up to $1.63\times$ reduction in Physical LUT utilization, with a test accuracy drop of no more than $0.01$ accuracy points.

cs.AR↗

arXiv2024

NeuraLUT: Hiding Neural Network Density in Boolean Synthesizable Functions

Field-Programmable Gate Array (FPGA) accelerators have proven successful in handling latency- and resource-critical deep neural network (DNN) inference tasks. Among the most computationally intensive operations in a neural network (NN) is the dot product between the feature and weight vectors. Thus, some previous FPGA acceleration works have proposed mapping neurons with quantized inputs and outputs directly to lookup tables (LUTs) for hardware implementation. In these works, the boundaries of the neurons coincide with the boundaries of the LUTs. We propose relaxing these boundaries and mapping entire sub-networks to a single LUT. As the sub-networks are absorbed within the LUT, the NN topology and precision within a partition do not affect the size of the lookup tables generated. Therefore, we utilize fully connected layers with floating-point precision inside each partition, which benefit from being universal function approximators, but with rigid sparsity and quantization enforced between partitions, where the NN topology becomes exposed to the circuit topology. Although cheap to implement, this approach can lead to very deep NNs, and so to tackle challenges like vanishing gradients, we also introduce skip connections inside the partitions. The resulting methodology can be seen as training DNNs with a specific FPGA hardware-inspired sparsity pattern that allows them to be mapped to much shallower circuit-level networks, thereby significantly improving latency. We validate our proposed method on a known latency-critical task, jet substructure tagging, and on the classical computer vision task, digit classification using MNIST. Our approach allows for greater function expressivity within the LUTs compared to existing work, leading to up to $4.3\times$ lower latency NNs for the same accuracy.

cs.AR↗

arXiv2024

Exploring FPGA designs for MX and beyond

A number of companies recently worked together to release the new Open Compute Project MX standard for low-precision computation, aimed at efficient neural network implementation. In this paper, we describe and evaluate the first open-source FPGA implementation of the arithmetic defined in the standard. Our designs fully support all the standard's concrete formats for conversion into and out of MX formats and for the standard-defined arithmetic operations, as well as arbitrary fixed-point and floating-point formats. Certain elements of the standard are left as implementation-defined, and we present the first concrete FPGA-inspired choices for these elements, which we outline in the paper. Our library of optimized hardware components is available open source, and can be used to build larger systems. For this purpose, we also describe and release an open-source Pytorch library for quantization into the new standard, integrated with the Brevitas library so that the community can develop novel neural network designs quantized with MX formats in mind. We demonstrate the usability and efficacy of our libraries via the implementation of example neural networks such as ResNet-18 on the ImageNet ILSVRC12 dataset. Our testing shows that MX is very effective for formats such as INT5 or FP6 which are not natively supported on GPUs. This gives FPGAs an advantage as they have the flexibility to implement a custom datapath and take advantage of the smaller area footprints offered by these formats.

cs.AR↗

arXiv2024

Unlocking the Global Synergies in Low-Rank Adapters

Low-rank Adaption (LoRA) has been the de-facto parameter-efficient fine-tuning technique for large language models. We present HeteroLoRA, a light-weight search algorithm that leverages zero-cost proxies to allocate the limited LoRA trainable parameters across the model for better fine-tuned performance. In addition to the allocation for the standard LoRA-adapted models, we also demonstrate the efficacy of HeteroLoRA by performing the allocation in a more challenging search space that includes LoRA modules and LoRA-adapted shortcut connections. Experiments show that HeteroLoRA enables improvements in model performance given the same parameter budge. For example, on MRPC, we see an improvement of 1.6% in accuracy with similar training parameter budget. We will open-source our algorithm once the paper is accepted.

cs.LG↗

arXiv2024

Optimised Grouped-Query Attention Mechanism for Transformers

Grouped-query attention (GQA) has been widely adopted in LLMs to mitigate the complexity of multi-head attention (MHA). To transform an MHA to a GQA, neighbour queries in MHA are evenly split into groups where each group shares the value and key layers. In this work, we propose AsymGQA, an activation-informed approach to asymmetrically grouping an MHA to a GQA for better model performance. Our AsymGQA outperforms the GQA within the same model size budget. For example, AsymGQA LLaMA-2-7B has an accuracy increase of 7.5% on MMLU compared to neighbour grouping. Our approach addresses the GQA's trade-off problem between model performance and hardware efficiency.

cs.LG↗

arXiv2024

ROVER: RTL Optimization via Verified E-Graph Rewriting

Manual RTL design and optimization remains prevalent across the semiconductor industry because commercial logic and high-level synthesis tools are unable to match human designs. Our experience in industrial datapath design demonstrates that manual optimization can typically be decomposed into a sequence of local equivalence preserving transformations. By formulating datapath optimization as a graph rewriting problem we automate design space exploration in a tool we call ROVER. We develop a set of mixed precision RTL rewrite rules inspired by designers at Intel and an accompanying automated validation framework. A particular challenge in datapath design is to determine a productive order in which to apply transformations as this can be design dependent. ROVER resolves this problem by building upon the e-graph data structure, which compactly represents a design space of equivalent implementations. By applying rewrites to this data structure, ROVER generates a set of efficient and functionally equivalent design options. From the ROVER generated e-graph we select an efficient implementation. To accurately model the circuit area we develop a theoretical cost metric and then an integer linear programming model to extract the optimal implementation. To build trust in the generated design ROVER also produces a back-end verification certificate that can be checked using industrial tools. We apply ROVER to both Intel-provided and open-source benchmarks, and see up to a 63% reduction in circuit area. ROVER is also able to generate a customized library of distinct implementations from a given parameterizable RTL design, improving circuit area across the range of possible instantiations.

cs.AR↗