The Cambrian Explosion of Mixed-Precision Matrix Multiplication for Quantized Deep Learning Inference

Recent advances in deep learning (DL) have led to a shift from traditional 64-bit floating point (FP64) computations toward reduced-precision formats, such as FP16, BF16, and 8- or 16-bit integers, combined with mixed-precision arithmetic. This transition enhances computational throughput, reduces memory and bandwidth usage, and improves energy efficiency, offering significant advantages for resource-constrained edge devices. To support this shift, hardware architectures have evolved accordingly, now including adapted ISAs (Instruction Set Architectures) that expose mixed-precision vector units and matrix engines tailored for DL workloads. At the heart of many DL and scientific computing tasks is the general matrix-matrix multiplication gemm, a fundamental kernel historically optimized using axpy vector instructions on SIMD (single instruction, multiple data) units. However, as hardware moves toward mixed-precision dot-product-centric operations optimized for quantized inference, these legacy approaches are being phased out. In response to this, our paper revisits traditional high-performance gemm and describes strategies for adapting it to mixed-precision integer (MIP) arithmetic across modern ISAs, including x86_64, ARM, and RISC-V. Concretely, we illustrate novel micro-kernel designs and data layouts that better exploit today's specialized hardware and demonstrate significant performance gains from MIP arithmetic over floating-point implementations across three representative CPU architectures. These contributions highlight a new era of gemm optimization-driven by the demands of DL inference on heterogeneous architectures, marking what we term as the "Cambrian period" for matrix multiplication.

Sinusoidal Initialization, Time for a New Start

Initialization plays a critical role in Deep Neural Network training, directly influencing convergence, stability, and generalization. Common approaches such as Glorot and He initializations rely on randomness, which can produce uneven weight distributions across layer connections. In this paper, we introduce the Sinusoidal initialization, a novel deterministic method that employs sinusoidal functions to construct structured weight matrices expressly to improve the spread and balance of weights throughout the network while simultaneously fostering a more uniform, well-conditioned distribution of neuron activation states from the very first forward pass. Because Sinusoidal initialization begins with weights and activations that are already evenly and efficiently utilized, it delivers consistently faster convergence, greater training stability, and higher final accuracy across a wide range of models, including convolutional neural networks, vision transformers, and large language models. On average, our experiments show an increase of 4.8 % in final validation accuracy and 20.9 % in convergence speed. By replacing randomness with structure, this initialization provides a stronger and more reliable foundation for Deep Learning systems.

OUI Need to Talk About Weight Decay: A New Perspective on Overfitting Detection

We introduce the Overfitting-Underfitting Indicator (OUI), a novel tool for monitoring the training dynamics of Deep Neural Networks (DNNs) and identifying optimal regularization hyperparameters. Specifically, we validate that OUI can effectively guide the selection of the Weight Decay (WD) hyperparameter by indicating whether a model is overfitting or underfitting during training without requiring validation data. Through experiments on DenseNet-BC-100 with CIFAR- 100, EfficientNet-B0 with TinyImageNet and ResNet-34 with ImageNet-1K, we show that maintaining OUI within a prescribed interval correlates strongly with improved generalization and validation scores. Notably, OUI converges significantly faster than traditional metrics such as loss or accuracy, enabling practitioners to identify optimal WD (hyperparameter) values within the early stages of training. By leveraging OUI as a reliable indicator, we can determine early in training whether the chosen WD value leads the model to underfit the training data, overfit, or strike a well-balanced trade-off that maximizes validation scores. This enables more precise WD tuning for optimal performance on the tested datasets and DNNs. All code for reproducing these experiments is available at https://github.com/AlbertoFdezHdez/OUI.

GreenLightningAI: An Efficient AI System with Decoupled Structural and Quantitative Knowledge

The number and complexity of artificial intelligence (AI) applications is growing relentlessly. As a result, even with the many algorithmic and mathematical advances experienced over past decades as well as the impressive energy efficiency and computational capacity of current hardware accelerators, training the most powerful and popular deep neural networks comes at very high economic and environmental costs. Recognising that additional optimisations of conventional neural network training is very difficult, this work takes a radically different approach by proposing GreenLightningAI, a new AI system design consisting of a linear model that is capable of emulating the behaviour of deep neural networks by subsetting the model for each particular sample. The new AI system stores the information required to select the system subset for a given sample (referred to as structural information) separately from the linear model parameters (referred to as quantitative knowledge). In this paper we present a proof of concept, showing that the structural information stabilises far earlier than the quantitative knowledge. Additionally, we show experimentally that the structural information can be kept unmodified when re-training the AI system with new samples while still achieving a validation accuracy similar to that obtained when re-training a neural network with similar size. Since the proposed AI system is based on a linear model, multiple copies of the model, trained with different datasets, can be easily combined. This enables faster and greener (re)-training algorithms, including incremental re-training and federated incremental re-training.

Automatic Generators for a Family of Matrix Multiplication Routines with Apache TVM

We explore the utilization of the Apache TVM open source framework to automatically generate a family of algorithms that follow the approach taken by popular linear algebra libraries, such as GotoBLAS2, BLIS and OpenBLAS, in order to obtain high-performance blocked formulations of the general matrix multiplication (GEMM). % In addition, we fully automatize the generation process, by also leveraging the Apache TVM framework to derive a complete variety of the processor-specific micro-kernels for GEMM. This is in contrast with the convention in high performance libraries, which hand-encode a single micro-kernel per architecture using Assembly code. % In global, the combination of our TVM-generated blocked algorithms and micro-kernels for GEMM 1)~improves portability, maintainability and, globally, streamlines the software life cycle; 2)~provides high flexibility to easily tailor and optimize the solution to different data types, processor architectures, and matrix operand shapes, yielding performance on a par (or even superior for specific matrix shapes) with that of hand-tuned libraries; and 3)~features a small memory footprint.

Toward matrix multiplication for deep learning inference on the Xilinx Versal

The remarkable positive impact of Deep Neural Networks on many Artificial Intelligence (AI) tasks has led to the development of various high performance algorithms as well as specialized processors and accelerators. In this paper we address this scenario by demonstrating that the principles underlying the modern realization of the general matrix multiplication (GEMM) in conventional processor architectures, are also valid to achieve high performance for the type of operations that arise in deep learning (DL) on an exotic accelerator such as the AI Engine (AIE) tile embedded in Xilinx Versal platforms. In particular, our experimental results with a prototype implementation of the GEMM kernel, on a Xilinx Versal VCK190, delivers performance close to 86.7% of the theoretical peak that can be expected on an AIE tile, for 16-bit integer operands.