Olga Krestinskaya

The rapid scaling of Language Models (LMs) has resulted in unprecedented computational, memory, and energy requirements, making their training and deployment increasingly unsustainable. Quantization has emerged as a crucial compression technique for reducing model size, alleviating memory bottlenecks, and accelerating inference. However, while uniform low-bit quantization (e.g., INT8, INT4) provides significant efficiency gains, it can degrade accuracy in sensitive components of transformer-based LMs. Mixed-precision quantization offers a promising alternative by selectively allocating precision across layers or within tensors to strike a balance between efficiency and accuracy. This survey provides a comprehensive overview of Mixed-Precision quantization frameworks for LMs (MXPLMs). We first review quantization fundamentals, including uniform and non-uniform quantizers, quantization granularity, and methods widely used in post-training quantization. We then categorize and compare recent MXPLM frameworks according to their bit allocation strategies and precision configurations across weights, activations, and key-value caches. A comparative analysis highlights differences in perplexity, zero-shot task performance, and deployment trade-offs. Furthermore, we contrast MXPLMs with earlier mixed-precision quantization methods for deep neural networks, identifying strategies that transfer and those that face challenges in the LM setting. Then, we discuss quantization-compatible hardware, KV cache and mixture-of-experts quantization, and summarize open issues and future directions, including hardware-aware design, activation quantization, and scalable optimization methods for billion-parameter models.

Software-hardware co-design is essential for optimizing in-memory computing (IMC) hardware accelerators for neural networks. However, most existing optimization frameworks target a single workload, leading to highly specialized hardware designs that do not generalize well across models and applications. In contrast, practical deployment scenarios require a single IMC platform that can efficiently support multiple neural network workloads. This work presents a joint hardware-workload co-optimization framework based on an optimized evolutionary algorithm for designing generalized IMC accelerators. By explicitly capturing cross-workload trade-offs rather than optimizing for a single model, the proposed approach significantly reduces the performance gap between workload-specific and generalized IMC designs. The framework is evaluated on both RRAM- and SRAM-based IMC architectures, demonstrating strong robustness and adaptability across diverse design scenarios. Compared to baseline methods, the optimized designs achieve energy-delay-area product (EDAP) reductions of up to 76.2% and 95.5% when optimizing across a small set (4 workloads) and a large set (9 workloads), respectively.

To maximize hardware efficiency and performance accuracy in Compute-In-Memory (CIM)-based neural network accelerators for Artificial Intelligence (AI) applications, co-optimizing both software and hardware design parameters is essential. Manual tuning is impractical due to the vast number of parameters and their complex interdependencies. To effectively automate the design and optimization of CIM-based neural network accelerators, hardware-aware neural architecture search (HW-NAS) techniques can be applied. This work introduces CIMNAS, a joint model-quantization-hardware optimization framework for CIM architectures. CIMNAS simultaneously searches across software parameters, quantization policies, and a broad range of hardware parameters, incorporating device-,circuit-, and architecture-level co-optimizations. CIMNAS experiments were conducted over a search space of 9.9×10⁸⁵ potential parameter combinations with the MobileNet model as a baseline and RRAM-based CIM architecture. Evaluated on the ImageNet dataset, CIMNAS achieved a reduction in energy-delay-area product (EDAP) ranging from 90.1× to 104.5×, an improvement in TOPS/W between 4.68× and 4.82×, and an enhancement in TOPS/mm² from 11.3× to 12.78× relative to various baselines, all while maintaining an accuracy of 73.81%. The adaptability and robustness of CIMNAS are demonstrated by extending the framework to support the SRAM-based ResNet50 architecture, achieving up to an 819.5× reduction in EDAP. Unlike other state-of-the-art methods, CIMNAS achieves EDAP-focused optimization without any accuracy loss, generating diverse software-hardware parameter combinations for high-performance CIM-based neural network designs. The source code of CIMNAS is available at https://github.com/OlgaKrestinskaya/CIMNAS.

To maximize hardware efficiency and performance accuracy in Compute-In-Memory (CIM)-based neural network accelerators for Artificial Intelligence (AI) applications, co-optimizing both software and hardware design parameters is essential. Manual tuning is impractical due to the vast number of parameters and their complex interdependencies. To effectively automate the design and optimization of CIM-based neural network accelerators, hardware-aware neural architecture search (HW-NAS) techniques can be applied. This work introduces CIMNAS, a joint model-quantization-hardware optimization framework for CIM architectures. CIMNAS simultaneously searches across software parameters, quantization policies, and a broad range of hardware parameters, incorporating device-, circuit-, and architecture-level co-optimizations. CIMNAS experiments were conducted over a search space of 9.9x10^85 potential parameter combinations with the MobileNet model as a baseline and RRAM-based CIM architecture. Evaluated on the ImageNet dataset, CIMNAS achieved a reduction in energy-delay-area product (EDAP) ranging from 90.1x to 104.5x, an improvement in TOPS/W between 4.68x and 4.82x, and an enhancement in TOPS/mm^2 from 11.3x to 12.78x relative to various baselines, all while maintaining an accuracy of 73.81%. The adaptability and robustness of CIMNAS are demonstrated by extending the framework to support the SRAM-based ResNet50 architecture, achieving up to an 819.5x reduction in EDAP. Unlike other state-of-the-art methods, CIMNAS achieves EDAP-focused optimization without any accuracy loss, generating diverse software-hardware parameter combinations for high-performance CIM-based neural network designs. The source code of CIMNAS is available at https://github.com/OlgaKrestinskaya/CIMNAS.

Artificial Intelligence (AI) is currently experiencing a bloom driven by deep learning (DL) techniques, which rely on networks of connected simple computing units operating in parallel. The low communication bandwidth between memory and processing units in conventional von Neumann machines does not support the requirements of emerging applications that rely extensively on large sets of data. More recent computing paradigms, such as high parallelization and near-memory computing, help alleviate the data communication bottleneck to some extent, but paradigm- shifting concepts are required. Memristors, a novel beyond-complementary metal-oxide-semiconductor (CMOS) technology, are a promising choice for memory devices due to their unique intrinsic device-level properties, enabling both storing and computing with a small, massively-parallel footprint at low power. Theoretically, this directly translates to a major boost in energy efficiency and computational throughput, but various practical challenges remain. In this work we review the latest efforts for achieving hardware-based memristive artificial neural networks (ANNs), describing with detail the working principia of each block and the different design alternatives with their own advantages and disadvantages, as well as the tools required for accurate estimation of performance metrics. Ultimately, we aim to provide a comprehensive protocol of the materials and methods involved in memristive neural networks to those aiming to start working in this field and the experts looking for a holistic approach.

The amount of data processed in the cloud, the development of Internet-of-Things (IoT) applications, and growing data privacy concerns force the transition from cloud-based to edge-based processing. Limited energy and computational resources on edge push the transition from traditional von Neumann architectures to In-memory Computing (IMC), especially for machine learning and neural network applications. Network compression techniques are applied to implement a neural network on limited hardware resources. Quantization is one of the most efficient network compression techniques allowing to reduce the memory footprint, latency, and energy consumption. This paper provides a comprehensive review of IMC-based Quantized Neural Networks (QNN) and links software-based quantization approaches to IMC hardware implementation. Moreover, open challenges, QNN design requirements, recommendations, and perspectives along with an IMC-based QNN hardware roadmap are provided.

The amount of data processed in the cloud, the development of Internet-of-Things (IoT) applications, and growing data privacy concerns force the transition from cloud-based to edge-based processing. Limited energy and computational resources on edge push the transition from traditional von Neumann architectures to In-memory Computing (IMC), especially for machine learning and neural network applications. Network compression techniques are applied to implement a neural network on limited hardware resources. Quantization is one of the most efficient network compression techniques allowing to reduce the memory footprint, latency, and energy consumption. This paper provides a comprehensive review of IMC-based Quantized Neural Networks (QNN) and links software-based quantization approaches to IMC hardware implementation. Moreover, open challenges, QNN design requirements, recommendations, and perspectives along with an IMC-based QNN hardware roadmap are provided.

Miniaturization and energy efficiency are essential for building reliable edge AI computing devices using memristive crossbar accelerators. We propose that stochastic dropouts in inference stages are essential to improve the operational energy efficiency. Further, we show that the binarization of weights under stochastic processing improves the robustness in a crossbar based architecture. The inference dropouts can reduce the power consumption of memristive crossbars without compromising the performance accuracy for up to 10-15% of dropped neurons, saving significant energy in the crossbar. The architecture shows robustness to variability, hardware noise, aging, and drifting of memristor levels and memristor failures.

Gated Pixel Convolution Neural Network (Pix-eICNN) is a computationally intensive network that is useful for generating visual data. The prediction and generating pixels is a challenging but useful task for many fields such as forensics, machine vision and robotics. However, implementing PixeICNN in edge devices is a challenging task due to learning complexity and computational limits. In this paper, we present the design of neuro-memristive circuits for computing PixelCNN cells in analog domain as a coprocessor unit in edge devices. The architecture was designed using 180nm CMOS technology and carbon-chalcogenide memristive devices. On-chip area of the proposed architecture unit is 24.756mm ² , while power depends on the size of the input image and the configuration of the overall network. The power required to generate the images sequentially is 154.336mW.