Ulf Schlichtmann

Philipp van Kempen, Rafael Stahl, Daniel Mueller-Gritschneder et al.

While there exist many ways to deploy machine learning models on microcontrollers, it is non-trivial to choose the optimal combination of frameworks and targets for a given application. Thus, automating the end-to-end benchmarking flow is of high relevance nowadays. A tool called MLonMCU is proposed in this paper and demonstrated by benchmarking the state-of-the-art TinyML frameworks TFLite for Microcontrollers and TVM effortlessly with a large number of configurations in a low amount of time.

Amro Eldebiky, Grace Li Zhang, Georg Boecherer et al.

The last decade has witnessed the breakthrough of deep neural networks (DNNs) in many fields. With the increasing depth of DNNs, hundreds of millions of multiply-and-accumulate (MAC) operations need to be executed. To accelerate such operations efficiently, analog in-memory computing platforms based on emerging devices, e.g., resistive RAM (RRAM), have been introduced. These acceleration platforms rely on analog properties of the devices and thus suffer from process variations and noise. Consequently, weights in neural networks configured into these platforms can deviate from the expected values, which may lead to feature errors and a significant degradation of inference accuracy. To address this issue, in this paper, we propose a framework to enhance the robustness of neural networks under variations and noise. First, a modified Lipschitz constant regularization is proposed during neural network training to suppress the amplification of errors propagated through network layers. Afterwards, error compensation is introduced at necessary locations determined by reinforcement learning to rescue the feature maps with remaining errors. Experimental results demonstrate that inference accuracy of neural networks can be recovered from as low as 1.69% under variations and noise back to more than 95% of their original accuracy, while the training and hardware cost are negligible.

Wenhao Sun, Bing Li, Grace Li Zhang et al.

While large language models (LLMs) have demonstrated the ability to generate hardware description language (HDL) code for digital circuits, they still suffer from the hallucination problem, which leads to the generation of incorrect HDL code or misunderstanding of specifications. In this work, we introduce a human-expert-inspired method to mitigate the hallucination of LLMs and improve the performance in HDL code generation. We first let LLMs classify the type of the circuit based on the specifications. Then, according to the type of the circuit, we split the tasks into several sub-procedures, including information extraction and human-like design flow using Electronic Design Automation (EDA) tools. Besides, we also use a search method to mitigate the variation in code generation. Experimental results show that our method can significantly improve the functional correctness of the generated Verilog and reduce the hallucination of LLMs.

Wenhao Sun, Grace Li Zhang, Huaxi Gu et al.

In deep neural networks (DNNs), there are a huge number of weights and multiply-and-accumulate (MAC) operations. Accordingly, it is challenging to apply DNNs on resource-constrained platforms, e.g., mobile phones. Quantization is a method to reduce the size and the computational complexity of DNNs. Existing quantization methods either require hardware overhead to achieve a non-uniform quantization or focus on model-wise and layer-wise uniform quantization, which are not as fine-grained as filter-wise quantization. In this paper, we propose a class-based quantization method to determine the minimum number of quantization bits for each filter or neuron in DNNs individually. In the proposed method, the importance score of each filter or neuron with respect to the number of classes in the dataset is first evaluated. The larger the score is, the more important the filter or neuron is and thus the larger the number of quantization bits should be. Afterwards, a search algorithm is adopted to exploit the different importance of filters and neurons to determine the number of quantization bits of each filter or neuron. Experimental results demonstrate that the proposed method can maintain the inference accuracy with low bit-width quantization. Given the same number of quantization bits, the proposed method can also achieve a better inference accuracy than the existing methods.

Richard Petri, Grace Li Zhang, Yiran Chen et al.

Deep neural networks (DNNs) have been successfully applied in various fields. A major challenge of deploying DNNs, especially on edge devices, is power consumption, due to the large number of multiply-and-accumulate (MAC) operations. To address this challenge, we propose PowerPruning, a novel method to reduce power consumption in digital neural network accelerators by selecting weights that lead to less power consumption in MAC operations. In addition, the timing characteristics of the selected weights together with all activation transitions are evaluated. The weights and activations that lead to small delays are further selected. Consequently, the maximum delay of the sensitized circuit paths in the MAC units is reduced even without modifying MAC units, which thus allows a flexible scaling of supply voltage to reduce power consumption further. Together with retraining, the proposed method can reduce power consumption of DNNs on hardware by up to 78.3% with only a slight accuracy loss.

Wenhao Sun, Grace Li Zhang, Xunzhao Yin et al.

Deep neural networks (DNNs) have successfully been applied in many fields in the past decades. However, the increasing number of multiply-and-accumulate (MAC) operations in DNNs prevents their application in resource-constrained and resource-varying platforms, e.g., mobile phones and autonomous vehicles. In such platforms, neural networks need to provide acceptable results quickly and the accuracy of the results should be able to be enhanced dynamically according to the computational resources available in the computing system. To address these challenges, we propose a design framework called SteppingNet. SteppingNet constructs a series of subnets whose accuracy is incrementally enhanced as more MAC operations become available. Therefore, this design allows a trade-off between accuracy and latency. In addition, the larger subnets in SteppingNet are built upon smaller subnets, so that the results of the latter can directly be reused in the former without recomputation. This property allows SteppingNet to decide on-the-fly whether to enhance the inference accuracy by executing further MAC operations. Experimental results demonstrate that SteppingNet provides an effective incremental accuracy improvement and its inference accuracy consistently outperforms the state-of-the-art work under the same limit of computational resources.

Chuangtao Chen, Grace Li Zhang, Xunzhao Yin et al.

Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable ``packets'' wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen2.5 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline.

Amro Eldebiky, Grace Li Zhang, Xunzhao Yin et al.

Deep neural networks (DNNs) have made breakthroughs in various fields including image recognition and language processing. DNNs execute hundreds of millions of multiply-and-accumulate (MAC) operations. To efficiently accelerate such computations, analog in-memory-computing platforms have emerged leveraging emerging devices such as resistive RAM (RRAM). However, such accelerators face the hurdle of being required to have sufficient on-chip crossbars to hold all the weights of a DNN. Otherwise, RRAM cells in the crossbars need to be reprogramed to process further layers, which causes huge time/energy overhead due to the extremely slow writing and verification of the RRAM cells. As a result, it is still not possible to deploy such accelerators to process large-scale DNNs in industry. To address this problem, we propose the BasisN framework to accelerate DNNs on any number of available crossbars without reprogramming. BasisN introduces a novel representation of the kernels in DNN layers as combinations of global basis vectors shared between all layers with quantized coefficients. These basis vectors are written to crossbars only once and used for the computations of all layers with marginal hardware modification. BasisN also provides a novel training approach to enhance computation parallelization with the global basis vectors and optimize the coefficients to construct the kernels. Experimental results demonstrate that cycles per inference and energy-delay product were reduced to below 1% compared with applying reprogramming on crossbars in processing large-scale DNNs such as DenseNet and ResNet on ImageNet and CIFAR100 datasets, while the training and hardware costs are negligible.

Chuangtao Chen, Grace Li Zhang, Xunzhao Yin et al.

Deep neural networks (DNNs) have been widely deployed across diverse domains such as computer vision and natural language processing. However, the impressive accomplishments of DNNs have been realized alongside extensive computational demands, thereby impeding their applicability on resource-constrained devices. To address this challenge, many researchers have been focusing on basic neuron structures, the fundamental building blocks of neural networks, to alleviate the computational and storage cost. In this work, an efficient quadratic neuron architecture distinguished by its enhanced utilization of second-order computational information is introduced. By virtue of their better expressivity, DNNs employing the proposed quadratic neurons can attain similar accuracy with fewer neurons and computational cost. Experimental results have demonstrated that the proposed quadratic neuron structure exhibits superior computational and storage efficiency across various tasks when compared with both linear and non-linear neurons in prior work.

Rafael Stahl, Daniel Mueller-Gritschneder, Ulf Schlichtmann

Memory optimization for deep neural network (DNN) inference gains high relevance with the emergence of TinyML, which refers to the deployment of DNN inference tasks on tiny, low-power microcontrollers. Applications such as audio keyword detection or radar-based gesture recognition are heavily constrained by the limited memory on such tiny devices because DNN inference requires large intermediate run-time buffers to store activations and other intermediate data, which leads to high memory usage. In this paper, we propose a new Fused Depthwise Tiling (FDT) method for the memory optimization of DNNs, which, compared to existing tiling methods, reduces memory usage without inducing any run time overhead. FDT applies to a larger variety of network layers than existing tiling methods that focus on convolutions. It improves TinyML memory optimization significantly by reducing memory of models where this was not possible before and additionally providing alternative design points for models that show high run time overhead with existing methods. In order to identify the best tiling configuration, an end-to-end flow with a new path discovery method is proposed, which applies FDT and existing tiling methods in a fully automated way, including the scheduling of the operations and planning of the layout of buffers in memory. Out of seven evaluated models, FDT achieved significant memory reduction for two models by 76.2% and 18.1% where existing tiling methods could not be applied. Two other models showed a significant run time overhead with existing methods and FDT provided alternative design points with no overhead but reduced memory savings.

Sijie Fei, Grace Li Zhang, Bing Li et al.

Distributed learning is widely used for training large models on large datasets by distributing parts of the model or dataset across multiple devices and aggregating the computed results for subsequent computations or parameter updates. Existing communication algorithms for distributed learning such as ring all-reduce result in heavy communication overhead between servers. Since communication in large-scale systems uses optical fibers, we propose an Optical In-Network-Computing (OptINC) architecture to offload the computation in servers onto the optical interconnects. To execute gradient averaging and quantization in the optical domain, we incorporate optical devices such as Mach-Zehnder-Interferometers (MZIs) into the interconnects. Such a de facto optical neural network (ONN) can effectively reduce the communication overhead in existing distributed training solutions. To reduce dataset complexity for training this neural network, a preprocessing algorithm implemented in the optical domain is also proposed. Hardware cost is lowered by approximating the weight matrices of the optical neural network with unitary and diagonal matrices, while the accuracy is maintained by a proposed hardware-aware training algorithm. The proposed solution was evaluated on real distributed learning tasks, including ResNet50 on CIFAR-100, and a LLaMA-based network on Wikipedia-1B. In both cases, the proposed framework can achieve comparable training accuracy to the ring all-reduce baseline, while eliminating communication overhead.

Kangwei Xu, Bing Li, Ulf Schlichtmann

As chip complexity increases and time-to-market pressures grow, front-end design has become a critical bottleneck in chip development. Recently, Large Language Models (LLMs) have shown great potential in Electronic Design Automation (EDA). Beyond specification understanding, LLMs show the potential to serve as a unified intelligent interface for hardware description language (HDL) generation, testbench construction, and design space exploration. The rise of agentic AI, represented by pioneering systems such as OpenClaw, offers a strategic roadmap for the next generation EDA. From this perspective, this paper discusses the evolution of EDA from localized assistance to autonomous agentic execution. Then, we review representative advances of LLMs in front-end design, focusing on key tasks such as circuit and testbench generation from a shared specification, as well as design quality improvement in established workflows such as high-level synthesis. Finally, we discuss the key challenges and limitations of integrating LLMs into EDA, and outline future opportunities for advancing LLM-enabled front-end design, offering a systematic perspective for researchers interested in leveraging agentic AI technologies for EDA.

Bo Liu, Grace Li Zhang, Xunzhao Yin et al.

Deep neural networks (DNNs) have achieved great breakthroughs in many fields such as image classification and natural language processing. However, the execution of DNNs needs to conduct massive numbers of multiply-accumulate (MAC) operations on hardware and thus incurs a large power consumption. To address this challenge, we propose a novel digital MAC design based on encoding. In this new design, the multipliers are replaced by simple logic gates to represent the results with a wide bit representation. The outputs of the new multipliers are added by bit-wise weighted accumulation and the accumulation results are compatible with existing computing platforms accelerating neural networks. Since the multiplication function is replaced by a simple logic representation, the critical paths in the resulting circuits become much shorter. Correspondingly, pipelining stages and intermediate registers used to store partial sums in the MAC array can be reduced, leading to a significantly smaller area as well as better power efficiency. The proposed design has been synthesized and verified by ResNet18- Cifar10, ResNet20-Cifar100, ResNet50-ImageNet, MobileNetV2-Cifar10, MobileNetV2-Cifar100, and EfficientNetB0-ImageNet. The experimental results confirmed the reduction of circuit area by up to 48.79% and the reduction of power consumption of executing DNNs by up to 64.41%, while the accuracy of the neural networks can still be well maintained. The open source code of this work can be found on GitHub with link https://github.com/Bo-Liu-TUM/EncodingNet/.

Samira Ahmadifarsani, Daniel Mueller-Gritschneder, Ulf Schlichtmann

The growing adoption of domain-specific architectures in edge computing platforms for deep learning has highlighted the efficiency of hardware accelerators. However, integrating custom accelerators into modern machine learning (ML) compilers remains a complex challenge due to the need for significant modifications in compilation layers and specialized scheduling techniques. Existing frameworks offer partial solutions and require users to navigate intricate compiler internals. In this paper, we introduce a TVM-based compilation integration approach that targets GEMM-based deep learning accelerators. Our approach abstracts the complexities of compiler integration, enabling seamless integration of accelerators without requiring in-depth knowledge of the underlying compiler. Furthermore, we extend and incorporate design space exploration tools, specifically CoSA, to automate efficient tensor scheduling, accounting for factors such as uneven mapping and double buffering. Our framework is benchmarked on the Gemmini accelerator, demonstrating performance comparable to its specialized manually implemented toolchain.

Chuangtao Chen, Grace Li Zhang, Xunzhao Yin et al.

In this paper, we introduce LiveMind, a novel low-latency inference framework for large language model (LLM) inference which enables LLMs to perform inferences with incomplete user input. By reallocating computational processes to the input phase, a substantial reduction in latency is achieved, thereby significantly enhancing the interactive experience for users of LLMs. The framework adeptly manages the visibility of the streaming input to the model, allowing it to infer from incomplete user input or await additional content. Compared with traditional inference methods on complete user input, our approach demonstrates an average reduction in response latency of 84.0% on the MMLU dataset and 71.6% on the MMLU-Pro dataset, while maintaining comparable accuracy. Additionally, our framework facilitates collaborative inference and output across different models. By employing an large LLM for inference and a small LLM for output, we achieve an average 37% reduction in response latency, alongside a 4.30% improvement in accuracy on the MMLU-Pro dataset compared with the baseline. The proposed LiveMind framework advances the field of human-AI interaction by enabling more responsive and efficient communication between users and AI systems.

Christopher Wolters, Xiaoxuan Yang, Ulf Schlichtmann et al.

Large language models (LLMs) have recently transformed natural language processing, enabling machines to generate human-like text and engage in meaningful conversations. This development necessitates speed, efficiency, and accessibility in LLM inference as the computational and memory requirements of these systems grow exponentially. Meanwhile, advancements in computing and memory capabilities are lagging behind, exacerbated by the discontinuation of Moore's law. With LLMs exceeding the capacity of single GPUs, they require complex, expert-level configurations for parallel processing. Memory accesses become significantly more expensive than computation, posing a challenge for efficient scaling, known as the memory wall. Here, compute-in-memory (CIM) technologies offer a promising solution for accelerating AI inference by directly performing analog computations in memory, potentially reducing latency and power consumption. By closely integrating memory and compute elements, CIM eliminates the von Neumann bottleneck, reducing data movement and improving energy efficiency. This survey paper provides an overview and analysis of transformer-based models, reviewing various CIM architectures and exploring how they can address the imminent challenges of modern AI computing systems. We discuss transformer-related operators and their hardware acceleration schemes and highlight challenges, trends, and insights in corresponding CIM designs.

Felix Last, Ceren Yeni, Ulf Schlichtmann

As the relative power, performance, and area (PPA) impact of embedded memories continues to grow, proper parameterization of each of the thousands of memories on a chip is essential. When the parameters of all memories of a product are optimized together as part of a single system, better trade-offs may be achieved than if the same memories were optimized in isolation. However, challenges such as a sparse solution space, conflicting objectives, and computationally expensive PPA estimation impede the application of common optimization heuristics. We show how the memory system optimization problem can be solved through computational intelligence. We apply a Pareto-based Differential Evolution to ensure unbiased optimization of multiple PPA objectives. To ensure efficient exploration of a sparse solution space, we repair individuals to yield feasible parameterizations. PPA is estimated efficiently in large batches by pre-trained regression neural networks. Our framework enables the system optimization of thousands of memories while keeping a small resource footprint. Evaluating our method on a tractable system, we find that our method finds diverse solutions which exhibit less than 0.5% distance from known global optima.

Grace Li Zhang, Bing Li, Meng Li et al.

With recent advances in reverse engineering, attackers can reconstruct a netlist to counterfeit chips by opening the die and scanning all layers of authentic chips. This relatively easy counterfeiting is made possible by the use of the standard simple clocking scheme, where all combinational blocks function within one clock period, so that a netlist of combinational logic gates and flip-flops is sufficient to duplicate a design. In this paper, we propose to invalidate the assumption that a netlist completely represents the function of a circuit with unconventional timing. With the introduced wave-pipelining paths, attackers have to capture gate and interconnect delays during reverse engineering, or to test a huge number of combinational paths to identify the wave-pipelining paths. To hinder the test-based attack, we construct false paths with wave-pipelining to increase the counterfeiting challenge. Experimental results confirm that wave-pipelining true paths and false paths can be constructed in benchmark circuits successfully with only a negligible cost, thus thwarting the potential attack techniques.