Mahdi Nazemi

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Seyedarmin Azizi, Mahdi Nazemi, Arash Fayyazi et al.

As the complexity and computational demands of deep learning models rise, the need for effective optimization methods for neural network designs becomes paramount. This work introduces an innovative search mechanism for automatically selecting the best bit-width and layer-width for individual neural network layers. This leads to a marked enhancement in deep neural network efficiency. The search domain is strategically reduced by leveraging Hessian-based pruning, ensuring the removal of non-crucial parameters. Subsequently, we detail the development of surrogate models for favorable and unfavorable outcomes by employing a cluster-based tree-structured Parzen estimator. This strategy allows for a streamlined exploration of architectural possibilities and swift pinpointing of top-performing designs. Through rigorous testing on well-known datasets, our method proves its distinct advantage over existing methods. Compared to leading compression strategies, our approach records an impressive 20% decrease in model size without compromising accuracy. Additionally, our method boasts a 12x reduction in search time relative to the best search-focused strategies currently available. As a result, our proposed method represents a leap forward in neural network design optimization, paving the way for quick model design and implementation in settings with limited resources, thereby propelling the potential of scalable deep learning solutions.

Soheil Nazar Shahsavani, Arash Fayyazi, Mahdi Nazemi et al.

Recent efforts for improving the performance of neural network (NN) accelerators that meet today's application requirements have given rise to a new trend of logic-based NN inference relying on fixed function combinational logic. Mapping such large Boolean functions with many input variables and product terms to digital signal processors (DSPs) on Field-programmable gate arrays (FPGAs) needs a novel framework considering the structure and the reconfigurability of DSP blocks during this process. The proposed methodology in this paper maps the fixed function combinational logic blocks to a set of Boolean functions where Boolean operations corresponding to each function are mapped to DSP devices rather than look-up tables (LUTs) on the FPGAs to take advantage of the high performance, low latency, and parallelism of DSP blocks. % This paper also presents an innovative design and optimization methodology for compilation and mapping of NNs, utilizing fixed function combinational logic to DSPs on FPGAs employing high-level synthesis flow. % Our experimental evaluations across several \REVone{datasets} and selected NNs demonstrate the comparable performance of our framework in terms of the inference latency and output accuracy compared to prior art FPGA-based NN accelerators employing DSPs.

Seyedarmin Azizi, Mahdi Nazemi, Massoud Pedram

As Vision Transformers (ViTs) increasingly set new benchmarks in computer vision, their practical deployment on inference engines is often hindered by their significant memory bandwidth and (on-chip) memory footprint requirements. This paper addresses this memory limitation by introducing an activation-aware model compression methodology that uses selective low-rank weight tensor approximations of different layers to reduce the parameter count of ViTs. The key idea is to decompose the weight tensors into a sum of two parameter-efficient tensors while minimizing the error between the product of the input activations with the original weight tensor and the product of the input activations with the approximate tensor sum. This approximation is further refined by adopting an efficient layer-wise error compensation technique that uses the gradient of the layer's output loss. The combination of these techniques achieves excellent results while it avoids being trapped in a shallow local minimum early in the optimization process and strikes a good balance between the model compression and output accuracy. Notably, the presented method significantly reduces the parameter count of DeiT-B by 60% with less than 1% accuracy drop on the ImageNet dataset, overcoming the usual accuracy degradation seen in low-rank approximations. In addition to this, the presented compression technique can compress large DeiT/ViT models to have about the same model size as smaller DeiT/ViT variants while yielding up to 1.8% accuracy gain. These results highlight the efficacy of our approach, presenting a viable solution for embedding ViTs in memory-constrained environments without compromising their performance.

Seyedarmin Azizi, Mahdi Nazemi, Mehdi Kamal et al.

This paper presents a mixed-computation neural network processing approach for edge applications that incorporates low-precision (low-width) Posit and low-precision fixed point (FixP) number systems. This mixed-computation approach employs 4-bit Posit (Posit4), which has higher precision around zero, for representing weights with high sensitivity, while it uses 4-bit FixP (FixP4) for representing other weights. A heuristic for analyzing the importance and the quantization error of the weights is presented to assign the proper number system to different weights. Additionally, a gradient approximation for Posit representation is introduced to improve the quality of weight updates in the backpropagation process. Due to the high energy consumption of the fully Posit-based computations, neural network operations are carried out in FixP or Posit/FixP. An efficient hardware implementation of a MAC operation with a first Posit operand and FixP for a second operand and accumulator is presented. The efficacy of the proposed low-precision mixed-computation approach is extensively assessed on vision and language models. The results show that, on average, the accuracy of the mixed-computation is about 1.5% higher than that of FixP with a cost of 0.19% energy overhead.

Mahdi Nazemi, Arash Fayyazi, Amirhossein Esmaili et al.

While there is a large body of research on efficient processing of deep neural networks (DNNs), ultra-low-latency realization of these models for applications with stringent, sub-microsecond latency requirements continues to be an unresolved, challenging problem. Field-programmable gate array (FPGA)-based DNN accelerators are gaining traction as a serious contender to replace graphics processing unit/central processing unit-based platforms considering their performance, flexibility, and energy efficiency. This paper presents NullaNet Tiny, an across-the-stack design and optimization framework for constructing resource and energy-efficient, ultra-low-latency FPGA-based neural network accelerators. The key idea is to replace expensive operations required to compute various filter/neuron functions in a DNN with Boolean logic expressions that are mapped to the native look-up tables (LUTs) of the FPGA device (examples of such operations are multiply-and-accumulate and batch normalization). At about the same level of classification accuracy, compared to Xilinx's LogicNets, our design achieves 2.36$\times$ lower latency and 24.42$\times$ lower LUT utilization.

Souvik Kundu, Mahdi Nazemi, Peter A. Beerel et al.

This paper presents a dynamic network rewiring (DNR) method to generate pruned deep neural network (DNN) models that are robust against adversarial attacks yet maintain high accuracy on clean images. In particular, the disclosed DNR method is based on a unified constrained optimization formulation using a hybrid loss function that merges ultra-high model compression with robust adversarial training. This training strategy dynamically adjusts inter-layer connectivity based on per-layer normalized momentum computed from the hybrid loss function. In contrast to existing robust pruning frameworks that require multiple training iterations, the proposed learning strategy achieves an overall target pruning ratio with only a single training iteration and can be tuned to support both irregular and structured channel pruning. To evaluate the merits of DNR, experiments were performed with two widely accepted models, namely VGG16 and ResNet-18, on CIFAR-10, CIFAR-100 as well as with VGG16 on Tiny-ImageNet. Compared to the baseline uncompressed models, DNR provides over20x compression on all the datasets with no significant drop in either clean or adversarial classification accuracy. Moreover, our experiments show that DNR consistently finds compressed models with better clean and adversarial image classification performance than what is achievable through state-of-the-art alternatives.

Mahdi Nazemi, Amirhossein Esmaili, Arash Fayyazi et al.

Machine learning models differ in terms of accuracy, computational/memory complexity, training time, and adaptability among other characteristics. For example, neural networks (NNs) are well-known for their high accuracy due to the quality of their automatic feature extraction while brain-inspired hyperdimensional (HD) learning models are famous for their quick training, computational efficiency, and adaptability. This work presents a hybrid, synergic machine learning model that excels at all the said characteristics and is suitable for incremental, on-line learning on a chip. The proposed model comprises an NN and a classifier. The NN acts as a feature extractor and is specifically trained to work well with the classifier that employs the HD computing framework. This work also presents a parameterized hardware implementation of the said feature extraction and classification components while introducing a compiler that maps any arbitrary NN and/or classifier to the aforementioned hardware. The proposed hybrid machine learning model has the same level of accuracy (i.e. $\pm$1%) as NNs while achieving at least 10% improvement in accuracy compared to HD learning models. Additionally, the end-to-end hardware realization of the hybrid model improves power efficiency by 1.60x compared to state-of-the-art, high-performance HD learning implementations while improving latency by 2.13x. These results have profound implications for the application of such synergic models in challenging cognitive tasks.

Souvik Kundu, Mahdi Nazemi, Massoud Pedram et al.

The high energy cost of processing deep convolutional neural networks impedes their ubiquitous deployment in energy-constrained platforms such as embedded systems and IoT devices. This work introduces convolutional layers with pre-defined sparse 2D kernels that have support sets that repeat periodically within and across filters. Due to the efficient storage of our periodic sparse kernels, the parameter savings can translate into considerable improvements in energy efficiency due to reduced DRAM accesses, thus promising significant improvements in the trade-off between energy consumption and accuracy for both training and inference. To evaluate this approach, we performed experiments with two widely accepted datasets, CIFAR-10 and Tiny ImageNet in sparse variants of the ResNet18 and VGG16 architectures. Compared to baseline models, our proposed sparse variants require up to 82% fewer model parameters with 5.6times fewer FLOPs with negligible loss in accuracy for ResNet18 on CIFAR-10. For VGG16 trained on Tiny ImageNet, our approach requires 5.8times fewer FLOPs and up to 83.3% fewer model parameters with a drop in top-5 (top-1) accuracy of only 1.2% (2.1%). We also compared the performance of our proposed architectures with that of ShuffleNet andMobileNetV2. Using similar hyperparameters and FLOPs, our ResNet18 variants yield an average accuracy improvement of 2.8%.

Mahdi Nazemi, Ghasem Pasandi, Massoud Pedram

Deep neural networks have been successfully deployed in a wide variety of applications including computer vision and speech recognition. However, computational and storage complexity of these models has forced the majority of computations to be performed on high-end computing platforms or on the cloud. To cope with computational and storage complexity of these models, this paper presents a training method that enables a radically different approach for realization of deep neural networks through Boolean logic minimization. The aforementioned realization completely removes the energy-hungry step of accessing memory for obtaining model parameters, consumes about two orders of magnitude fewer computing resources compared to realizations that use floatingpoint operations, and has a substantially lower latency.

Mahdi Nazemi, Massoud Pedram

Major advancements in building general-purpose and customized hardware have been one of the key enablers of versatility and pervasiveness of machine learning models such as deep neural networks. To sustain this ubiquitous deployment of machine learning models and cope with their computational and storage complexity, several solutions such as low-precision representation of model parameters using fixed-point representation and deploying approximate arithmetic operations have been employed. Studying the potency of such solutions in different applications requires integrating them into existing machine learning frameworks for high-level simulations as well as implementing them in hardware to analyze their effects on power/energy dissipation, throughput, and chip area. Lop is a library for design space exploration that bridges the gap between machine learning and efficient hardware realization. It comprises a Python module, which can be integrated with some of the existing machine learning frameworks and implements various customizable data representations including fixed-point and floating-point as well as approximate arithmetic operations.Furthermore, it includes a highly-parameterized Scala module, which allows synthesizing hardware based on the said data representations and arithmetic operations. Lop allows researchers and designers to quickly compare quality of their models using various data representations and arithmetic operations in Python and contrast the hardware cost of viable representations by synthesizing them on their target platforms (e.g., FPGA or ASIC). To the best of our knowledge, Lop is the first library that allows both software simulation and hardware realization using customized data representations and approximate computing techniques.

Mahdi Nazemi, Amir Erfan Eshratifar, Massoud Pedram

With ever-increasing application of machine learning models in various domains such as image classification, speech recognition and synthesis, and health care, designing efficient hardware for these models has gained a lot of popularity. While the majority of researches in this area focus on efficient deployment of machine learning models (a.k.a inference), this work concentrates on challenges of training these models in hardware. In particular, this paper presents a high-performance, scalable, reconfigurable solution for both training and deployment of different dimensionality reduction models in hardware by introducing a hardware-friendly algorithm. Compared to state-of-the-art implementations, our proposed algorithm and its hardware realization decrease resource consumption by 50\% without any degradation in accuracy.

Sheng Lin, Ning Liu, Mahdi Nazemi et al.

Deep learning has delivered its powerfulness in many application domains, especially in image and speech recognition. As the backbone of deep learning, deep neural networks (DNNs) consist of multiple layers of various types with hundreds to thousands of neurons. Embedded platforms are now becoming essential for deep learning deployment due to their portability, versatility, and energy efficiency. The large model size of DNNs, while providing excellent accuracy, also burdens the embedded platforms with intensive computation and storage. Researchers have investigated on reducing DNN model size with negligible accuracy loss. This work proposes a Fast Fourier Transform (FFT)-based DNN training and inference model suitable for embedded platforms with reduced asymptotic complexity of both computation and storage, making our approach distinguished from existing approaches. We develop the training and inference algorithms based on FFT as the computing kernel and deploy the FFT-based inference model on embedded platforms achieving extraordinary processing speed.

Mahdi Nazemi, Shahin Nazarian, Massoud Pedram

Independent Component Analysis (ICA) is a dimensionality reduction technique that can boost efficiency of machine learning models that deal with probability density functions, e.g. Bayesian neural networks. Algorithms that implement adaptive ICA converge slower than their nonadaptive counterparts, however, they are capable of tracking changes in underlying distributions of input features. This intrinsically slow convergence of adaptive methods combined with existing hardware implementations that operate at very low clock frequencies necessitate fundamental improvements in both algorithm and hardware design. This paper presents an algorithm that allows efficient hardware implementation of ICA. Compared to previous work, our FPGA implementation of adaptive ICA improves clock frequency by at least one order of magnitude and throughput by at least two orders of magnitude. Our proposed algorithm is not limited to ICA and can be used in various machine learning problems that use stochastic gradient descent optimization.