Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching

Large language models (LLMs) have emerged due to their capability to generate high-quality content across diverse contexts. To reduce their explosively increasing demands for computing resources, a mixture of experts (MoE) has emerged. The MoE layer enables exploiting a huge number of parameters with less computation. Applying state-of-the-art continuous batching increases throughput; however, it leads to frequent DRAM access in the MoE and attention layers. We observe that conventional computing devices have limitations when processing the MoE and attention layers, which dominate the total execution time and exhibit low arithmetic intensity (Op/B). Processing MoE layers only with devices targeting low-Op/B such as processing-in-memory (PIM) architectures is challenging due to the fluctuating Op/B in the MoE layer caused by continuous batching. To address these challenges, we propose Duplex, which comprises xPU tailored for high-Op/B and Logic-PIM to effectively perform low-Op/B operation within a single device. Duplex selects the most suitable processor based on the Op/B of each layer within LLMs. As the Op/B of the MoE layer is at least 1 and that of the attention layer has a value of 4-8 for grouped query attention, prior PIM architectures are not efficient, which place processing units inside DRAM dies and only target extremely low-Op/B (under one) operations. Based on recent trends, Logic-PIM adds more through-silicon vias (TSVs) to enable high-bandwidth communication between the DRAM die and the logic die and place powerful processing units on the logic die, which is best suited for handling low-Op/B operations ranging from few to a few dozens. To maximally utilize the xPU and Logic-PIM, we propose expert and attention co-processing.

NeuJeans: Private Neural Network Inference with Joint Optimization of Convolution and Bootstrapping

Fully homomorphic encryption (FHE) is a promising cryptographic primitive for realizing private neural network inference (PI) services by allowing a client to fully offload the inference task to a cloud server while keeping the client data oblivious to the server. This work proposes NeuJeans, an FHE-based solution for the PI of deep convolutional neural networks (CNNs). NeuJeans tackles the critical problem of the enormous computational cost for the FHE evaluation of convolutional layers (conv2d), mainly due to the high cost of data reordering and bootstrapping. We first propose an encoding method introducing nested structures inside encoded vectors for FHE, which enables us to develop efficient conv2d algorithms with reduced data reordering costs. However, the new encoding method also introduces additional computations for conversion between encoding methods, which could negate its advantages. We discover that fusing conv2d with bootstrapping eliminates such computations while reducing the cost of bootstrapping. Then, we devise optimized execution flows for various types of conv2d and apply them to end-to-end implementation of CNNs. NeuJeans accelerates the performance of conv2d by up to 5.68 times compared to state-of-the-art FHE-based PI work and performs the PI of a CNN at the scale of ImageNet (ResNet18) within a mere few seconds

HyPHEN: A Hybrid Packing Method and Optimizations for Homomorphic Encryption-Based Neural Networks

Convolutional neural network (CNN) inference using fully homomorphic encryption (FHE) is a promising private inference (PI) solution due to the capability of FHE that enables offloading the whole computation process to the server while protecting the privacy of sensitive user data. However, prior FHEbased CNN (HCNN) implementations are far from being practical due to the high computational and memory overheads of FHE. To overcome this limitation, we present HyPHEN, a deep HCNN construction that features an efficient FHE convolution algorithm, data packing methods (hybrid packing and image slicing), and FHE-specific optimizations. Such enhancements enable HyPHEN to substantially reduce the memory footprint and the number of expensive homomorphic operations, such as ciphertext rotation and bootstrapping. As a result, HyPHEN brings the latency of HCNN CIFAR-10 inference down to a practical level at 1.40s (ResNet20) and demonstrates HCNN ImageNet inference for the first time at 16.87s (ResNet18).

AESPA: Accuracy Preserving Low-degree Polynomial Activation for Fast Private Inference

Hybrid private inference (PI) protocol, which synergistically utilizes both multi-party computation (MPC) and homomorphic encryption, is one of the most prominent techniques for PI. However, even the state-of-the-art PI protocols are bottlenecked by the non-linear layers, especially the activation functions. Although a standard non-linear activation function can generate higher model accuracy, it must be processed via a costly garbled-circuit MPC primitive. A polynomial activation can be processed via Beaver's multiplication triples MPC primitive but has been incurring severe accuracy drops so far. In this paper, we propose an accuracy preserving low-degree polynomial activation function (AESPA) that exploits the Hermite expansion of the ReLU and basis-wise normalization. We apply AESPA to popular ML models, such as VGGNet, ResNet, and pre-activation ResNet, to show an inference accuracy comparable to those of the standard models with ReLU activation, achieving superior accuracy over prior low-degree polynomial studies. When applied to the all-RELU baseline on the state-of-the-art Delphi PI protocol, AESPA shows up to 42.1x and 28.3x lower online latency and communication cost.

Restructuring Batch Normalization to Accelerate CNN Training

Because CNN models are compute-intensive, where billions of operations can be required just for an inference over a single input image, a variety of CNN accelerators have been proposed and developed. For the early CNN models, the research mostly focused on convolutional and fully-connected layers because the two layers consumed most of the computation cycles. For more recent CNN models, however, non-convolutional layers have become comparably important because of the popular use of newly designed non-convolutional layers and because of the reduction in the number and size of convolutional filters. Non-convolutional layers, including batch normalization (BN), typically have relatively lower computational intensity compared to the convolutional or fully-connected layers, and hence are often constrained by main-memory bandwidth. In this paper, we focus on accelerating the BN layers among the non-convolutional layers, as BN has become a core design block of modern CNNs. A typical modern CNN has a large number of BN layers. BN requires mean and variance calculations over each mini-batch during training. Therefore, the existing memory-access reduction techniques, such as fusing multiple CONV layers, are not effective for accelerating BN due to their inability to optimize mini-batch related calculations. To address this increasingly important problem, we propose to restructure BN layers by first splitting it into two sub-layers and then combining the first sub-layer with its preceding convolutional layer and the second sub-layer with the following activation and convolutional layers. The proposed solution can significantly reduce main-memory accesses while training the latest CNN models, and the experiments on a chip multiprocessor with our modified Caffe implementation show that the proposed BN restructuring can improve the performance of DenseNet with 121 convolutional layers by 28.4%.