Trusted Executing Environments (TEEs) in hardware designs protect program execution from other untrusted software programs in the processor as well as untrusted off-chip hardware components. Meanwhile, Near-Memory Processing (NMP) has shown performance and energy benefits on memory-intensive workloads. Recently, novel memory encryption schemes have been proposed to allow TEEs to leverage the benefits of NMP without requiring trust in the NMP components. In this paper, we present a system design of confidential computing with NMP that can be directly used in Intel SGX, a TEE platform available in commercial processors today. We develop the full software stack and evaluate the results on commercial processors with the emulated AxDIMM, an FPGA-based NMP platform. In our case study on personalized Deep Learning Recommendation Model (DLRM) inference, the proposed confidential computing in NMP achieves up to 1.51× latency reduction and up to 2.57× throughput improvement.

Mixture-of-Experts (MoE) models have recently gained steam in achieving the state-of-the-art performance in a wide range of tasks in computer vision and natural language processing. They effectively expand the model capacity while incurring a minimal increase in computation cost during training. However, deploying such models for inference is difficult due to their large model size and complex communication pattern. In this work, we provide a characterization of two MoE workloads, namely Language Modeling (LM) and Machine Translation (MT) and identify their sources of inefficiencies at deployment. We propose three optimization techniques to mitigate sources of inefficiencies, namely (1) Dynamic gating, (2) Expert Buffering, and (3) Expert load balancing. We show that dynamic gating improves maximum throughput by 6.21-11.55$\times$ for LM, 5.75-10.98$\times$ for MT Encoder and 2.58-5.71$\times$ for MT Decoder. It also reduces memory usage by up to 1.36$\times$ for LM and up to 1.1$\times$ for MT. We further propose Expert Buffering, a new caching mechanism that only keeps hot, active experts in GPU memory while buffering the rest in CPU memory. This reduces static memory allocation by 1.47$\times$. Finally, we propose a load balancing methodology that provides additional robustness to the workload. Our code is available at https://github.com/hyhuang00/moe_inference.

Personalized recommendation is an important class of deep-learning applications that powers a large collection of internet services and consumes a considerable amount of datacenter resources. As the scale of production-grade recommendation systems continues to grow, optimizing their serving performance and efficiency in a heterogeneous datacenter is important and can translate into infrastructure capacity saving. In this paper, we propose Hercules, an optimized framework for personalized recommendation inference serving that targets diverse industry-representative models and cloud-scale heterogeneous systems. Hercules performs a two-stage optimization procedure — offline profiling and online serving. The first stage searches the large under-explored task scheduling space with a gradient-based search algorithm achieving up to 9.0× latency-bounded throughput improvement on individual servers; it also identifies the optimal heterogeneous server architecture for each recommendation workload. The second stage performs heterogeneity-aware cluster provisioning to optimize resource mapping and allocation in response to fluctuating diurnal loads. The proposed cluster scheduler in Hercules achieves 47.7% cluster capacity saving and reduces the provisioned power by 23.7% over a state-of-the-art greedy scheduler.

Today’s data-intensive applications increasingly suffer from significant performance bottlenecks due to the limited memory bandwidth of the classical von Neumann architecture. Near-Data Processing (NDP) has been proposed to perform computation near memory or data storage to reduce data movement for improving performance and energy consumption. However, the untrusted NDP processing units (PUs) bring in new threats to workloads that are private and sensitive, such as private database queries and private machine learning inferences. Meanwhile, most existing secure hardware designs do not consider off-chip components trustworthy. Once data leaving the processor, they must be protected, e.g., via block cipher encryption. Unfortunately, current encryption schemes do not support computation over encrypted data stored in memory or storage, hindering the adoption of NDP techniques for sensitive workloads.In this paper, we propose SecNDP, a lightweight encryption and verification scheme for untrusted NDP devices to perform computation over ciphertext and verify the correctness of linear operations. Our encryption scheme leverages arithmetic secret sharing in secure Multi-Party Computation (MPC) to support operations over ciphertext, and uses counter-mode encryption to reduce the decryption latency. The security of the encryption and verification algorithm is formally proven. Compared with a non-NDP baseline, secure computation with SecNDP significantly reduces the memory bandwidth usage while providing security guarantees. We evaluate SecNDP for two workloads of distinct memory access patterns. In the setting of eight NDP units, we show a speedup up to 7.46× and energy savings of 18% over an unprotected non-NDP baseline, approaching the performance gain attained by native NDP without protection. Furthermore, SecNDP does not require any security assumption on NDP to hold, thus, using the same threat model as existing secure processors. SecNDP can be implemented without changing the NDP protocols and their inherent hardware design.

Near-memory processing has been extensively studied to optimize memory intensive workloads. However, none of the proposed designs address sparse matrix transposition, an important building block in sparse linear algebra applications. Prior work shows that sparse matrix transposition does not scale as well as other sparse primitives such as sparse matrix vector multiplication (SpMV) and hence has become a growing bottleneck in common applications. Sparse matrix transposition is highly memory intensive but low in computational intensity, making it a promising candidate for near-memory processing. In this work, we propose MeNDA, a scalable near-DRAM multi-way merge accelerator that eliminates the off-chip memory interface bottleneck and exposes the high internal memory bandwidth to improve performance and reduce energy consumption for sparse matrix transposition. MeNDA adopts a merge sort based algorithm, exploiting spatial locality, and proposes a near-memory processing unit (PU) featuring a high-performance hardware merge tree. Because of the wide application of merge sort in sparse linear algebra, MeNDA is an extensible solution that can be easily adapted to support other sparse primitives such as SpMV. Techniques including seamless back-to-back merge sort, stall reducing prefetching and request coalescing are further explored to take full advantage of the increased system memory bandwidth. Compared to two state-of-the-art implementations of sparse matrix transposition on a CPU and a sparse library on a GPU, MeNDA is able to achieve a speedup of 19.1X, 12.0X, and 7.7x, respectively. MeNDA also shows an efficiency gain of 3.8x over a recent SpMV accelerator integrated with HBM. Incurring a power consumption of only 78.6 mW, a MeNDA PU can be easily accommodated by commodity DIMMs.

Recent work has demonstrated great potentials of neural network-inspired analog-to-digital converters (NNADCs) in many emerging applications. These NNADCs often rely on resistive random-access memory (RRAM) devices to realize basic NN operations, and usually need high-precision RRAM (6-12 b) to achieve moderate quantization resolutions (4-8 b). Such an optimistic assumption of RRAM precision, however, is not well supported by practical RRAM arrays in the large-scale production process. In this article, we evaluate two new designs of NNADC with low-precision RRAM devices. They take advantage of traditional two-stage/pipelined hardware architecture and a custom deep-learning-based building block design methodology. Results obtained from SPICE simulations demonstrate a robust design of an 8-b subranging NNADC using 4-b RRAM devices, as well as a 14-b pipelined NNADC using 3-b RRAM devices. The evaluations on the two NNADCs suggest that pipelined architecture is better to achieve higher-resolution using lower precision RRAM. We also perform design space exploration on the building blocks of NNADCs to achieve a balanced performance tradeoff. Comprehensive comparisons reveal improved power, speed performance, and competitive figure of merits (FoMs) of the pipelined NNADC, compared with state-of-the-art NNADCs and traditional ADCs. In addition, the proposed pipelined NNADC can support reconfigurable high-resolution nonlinear quantization with high conversion speed and low conversion energy, enabling intelligent analog-to-information interfaces for near-sensor processing.

Personalized recommendation systems leverage deep learning models and account for the majority of data center AI cycles. Their performance is dominated by memory-bound sparse embedding operations with unique irregular memory access patterns that pose a fundamental challenge to accelerate. This paper proposes a lightweight, commodity DRAM compliant, near-memory processing solution to accelerate personalized recommendation inference. The in-depth characterization of production-grade recommendation models shows that embedding operations with high model-, operator and data-level parallelism lead to memory bandwidth saturation, limiting recommendation inference performance. We propose RecNMP which provides a scalable solution to improve system throughput, supporting a broad range of sparse embedding models. RecNMP is specifically tailored to production environments with heavy co-location of operators on a single server. Several hardware/software cooptimization techniques such as memory-side caching, tableaware packet scheduling, and hot entry profiling are studied, providing up to 9.8× memory latency speedup over a highly-optimized baseline. Overall, RecNMP offers 4.2× throughput improvement and 45.8% memory energy savings.

To reduce power-hungry floating point operations and memory accesses in deep neural networks, quantized neural networks are proposed that replace floating point multiplications with simplified reduced-precision operations. To compensate for the accuracy loss due to the high degree of quantization, wider neural network layers with three or more times as many feature maps are employed. One by-product from these inflated layers is increased redundancy in the network. To further improve computational efficiency and leverage this inherent redundancy, we propose a joint optimization approach that simultaneously explores hardware-oriented training and efficient accelerator implementation of binary neural networks (BNN) in FPGAs. More specifically, our SparseBNN method consists of two parts. First, SparseBNN-SW is a training algorithm developed to enhance the structured sparsity of BNNs by 1) training for zero-valued ternary weights instead of binary that are more amenable to pruning and 2) regulating the sparsity for more efficient hardware deployment. Next, we present SparseBNN-HW, an accelerator architecture designed to directly execute the inference on the sparse-encoded format to save both memory access and computations. Experimental results on various representative datasets demonstrate that SparseBNN improves the power efficiency (GOPS/Watt) and resource efficiency (GOPS/kLUT) over the baseline BNN FPGA implementation by 1.70X and 2.22X.

Recent works propose neural network- (NN-) inspired analog-to-digital converters (NNADCs) and demonstrate their great potentials in many emerging applications. These NNADCs often rely on resistive random-access memory (RRAM) devices to realize the NN operations and require high-precision RRAM cells (6~12-bit) to achieve a moderate quantization resolution (4~8-bit). Such optimistic assumption of RRAM resolution, however, is not supported by fabrication data of RRAM arrays in large-scale production process. In this paper, we propose an NN-inspired super-resolution ADC based on low-precision RRAM devices by taking the advantage of a co-design methodology that combines a pipelined hardware architecture with a custom NN training framework. Results obtained from SPICE simulations demonstrate that our method leads to robust design of a 14-bit super-resolution ADC using 3-bit RRAM devices with improved power and speed performance and competitive figure-of-merits (FoMs). In addition to the linear uniform quantization, the proposed ADC can also support configurable high-resolution nonlinear quantization with high conversion speed and low conversion energy, enabling future intelligent analog-to-information interfaces for near-sensor analytics and processing.

The intrinsic error tolerance of neural network (NN) makes approximate computing a promising technique to improve the energy efficiency of NN inference. Conventional approximate computing focuses on balancing the efficiency-accuracy trade-off for existing pre-trained networks, which can lead to suboptimal solutions. In this paper, we propose AxTrain, a hardware-oriented training framework to facilitate approximate computing for NN inference. Specifically, AxTrain leverages the synergy between two orthogonal methods---one actively searches for a network parameters distribution with high error tolerance, and the other passively learns resilient weights by numerically incorporating the noise distributions of the approximate hardware in the forward pass during the training phase. Experimental results from various datasets with near-threshold computing and approximation multiplication strategies demonstrate AxTrain's ability to obtain resilient neural network parameters and system energy efficiency improvement.

Deep neural network (DNN) has achieved spectacular success in recent years. In response to DNN's enormous computation demand and memory footprint, numerous inference accelerators have been proposed. However, the diverse nature of DNNs, both at the algorithm level and the parallelization level, makes it hard to arrive at an "one-size-fits-all" hardware design. In this paper, we develop NNest, an early-stage design space exploration tool that can speedily and accurately estimate the area/performance/energy of DNN inference accelerators based on high-level network topology and architecture traits, without the need for low-level RTL codes. Equipped with a generalized spatial architecture framework, NNest is able to perform fast high-dimensional design space exploration across a wide spectrum of architectural/micro-architectural parameters. Our proposed novel date movement strategies and multi-layer fitting schemes allow NNest to more effectively exploit parallelism inherent in DNN. Results generated by NNest demonstrate: 1) previously-undiscovered accelerator design points that can outperform state-of-the-art implementation by 39.3% in energy efficiency; 2) Pareto frontier curves that comprehensively and quantitatively reveal the multi-objective tradeoffs in custom DNN accelerators; 3) holistic design exploration of different level of quantization techniques including recently-proposed binary neural network (BNN).