The constant growth of DNNs makes them challenging to implement and run efficiently on traditional compute-centric architectures. Some accelerators have attempted to add more compute units and on-chip buffers to solve the memory wall problem without much success, and sometimes even worsening the issue since more compute units also require higher memory bandwidth. Prior works have proposed the design of memory-centric architectures based on the Near-Data Processing (NDP) paradigm. NDP seeks to break the memory wall by moving the computations closer to the memory hierarchy, reducing the data movements and their cost as much as possible. The 3D-stacked memory is especially appealing for DNN accelerators due to its high-density/low-energy storage and near-memory computation capabilities to perform the DNN operations massively in parallel. However, memory accesses remain as the main bottleneck for running modern DNNs efficiently. To improve the efficiency of DNN inference we present QeiHaN, a hardware accelerator that implements a 3D-stacked memory-centric weight storage scheme to take advantage of a logarithmic quantization of activations. In particular, since activations of FC and CONV layers of modern DNNs are commonly represented as powers of two with negative exponents, QeiHaN performs an implicit in-memory bit-shifting of the DNN weights to reduce memory activity. Only the meaningful bits of the weights required for the bit-shift operation are accessed. Overall, QeiHaN reduces memory accesses by 25\% compared to a standard memory organization. We evaluate QeiHaN on a popular set of DNNs. On average, QeiHaN provides $4.3x$ speedup and $3.5x$ energy savings over a Neurocube-like accelerator. less

Modern GPUs require an enormous register file (RF) to store the context of thousands of active threads. It consumes considerable energy and contains multiple large banks to provide enough throughput. Thus, a RF caching mechanism can significantly improve the performance and energy consumption of the GPUs by avoiding reads from the large banks that consume significant energy and may cause port conflicts. This paper introduces an energy-efficient RF caching mechanism called Malekeh that repurposes an existing component in GPUs' RF to operate as a cache in addition to its original functionality. In this way, Malekeh minimizes the overhead of adding a RF cache to GPUs. Besides, Malekeh leverages an issue scheduling policy that utilizes the reuse distance of the values in the RF cache and is controlled by a dynamic algorithm. The goal is to adapt the issue policy to the runtime program characteristics to maximize the GPU's performance and the hit ratio of the RF cache. The reuse distance is approximated by the compiler using profiling and is used at run time by the proposed caching scheme. We show that Malekeh reduces the number of reads to the RF banks by 46.4% and the dynamic energy of the RF by 28.3%. Besides, it improves performance by 6.1% while adding only 2KB of extra storage per core to the baseline RF of 256KB, which represents a negligible overhead of 0.78%. less

Spiking Neural Networks (SNNs) offer a promising alternative to Artificial Neural Networks (ANNs) for deep learning applications, particularly in resource-constrained systems. This is largely due to their inherent sparsity, influenced by factors such as the input dataset, the length of the spike train, and the network topology. While a few prior works have demonstrated the advantages of incorporating sparsity into the hardware design, especially in terms of reducing energy consumption, the impact on hardware resources has not yet been explored. This is where design space exploration (DSE) becomes crucial, as it allows for the optimization of hardware performance by tailoring both the hardware and model parameters to suit specific application needs. However, DSE can be extremely challenging given the potentially large design space and the interplay of hardware architecture design choices and application-specific model parameters. In this paper, we propose a flexible hardware design that leverages the sparsity of SNNs to identify highly efficient, application-specific accelerator designs. We develop a high-level, cycle-accurate simulation framework for this hardware and demonstrate the framework's benefits in enabling detailed and fine-grained exploration of SNN design choices, such as the layer-wise logical-to-hardware ratio (LHR). Our experimental results show that our design can (i) achieve up to $76\%$ reduction in hardware resources and (ii) deliver a speed increase of up to $31.25\times$, while requiring $27\%$ fewer hardware resources compared to sparsity-oblivious designs. We further showcase the robustness of our framework by varying spike train lengths with different neuron population sizes to find the optimal trade-off points between accuracy and hardware latency. less

Recognizing the explosive increase in the use of DNN-based applications, several industrial companies developed a custom ASIC (e.g., Google TPU, IBM RaPiD, Intel NNP-I/NNP-T) and constructed a hyperscale cloud infrastructure with it. The ASIC performs operations of the inference or training process of DNN models which are requested by users. Since the DNN models have different data formats and types of operations, the ASIC needs to support diverse data formats and generality for the operations. However, the conventional ASICs do not fulfill these requirements. To overcome the limitations of it, we propose a flexible DNN accelerator called All-rounder. The accelerator is designed with an area-efficient multiplier supporting multiple precisions of integer and floating point datatypes. In addition, it constitutes a flexibly fusible and fissionable MAC array to support various types of DNN operations efficiently. We implemented the register transfer level (RTL) design using Verilog and synthesized it in 28nm CMOS technology. To examine practical effectiveness of our proposed designs, we designed two multiply units and three state-of-the-art DNN accelerators. We compare our multiplier with the multiply units and perform architectural evaluation on performance and energy efficiency with eight real-world DNN models. Furthermore, we compare benefits of the All-rounder accelerator to a high-end GPU card, i.e., NVIDIA GeForce RTX30390. The proposed All-rounder accelerator universally has speedup and high energy efficiency in various DNN benchmarks than the baselines. less

To ease the programmability of PIM architectures, we propose DaPPA(data-parallel processing-in-memory architecture), a framework that can, for a given application, automatically distribute input and gather output data, handle memory management, and parallelize work across the DPUs. The key idea behind DaPPA is to remove the responsibility of managing hardware resources from the programmer by providing an intuitive data-parallel pattern-based programming interface that abstracts the hardware components of the UPMEM system. Using this key idea, DaPPA transforms a data-parallel pattern-based application code into the appropriate UPMEM-target code, including the required APIs for data management and code partition, which can then be compiled into a UPMEM-based binary transparently from the programmer. While generating UPMEM-target code, DaPPA implements several code optimizations to improve end-to-end performance. less

We present a low-power, energy efficient 32-bit RISC-V microprocessor unit (MCU) in 22 nm FD-SOI. It achieves ultra-low leakage,even at high temperatures, by using an adaptive reverse body biasing aware sign-off approach, a low-power optimized physical implementation, and custom SRAM macros with retention mode. We demonstrate the robustness of the chip with measurements over the full industrial temperature range, from -40 {\deg}C to 125 {\deg}C. Our results match the state of the art (SOTA) with 4.8 uW / MHz at 50 MHz in active mode and surpass the SOTA in ultra-low-power retention mode. less

In this paper, we present a new analytical 3D placement framework with a bistratal wirelength model for F2F-bonded 3D ICs with heterogeneous technology nodes based on the electrostatic-based density model. The proposed framework, enabled GPU-acceleration, is capable of efficiently determining node partitioning and locations simultaneously, leveraging the dedicated 3D wirelength model and density model. The experimental results on ICCAD 2022 contest benchmarks demonstrate that our proposed 3D placement framework can achieve up to 6.1% wirelength improvement and 4.1% on average compared to the first-place winner with much fewer vertical interconnections and up to 9.8x runtime speedup. Notably, the proposed framework also outperforms the state-of-the-art 3D analytical placer by up to 3.3% wirelength improvement and 2.1% on average with up to 8.8x acceleration on large cases using GPUs. less

Radar-based techniques for detecting vital signs have shown promise for continuous contactless vital sign sensing and healthcare applications. However, real-world indoor environments face significant challenges for existing vital sign monitoring systems. These include signal blockage in non-line-of-sight (NLOS) situations, movement of human subjects, and alterations in location and orientation. Additionally, these existing systems failed to address the challenge of tracking multiple targets simultaneously. To overcome these challenges, we present MEDUSA, a novel coherent ultra-wideband (UWB) based distributed multiple-input multiple-output (MIMO) radar system, especially it allows users to customize and disperse the $16 \times 16$ into sub-arrays. MEDUSA takes advantage of the diversity benefits of distributed yet wirelessly synchronized MIMO arrays to enable robust vital sign monitoring in real-world and daily living environments where human targets are moving and surrounded by obstacles. We've developed a scalable, self-supervised contrastive learning model which integrates seamlessly with our hardware platform. Each attention weight within the model corresponds to a specific antenna pair of Tx and Rx. The model proficiently recovers accurate vital sign waveforms by decomposing and correlating the mixed received signals, including comprising human motion, mobility, noise, and vital signs. Through extensive evaluations involving 21 participants and over 200 hours of collected data (3.75 TB in total, with 1.89 TB for static subjects and 1.86 TB for moving subjects), MEDUSA's performance has been validated, showing an average gain of 20% compared to existing systems employing COTS radar sensors. This demonstrates MEDUSA's spatial diversity gain for real-world vital sign monitoring, encompassing target and environmental dynamics in familiar and unfamiliar indoor environments. less

Address translation is a performance bottleneck in data-intensive workloads due to large datasets and irregular access patterns that lead to frequent high-latency page table walks (PTWs). PTWs can be reduced by using (i) large hardware TLBs or (ii) large software-managed TLBs. Unfortunately, both solutions have significant drawbacks: increased access latency, power and area (for hardware TLBs), and costly memory accesses, the need for large contiguous memory blocks, and complex OS modifications (for software-managed TLBs). We present Victima, a new software-transparent mechanism that drastically increases the translation reach of the processor by leveraging the underutilized resources of the cache hierarchy. The key idea of Victima is to repurpose L2 cache blocks to store clusters of TLB entries, thereby providing an additional low-latency and high-capacity component that backs up the last-level TLB and thus reduces PTWs. Victima has two main components. First, a PTW cost predictor (PTW-CP) identifies costly-to-translate addresses based on the frequency and cost of the PTWs they lead to. Second, a TLB-aware cache replacement policy prioritizes keeping TLB entries in the cache hierarchy by considering (i) the translation pressure (e.g., last-level TLB miss rate) and (ii) the reuse characteristics of the TLB entries. Our evaluation results show that in native (virtualized) execution environments Victima improves average end-to-end application performance by 7.4% (28.7%) over the baseline four-level radix-tree-based page table design and by 6.2% (20.1%) over a state-of-the-art software-managed TLB, across 11 diverse data-intensive workloads. Victima (i) is effective in both native and virtualized environments, (ii) is completely transparent to application and system software, and (iii) incurs very small area and power overheads on a modern high-end CPU. less

Basecalling, an essential step in many genome analysis studies, relies on large Deep Neural Networks (DNNs) to achieve high accuracy. Unfortunately, these DNNs are computationally slow and inefficient, leading to considerable delays and resource constraints in the sequence analysis process. A Computation-In-Memory (CIM) architecture using memristors can significantly accelerate the performance of DNNs. However, inherent device non-idealities and architectural limitations of such designs can greatly degrade the basecalling accuracy, which is critical for accurate genome analysis. To facilitate the adoption of memristor-based CIM designs for basecalling, it is important to (1) conduct a comprehensive analysis of potential CIM architectures and (2) develop effective strategies for mitigating the possible adverse effects of inherent device non-idealities and architectural limitations. This paper proposes Swordfish, a novel hardware/software co-design framework that can effectively address the two aforementioned issues. Swordfish incorporates seven circuit and device restrictions or non-idealities from characterized real memristor-based chips. Swordfish leverages various hardware/software co-design solutions to mitigate the basecalling accuracy loss due to such non-idealities. To demonstrate the effectiveness of Swordfish, we take Bonito, the state-of-the-art (i.e., accurate and fast), open-source basecaller as a case study. Our experimental results using Sword-fish show that a CIM architecture can realistically accelerate Bonito for a wide range of real datasets by an average of 25.7x, with an accuracy loss of 6.01%. less