Search

The advent of Transformers has revolutionized computer vision, offering a powerful alternative to convolutional neural networks (CNNs), especially with the local attention mechanism that excels at capturing local structures within the input and achieve state-of-the-art performance. Processing in-memory (PIM) architecture offers extensive parallelism, low data movement costs, and scalable memory bandwidth, making it a promising solution to accelerate Transformer with memory-intensive operations. However, the crucial challenge lies in efficiently deploying the entire model onto a resource-limited PIM system while parallelizing each transformer block with potentially many computational branches based on local attention mechanisms. We present Allspark, which focuses on workload orchestration for visual Transformers on PIM systems, aiming at minimizing inference latency. Firstly, to fully utilize the massive parallelism of PIM, Allspark empolys a finer-grained partitioning scheme for computational branches, and format a systematic layout and interleaved dataflow with maximized data locality and reduced data movement. Secondly, Allspark formulates the scheduling of the complete model on a resource-limited distributed PIM system as an integer linear programming (ILP) problem. Thirdly, as local-global data interactions exhibit complex yet regular dependencies, Allspark provides a greedy-based mapping method to allocate computational branches onto the PIM system and minimize NoC communication costs. Extensive experiments on 3D-stacked DRAM-based PIM systems show that Allspark brings 1.2x-24.0x inference speedup for various visual Transformers over baselines, and that Allspark-enriched PIM system yields average speedups of 2.3x and energy savings of 20x-55x over Nvidia V100 GPU., Comment: The article is currently under review by IEEE Transactions on Computers, and has been submitted to HPCA'2024 and ISCA'2024

Quantum operations by utilizing the underlying geometric phases produced in physical systems are favoured due to its potential robustness. When a system in a non-degenerate eigenstate undergoes an adiabatically cyclic evolution dominated by its Hamiltonian, it will get a geometric phase, referred to as the Berry Phase. While a non-adiabatically cyclic evolution produces an Aharonov-Anandan geometric phase. The two types of Abelian geometric phases are extended to the non-Abelian cases, where the phase factors become matrix-valued and the transformations associated with different loops are non-commutable. Abelian and non-Abelian (holonomic) operations are prevalent in discrete variable systems, whose limited (say, two) energy levels, form the qubit. While their developments in continuous systems have also been investigated, mainly due to that, bosonic modes (in, such as, cat states) with large Hilbert spaces, provide potential advantages in fault-tolerant quantum computation. Here we propose a feasible scheme to realize non-adiabatic holonomic quantum logic operations in continuous variable systems with cat codes. We construct arbitrary single-qubit (two-qubit) gates with the combination of single- and two-photon drivings applied to a Kerr Parametric Oscillator (KPO) (the coupled KPOs). Our scheme relaxes the requirements of the previously proposed quantum geometric operation strategies in continuous variable systems, providing an effective way for quantum control.

In recent years, analog circuits have received extensive attention and are widely used in many emerging applications. The high demand for analog circuits necessitates shorter circuit design cycles. To achieve the desired performance and specifications, various geometrical symmetry constraints must be carefully considered during the analog layout process. However, the manual labeling of these constraints by experienced analog engineers is a laborious and time-consuming process. To handle the costly runtime issue, we propose a graph-based learning framework to automatically extract symmetric constraints in analog circuit layout. The proposed framework leverages the connection characteristics of circuits and the devices' information to learn the general rules of symmetric constraints, which effectively facilitates the extraction of device-level constraints on circuit netlists. The experimental results demonstrate that compared to state-of-the-art symmetric constraint detection approaches, our framework achieves higher accuracy and F1-score., Comment: 11 pages,10 figures, 6 tables, 1 algorithm

Sparse matrix multiplication (SpMM) is widely applied to numerous domains, such as graph processing, machine learning, and data analytics. However, inner product based SpMM induces redundant zero-element computing for mismatched nonzero operands, while outer product based approach lacks input reuse across Process Elements (PEs) and poor output locality for accumulating partial sum (psum) matrices. Besides, current works only focus on sparse-sparse matrix multiplication (SSMM) or sparse-dense matrix multiplication (SDMM), rarely performing efficiently for both. To address these problems, this paper proposes an unified SpMM accelerator, called IOPS, hybridizing inner with outer products. It reuses the input matrix among PEs with inner product dataflow, and removes zero-element calculations with outer product approach in each PE, which can efficiently process SSMM and SDMM. Moreover, an address mapping method is designed to accumulate the irregular sparse psum matrices, reducing the latency and DRAM access of psum accumulating. Furthermore, an adaptive partition strategy is proposed to tile the input matrices based on their sparsity ratios, effectively utilizing the storage of architecture and reducing DRAM access. Compared with the SSMM accelerator, SpArch, we achieve 1.7x~6.3x energy efficiency and 1.2x~4.4x resource efficiency, with 1.4x~2.1x DRAM access saving.

Analog in-memory computing (AiMC) is an emerging technology that shows fantastic performance superiority for neural network acceleration. However, as the computational bit-width and scale increase, high-precision data conversion and long-distance data routing will result in unacceptable energy and latency overheads in the AiMC system. In this work, we focus on the potential of in-charge computing and in-time interconnection and show an innovative AiMC architecture, named AiDAC, with three key contributions: (1) AiDAC enhances multibit computing efficiency and reduces data conversion times by grouping capacitors technology; (2) AiDAC first adopts row drivers and column time accumulators to achieve large-scale AiMC arrays integration while minimizing the energy cost of data movements. (3) AiDAC is the first work to support large-scale all-analog multibit vector-matrix multiplication (VMM) operations. The evaluation shows that AiDAC maintains high-precision calculation (less than 0.79% total computing error) while also possessing excellent performance features, such as high parallelism (up to 26.2TOPS), low latency (<20ns/VMM), and high energy efficiency (123.8TOPS/W), for 8bits VMM with 1024 input channels., Comment: 6 pages, 8 figures

This paper proposes an application mapping algorithm, BandMap, for coarse-grained reconfigurable array (CGRA), which allocates the bandwidth in PE array according to the transferring demands of data, especially the data with high spatial reuse, to reduce the routing PEs. To cover bandwidth allocation, BandMap maps the data flow graphs (DFGs), abstracted from applications, by solving the maximum independent set (MIS) on a mixture of tuple and quadruple resource occupation conflict graph. Compared to a state-of-art BusMap work, Bandmap can achieve reduced routing PEs with the same or even smaller initiation interval (II).

With the widespread use of deep neural networks(DNNs) in intelligent systems, DNN accelerators with high performance and energy efficiency are greatly demanded. As one of the feasible processing-in-memory(PIM) architectures, 3D-stacked-DRAM-based PIM(DRAM-PIM) architecture enables large-capacity memory and low-cost memory access, which is a promising solution for DNN accelerators with better performance and energy efficiency. However, the low-cost characteristics of stacked DRAM and the distributed manner of memory access and data storing require us to rebalance the hardware design and DNN mapping. In this paper, we propose NicePIM to efficiently explore the design space of hardware architecture and DNN mapping of DRAM-PIM accelerators, which consists of three key components: PIM-Tuner, PIM-Mapper and Data-Scheduler. PIM-Tuner optimizes the hardware configurations leveraging a DNN model for classifying area-compliant architectures and a deep kernel learning model for identifying better hardware parameters. PIM-Mapper explores a variety of DNN mapping configurations, including parallelism between branches of DNN, DNN layer partitioning, DRAM capacity allocation and data layout pattern in DRAM to generate high-hardware-utilization DNN mapping schemes for various hardware configurations. The Data-Scheduler employs an integer-linear-programming-based data scheduling algorithm to alleviate the inter-PIM-node communication overhead of data-sharing brought by DNN layer partitioning. Experimental results demonstrate that NicePIM can optimize hardware configurations for DRAM-PIM systems effectively and can generate high-quality DNN mapping schemes with latency and energy cost reduced by 37% and 28% on average respectively compared to the baseline method., Comment: 13 pages

Bit-serial architectures can handle Neural Networks (NNs) with different weight precisions, achieving higher resource efficiency compared with bit-parallel architectures. Besides, the weights contain abundant zero bits owing to the fault tolerance of NNs, indicating that bit sparsity of NNs can be further exploited for performance improvement. However, the irregular proportion of zero bits in each weight causes imbalanced workloads in the Processing Element (PE) array, which degrades performance or induces overhead for sparse processing. Thus, this paper proposed a bit-sparsity quantization method to maintain the bit sparsity ratio of each weight to no more than a certain value for balancing workloads, with little accuracy loss. Then, we co-designed a sparse bit-serial architecture, called Bit-balance, to improve overall performance, supporting weight-bit sparsity and adaptive bitwidth computation. The whole design was implemented with 65nm technology at 1 GHz and performs at 326-, 30-, 56-, and 218-frame/s for AlexNet, VGG-16, ResNet-50, and GoogleNet respectively. Compared with sparse bit-serial accelerator, Bitlet, Bit-balance achieves 1.8x~2.7x energy efficiency (frame/J) and 2.1x~3.7x resource efficiency (frame/mm2)., Comment: 10 pages 17 figures

Modern field programmable gate array(FPGA) can be partially dynamically reconfigurable with heterogeneous resources distributed on the chip. And FPGA-based partially dynamically reconfigurable system(FPGA-PDRS) can be used to accelerate computing and improve computing flexibility. However, the traditional design of FPGA-PDRS is based on manual design. Implementing the automation of FPGA-PDRS needs to solve the problems of task modules partitioning, scheduling, and floorplanning on heterogeneous resources. Existing works only partly solve problems for the automation process of FPGA-PDRS or model homogeneous resource for FPGA-PDRS. To better solve the problems in the automation process of FPGA-PDRS and narrow the gap between algorithm and application, in this paper, we propose a complete workflow including three parts, pre-processing to generate the list of task modules candidate shapes according to the resources requirements, exploration process to search the solution of task modules partitioning, scheduling, and floorplanning, and post-optimization to improve the success rate of floorplan. Experimental results show that, compared with state-of-the-art work, the proposed complete workflow can improve performance by 18.7\%, reduce communication cost by 8.6\%, on average, with improving the resources reuse rate of the heterogeneous resources on the chip. And based on the solution generated by the exploration process, the post-optimization can improve the success rate of the floorplan by 14\%.

Sparsity is an intrinsic property of convolutional neural network(CNN) and worth exploiting for CNN accelerators, but extra processing comes with hardware overhead, causing many architectures suffering from only minor profit. Meanwhile, systolic array has been increasingly competitive on CNNs acceleration for its high spatiotemporal locality and low hardware overhead. However, the irregularity of sparsity induces imbalanced workload under the rigid systolic dataflow, causing performance degradation. Thus, this paper proposed a systolicarray-based architecture, called Sense, for sparse CNN acceleration by model-hardware co-design, achieving large performance improvement. To balance input feature map(IFM) and weight loads across Processing Element(PE) array, we applied channel clustering to gather IFMs with approximate sparsity for array computation, and co-designed a load-balancing weight pruning method to keep the sparsity ratio of each kernel at a certain value with little accuracy loss, improving PE utilization and overall performance. Additionally, Adaptive Dataflow Configuration is applied to determine the computing strategy based on the storage ratio of IFMs and weights, lowering 1.17x-1.8x DRAM access compared with Swallow and further reducing system energy consumption. The whole design is implemented on ZynqZCU102 with 200MHz and performs at 471-, 34-, 53- and 191-image/s for AlexNet, VGG-16, ResNet-50 and GoogleNet respectively. Compared against sparse systolic-array-based accelerators, Swallow, FESA and SPOTS, Sense achieves 1x-2.25x, 1.95x-2.5x and 1.17x-2.37x performance improvement on these CNNs respectively with reasonable overhead., Comment: 14 pages, 29 figures, 6 tables, IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS