Abstract Realizing increasingly complex artificial intelligence (AI) functionalities directly on edge devices calls for unprecedented energy efficiency of edge hardware. Compute-in-memory (CIM) based on resistive random-access memory (RRAM) 1 promises to meet such demand by storing AI model weights in dense, analogue and non-volatile RRAM devices, and by performing AI computation directly within RRAM, thus eliminating power-hungry data movement between separate compute and memory 2–5 . Although recent studies have demonstrated in-memory matrix-vector multiplication on fully integrated RRAM-CIM hardware 6–17 , it remains a goal for a RRAM-CIM chip to simultaneously deliver high energy efficiency, versatility to support diverse models and software-comparable accuracy. Although efficiency, versatility and accuracy are all indispensable for broad adoption of the technology, the inter-related trade-offs among them cannot be addressed by isolated improvements on any single abstraction level of the design. Here, by co-optimizing across all hierarchies of the design from algorithms and architecture to circuits and devices, we present NeuRRAM—a RRAM-based CIM chip that simultaneously delivers versatility in reconfiguring CIM cores for diverse model architectures, energy efficiency that is two-times better than previous state-of-the-art RRAM-CIM chips across various computational bit-precisions, and inference accuracy comparable to software models quantized to four-bit weights across various AI tasks, including accuracy of 99.0 percent on MNIST 18 and 85.7 percent on CIFAR-10 19 image classification, 84.7-percent accuracy on Google speech command recognition 20 , and a 70-percent reduction in image-reconstruction error on a Bayesian image-recovery task.
more »
« less
LSTMs for Keyword Spotting with ReRAM-based Compute-In-Memory Architectures
The increasingly central role of speech based human computer interaction necessitates on-device, low-latency, low-power, high-accuracy key word spotting (KWS). State-of-the-art accuracies on speech-related tasks have been achieved by long short-term memory (LSTM) neural network (NN) models. Such models are typically computationally intensive because of their heavy use of Matrix vector multiplication (MVM) operations. Compute-in-Memory (CIM) architectures, while well suited to MVM operations, have not seen widespread adoption for LSTMs. In this paper we adapt resistive random access memory based CIM architectures for KWS using LSTMs. We find that a hybrid system composed of CIM cores and digital cores achieves 90% test accuracy on the google speech data set at the cost of 25 uJ/decision. Our optimized architecture uses 5-bit inputs, and analog weights to produce 6-bit outputs. All digital computation are performed with 8-bit precision leading to a 3.7× improvement in computational efficiency compared to equivalent digital systems at that accuracy.
more »
« less
- Award ID(s):
- 2002921
- NSF-PAR ID:
- 10252999
- Date Published:
- Journal Name:
- 2021 IEEE International Symposium on Circuits and Systems
- Page Range / eLocation ID:
- 1 to 5
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
In this work, we leverage the uni-polar switching behavior of Spin-Orbit Torque Magnetic Random Access Memory (SOT-MRAM) to develop an efficient digital Computing-in-Memory (CiM) platform named XOR-CiM. XOR-CiM converts typical MRAM sub-arrays to massively parallel computational cores with ultra-high bandwidth, greatly reducing energy consumption dealing with convolutional layers and accelerating X(N)OR-intensive Binary Neural Networks (BNNs) inference. With a similar inference accuracy to digital CiMs, XOR-CiM achieves ∼4.5× and 1.8× higher energy-efficiency and speed-up compared to the recent MRAM-based CiM platforms.more » « less
-
Nowadays, research topics on AI accelerator designs have attracted great interest, where accelerating Deep Neural Network (DNN) using Processing-in-Memory (PIM) platforms is an actively-explored direction with great potential. PIM platforms, which simultaneously aims to address power- and memory-wall bottlenecks, have shown orders of performance enhancement in comparison to the conventional computing platforms with Von-Neumann architecture. As one direction of accelerating DNN in PIM, resistive memory array (aka. crossbar) has drawn great research interest owing to its analog current-mode weighted summation operation which intrinsically matches the dominant Multiplication-and-Accumulation (MAC) operation in DNN, making it one of the most promising candidates. An alternative direction for PIM-based DNN acceleration is through bulk bit-wise logic operations directly performed on the content in digital memories. Thanks to the high fault-tolerant characteristic of DNN, the latest algorithmic progression successfully quantized DNN parameters to low bit-width representations, while maintaining competitive accuracy levels. Such DNN quantization techniques essentially convert MAC operation to much simpler addition/subtraction or comparison operations, which can be performed by bulk bit-wise logic operations in a highly parallel fashion. In this paper, we build a comprehensive evaluation framework to quantitatively compare and analyze aforementioned PIM based analog and digital approaches for DNN acceleration.more » « less
-
Dense matrix multiply (MM) serves as one of the most heavily used kernels in deep learning applications. To cope with the high computation demands of these applications, heterogeneous architectures featuring both FPGA and dedicated ASIC accelerators have emerged as promising platforms. For example, the AMD/Xilinx Versal ACAP architecture combines general-purpose CPU cores and programmable logic (PL) with AI Engine processors (AIE) optimized for AI/ML. An array of 400 AI Engine processors executing at 1 GHz can theoretically provide up to 6.4 TFLOPs performance for 32-bit floating-point (fp32) data. However, machine learning models often contain both large and small MM operations. While large MM operations can be parallelized efficiently across many cores, small MM operations typically cannot. In our investigation, we observe that executing some small MM layers from the BERT natural language processing model on a large, monolithic MM accelerator in Versal ACAP achieved less than 5% of the theoretical peak performance. Therefore, one key question arises: How can we design accelerators to fully use the abundant computation resources under limited communication bandwidth for end-to-end applications with multiple MM layers of diverse sizes? We identify the biggest system throughput bottleneck resulting from the mismatch of massive computation resources of one monolithic accelerator and the various MM layers of small sizes in the application. To resolve this problem, we propose the CHARM framework to compose multiple diverse MM accelerator architectures working concurrently towards different layers within one application. CHARM includes analytical models which guide design space exploration to determine accelerator partitions and layer scheduling. To facilitate the system designs, CHARM automatically generates code, enabling thorough onboard design verification. We deploy the CHARM framework for four different deep learning applications, including BERT, ViT, NCF, MLP, on the AMD/Xilinx Versal ACAP VCK190 evaluation board. Our experiments show that we achieve 1.46 TFLOPs, 1.61 TFLOPs, 1.74 TFLOPs, and 2.94 TFLOPs inference throughput for BERT, ViT, NCF, MLP, respectively, which obtain 5.40x, 32.51x, 1.00x and 1.00x throughput gains compared to one monolithic accelerator.more » « less
-
Binary neural networks (BNNs) substitute complex arithmetic operations with simple bit-wise operations. The binarized weights and activations in BNNs can drastically reduce memory requirement and energy consumption, making it attractive for edge ML applications with limited resources. However, the severe memory capacity and energy constraints of low-power edge devices call for further reduction of BNN models beyond binarization. Weight pruning is a proven solution for reducing the size of many neural network (NN) models, but the binary nature of BNN weights make it difficult to identify insignificant weights to remove. In this paper, we present a pruning method based on latent weight with layer-level pruning sensitivity analysis which reduces the over-parameterization of BNNs, allowing for accuracy gains while drastically reducing the model size. Our method advocates for a heuristics that distinguishes weights by their latent weights, a real-valued vector used to compute the pseudogradient during backpropagation. It is tested using three different convolutional NNs on the MNIST, CIFAR-10, and Imagenette datasets with results indicating a 33%--46% reduction in operation count, with no accuracy loss, improving upon previous works in accuracy, model size, and total operation count.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/ISCAS51556.2021.9401295
