Advances in deep neural networks (DNNs) and the availability of massive real-world data have enabled superhuman levels of accuracy on many AI tasks and ushered the explosive growth of AI workloads across the spectrum of computing devices. However, their superior accuracy comes at a high computational cost, which necessitates approaches beyond traditional computing paradigms to improve their operational efficiency. Leveraging the application-level insight of error resilience, we demonstrate how approximate computing (AxC) can significantly boost the efficiency of AI platforms and play a pivotal role in the broader adoption of AI-based applications and services. To this end, we present RaPiD, a multi-tera operations per second (TOPS) AI hardware accelerator core (fabricated at 14-nm technology) that we built from the ground-up using AxC techniques across the stack including algorithms, architecture, programmability, and hardware. We highlight the workload-guided systematic explorations of AxC techniques for AI, including custom number representations, quantization/pruning methodologies, mixed-precision architecture design, instruction sets, and compiler technologies with quality programmability, employed in the RaPiD accelerator.

A processor core is presented for AI training and inference products. Leading-edge compute efficiency is achieved for robust fp16 training via efficient heterogeneous 2-D systolic array-SIMD compute engines leveraging compact DLFloat16 FPUs. Architectural flexibility is maintained for very high compute utilization across neural network topologies. A modular dual-corelet architecture with a shared scratchpad and a software-controlled network/memory interface enables scalability to many-core SoCs and large-scale systems. The 14nm AI core achieves fp16 peak performance of 3.0 TFLOPS at 0.62V and 1.4 TFLOPS/W at 0.54V.

A multi-TOPS AI core is presented for acceleration of deep learning training and inference in systems from edge devices to data centers. With a programmable architecture and custom ISA, this engine achieves >90% sustained utilization across the range of neural network topologies by employing a dataflow architecture and an on-chip scratchpad hierarchy. Compute precision is optimized at 16b floating point (fp 16) for high model accuracy in training and inference as well as 1b/2b (bi-nary/ternary) integer for aggressive inference performance. At 1.5 GHz, the AI core prototype achieves 1.5 TFLOPS fp 16, 12 TOPS ternary, or 24 TOPS binary peak performance in 14nm CMOS.

The combination of growth in compute capabilities and availability of large datasets has led to a re-birth of deep learning. Deep Neural Networks (DNNs) have become state-of-the-art in a variety of machine learning tasks spanning domains across vision, speech, and machine translation. Deep Learning (DL) achieves high accuracy in these tasks at the expense of 100s of ExaOps of computation; posing significant challenges to efficient large-scale deployment in both resource-constrained environments and data centers. One of the key enablers to improve operational efficiency of DNNs is the observation that when extracting deep insight from vast quantities of structured and unstructured data the exactness imposed by traditional computing is not required. Relaxing the "exactness" constraint enables exploiting opportunities for approximate computing across all layers of the system stack. In this talk we present a multi-TOPS AI core [3] for acceleration of deep learning training and inference in systems from edge devices to data centers. We demonstrate that to derive high sustained utilization and energy efficiency from the AI core requires ground-up re-thinking to exploit approximate computing across the stack including algorithms, architecture, programmability, and hardware. Model accuracy is the fundamental measure of deep learning quality. The compute engine precision in our AI core is carefully calibrated to realize significant reduction in area and power while not compromising numerical accuracy. Our research at the DL algorithms/applications-level [2] shows that it is possible to carefully tune the precision of both weights and activations to as low as 2-bits for inference and was used to guide the choices of compute precision supported in the architecture and hardware for both training and inference. Similarly, distributed DL training's scalability is impacted by the communication overhead to exchange gradients and weights after each mini-batch. Our research on gradient compression [1] shows by selectively sending gradients larger than a threshold, and by further choosing the threshold based on the importance of the gradient we achieve achieve compression ratio of 40X for convolutional layers, and up to 200X for fully-connected layers of the network without losing model accuracy. These results guide the choice of interconnection network topology exploration for a system of accelerators built using the AI core. Overall, our work shows how the benefits from exploiting approximation using algorithm/application's robustness to tolerate reduced precision, and compressed data communication can be combined effectively with the architecture and hardware of the accelerator designed to support these reduced-precision computation and compressed data communication. Our results demonstate improved end-to-end efficiency of the DL accelerator across different metrics such as high sustained TOPs, high TOPs/watt and TOPs/mm2 catering to different operating environments for both training and inference.

This letter presents a multi-TOPS AI accelerator core for deep learning training and inference. With a programmable architecture and custom ISA, this engine achieves >90% sustained utilization across the range of neural network topologies by employing a dataflow architecture to provide high throughput and an on-chip scratchpad hierarchy to meet the bandwidth demands of the compute units. A custom 16b floating point (fp16) representation with 1 sign bit, 6 exponent bits, and 9 mantissa bits has also been developed for high model accuracy in training and inference as well as 1b/2b (binary/ternary) integer for aggressive inference performance. At 1.5 GHz, the AI core prototype achieves 1.5 TFLOPS fp16, 12 TOPS ternary, or 24 TOPS binary peak performance in 14-nm CMOS.

The scale of technology node increases power-density dynamically. Various techniques are proposed to reduce the power consumption. One approach is Dual-Supply Voltage (DSV). DSV is to apply a lower supply voltage on selected non-critical gates for power saving while maintaining chip performance at the same time. In order to facilitate the power design in DSV, the same voltage gates are grouped to form islands. [21] presents a flow to generate and place voltage islands. However, after relocating gates to voltage islands, the original placement is changed and the timing might become worse. In this paper, we propose algorithms to redistribute gates for performance improvement. At the same time, all the DSV island/gate constraints are satisfied. On tested designs, our algorithm greatly improved the timing, and the final worst slack is less than 10ps worse than that of the original design.

Power consumption has become a major consideration in nanometer chip design. Since the dynamic power is proportional to Vdd 2 , and the static power is proportional to Vdd, lowering power supply voltage is an efficient method to reduce the power usage. In this paper, we adopt the row-based dual-supply voltage (DSV) scheme. DSV assigns a low power supply voltage to timing noncritical gates for the power saving. Contrary to the traditional region-based voltage island works, the row-based approach creates voltage islands along the circuit rows. This kind of fine grid voltage islands give more flexibility on gate voltage assignment such that the low voltage gates can be selected primarily based on timing and design logic, which in turn minimizes the shifter insertion. We present a two-stage flow to generate voltage islands for every two mirrored circuit rows and place gates legally inside each island. To the best of our knowledge, this is also the first work to cover latch and LCB (local clock buffer) handling and shifter placement in voltage island generation. The experimental results demonstrate the effectiveness and efficiency of our approach.