31 October 2022 in IEEE Journal of Solid-State Circuits
DIgital-ANAlog (DIANA), a heterogeneous multi-core accelerator, combines a reduced instruction set computer -five (RISC-V) host processor with an analog in-memory computing (AIMC) artificial intelligence (AI) accelerator and a digital reconfigurable deep neural network (DNN) accelerator in a single system-on-chip (SoC) to support a wide variety of neural network (NN) workloads. AIMC cores can bring extreme computational parallelism and efficiency at the expense of accuracy and dataflow flexibility. Digital AI co-processors, on the other hand, guarantee accuracy through deterministic compute, but cannot achieve the same computational density and efficiency. DIANA exploits this fundamental tradeoff by integrating both types of cores in a shared and optimized memory system, to enable seamless execution of the workloads on the parallel cores. The system’s performance benefits further from pipelined parallel execution across both accelerator cores and enhanced AIMC spatial unrolling techniques, leading to drastically reduced execution latency and reduced memory footprints. The design has been implemented in a 22-nm technology and achieves peak efficiencies of 600 TOP/s/W for the AIMC core (I/W/O: 7/1.5/6 bit) and 14 TOP/s/W (I/W/O: 8/8/8 bit) for the digital accelerator, respectively. End-to-end performance evaluation of CIFAR-10 and ImageNet classification workloads is carried out on the chip, reporting 7.02 and 5.56 TOP/s/W, respectively, at the system level.
5 October 2020 in IEEE Transactions on Electron Devices
We propose, for the first time, an indium gallium zinc oxide (IGZO)-based 2T1C compute cell (IGZO-cell) for analog in-memory computing. To assess the impact of an IGZO-cell-based array including the periphery on power and accuracy, a PyTorch framework was developed to analytically modeled analog components. The results are reported for a ResNet20 network on the Canadian Institute For Advanced Research-10 (CIFAR-10) benchmark. The state-of-the-art energy efficiency of 15 peta operations per second (POPS)/W including the periphery is achieved by using our proposed IGZO-cell with CMOS compatibility. Finally, it is shown that, with a properly trained neural network model, there is no degradation of test accuracy with 10% device to device variability for the IGZO devices.
1 June 2018 in Integration
9 April 2019 in IEEE Transactions on Circuits and Systems II: Express Briefs
The manufacturing cost of contactless 3-D circuits with inductive links is explored for the first time. The high-volume manufacturing (HVM) cost of contactless 3-D ICs is modeled by considering the on-chip inductor area and other important parameters, such as the coupling efficiency, which depend upon specific steps of the fabrication process. Inductive links based on current mode logic (CML) transceivers are simulated across a number of integration scenarios in a commercial 65-nm CMOS technology process, thereby evaluating the cost of these manufacturing options for performance and power objectives. Assuming a 10 Gb/s data rate, the power and area of the inductive link are determined for wafer-to-wafer and die-to-wafer integration and communication distances in the range of 5 $\mu \text{m}$ (face-to-face) to 120 $\mu \text{m}$ (face-to-back). An HVM cost estimation for each investigated inductive link scheme is provided for these integration approaches. Additional processing steps attributed to 3-D integration increase the aggregate cost by 2.5%–4.5% for a case study of a three-tier stack, compared to a conventional 2-D process. Moreover, the yield is explored for each integration approach considering the inductor area for contactless 3-D ICs.