With the rise of deep learning (DL), our world braces for artificial intelligence (AI) in every edge device, creating an urgent need for edge-AI SoCs. This SoC hardware needs to support high throughput, reliable and secure AI processing at ultra-low power (ULP), with a very short time to market. With its strong legacy in edge solutions and open processing platforms, the EU is well-positioned to become a leader in this SoC market. However, this requires AI edge processing to become at least 100 times more energy-efficient, while offering sufficient flexibility and scalability to deal with AI as a fast-moving target. Since the design space of these complex SoCs is huge, advanced tooling is needed to make their design tractable. The CONVOLVE project (currently in Inital stage) addresses these roadblocks. It takes a holistic approach with innovations at all levels of the design hierarchy. Starting with an overview of SOTA DL processing support and our project methodology, this paper presents 8 important design choices largely impacting the energy efficiency and flexibility of DL hardware. Finding good solutions is key to making smart-edge computing a reality.

Transformer networks have emerged as the state-of-the-art approach for natural language processing tasks and are gaining popularity in other domains such as computer vision and audio processing. However, the efficient hardware acceleration of transformer models poses new challenges due to their high arithmetic intensities, large memory requirements, and complex dataflow dependencies. In this work, we propose ITA, a novel accelerator architecture for transformers and related models that targets efficient inference on embedded systems by exploiting 8-bit quantization and an innovative softmax implementation that operates exclusively on integer values. By computing on-the-fly in streaming mode, our softmax implementation minimizes data movement and energy consumption. ITA achieves competitive energy efficiency with respect to state-of-the-art transformer accelerators with 16.9 TOPS/W, while outperforming them in area efficiency with 5.93 TOPS/mm2in 22 nm fully-depleted silicon-on-insulator technology at 0.8 V.

Low-precision formats have recently driven major breakthroughs in neural network (NN) training and inference by reducing the memory footprint of the NN models and improving the energy efficiency of the underlying hardware architectures. Narrow integer data types have been vastly investigated for NN inference and have successfully been pushed to the extreme of ternary and binary representations. In contrast, most training-oriented platforms use at least 16-bit floating-point (FP) formats. Lower-precision data types such as 8-bit FP formats and mixed-precision techniques have only recently been explored in hardware implementations. We present MiniFloat-NN, a RISC-V instruction set architecture extension for low-precision NN training, providing support for two 8-bit and two 16-bit FP formats and expanding operations. The extension includes sum-of-dot-product instructions that accumulate the result in a larger format and three-term additions in two variations: expanding and non-expanding. We implement an ExSdotp unit to efficiently support in hardware both instruction types. The fused nature of the ExSdotp module prevents precision losses generated by the non-associativity of two consecutive FP additions while saving around 30% of the area and critical path compared to a cascade of two expanding fused multiply-add units. We replicate the ExSdotp module in a SIMD wrapper and integrate it into an open-source floating-point unit, which, coupled to an open-source RISC-V core, lays the foundation for future scalable architectures targeting low-precision and mixed-precision NN training. A cluster containing eight extended cores sharing a scratchpad memory, implemented in 12 nm FinFET technology, achieves up to 575 GFLOPS/W when computing FP8-to-FP16 GEMMs at 0.8 V, 1.26 GHz.

Recurrent neural networks such as Long Short-Term Memories (LSTMs) learn temporal dependencies by keeping an internal state, making them ideal for time-series problems such as speech recognition. However, the output-to-input feedback creates distinctive memory bandwidth and scalability challenges in designing accelerators for RNNs. We present Muntaniala, an RNN accelerator architecture for LSTM inference with a silicon-measured energy-efficiency of 3.25TOP/s/W and performance of 30.53GOP/s in UMC 65nm technology. The scalable design of Muntaniala allows running large RNN models by combining multiple tiles in a systolic array. We keep all parameters stationary on every die in the array, drastically reducing the I/O communication to only loading new features and sharing partial results with other dies. For quantifying the overall system power, including I/O power, we built Vau da Muntanialas, to the best of our knowledge, the first demonstration of a systolic multi-chip-on-PCB array of RNN accelerator. Our multi-die prototype performs LSTM inference with 192 hidden states in 330$\mu \text{s}$with a total system power of 9.0 mW at 10 MHz consuming 2.95$\mu \text{J}$. Targeting the 8/16-bit quantization implemented in Muntaniala, we show a phoneme error rate (PER) drop of approximately 3% with respect to floating-point (FP) on a 3L-384NH-123NI LSTM network on the TIMIT dataset.

Radio resource management (RRM) is critical in 5G mobile communications due to its ubiquity on every radio device and its low latency constraints. The rapidly evolving RRM algorithms with low latency requirements combined with the dense and massive 5G base station deployment ask for an on-the-edge RRM acceleration system with a tradeoff between flexibility, efficiency, and cost-making application-specific instruction-set processors (ASIPs) an optimal choice. In this work, we start from a baseline, simple RISC-V core and introduce instruction extensions coupled with software optimizations for maximizing the throughput of a selected set of recently proposed RRM algorithms based on models using multilayer perceptrons (MLPs) and recurrent neural networks (RNNs). Furthermore, we scale from a single-ASIP to a multi-ASIP acceleration system to further improve RRM throughput. For the single-ASIP system, we demonstrate an energy efficiency of 218 GMAC/s/W and a throughput of 566 MMAC/s corresponding to an improvement of 10× and 10.6×, respectively, over the single-core system with a baseline RV32IMC core. For the multi-ASIP system, we analyze the parallel speedup dependency on the input and output feature map (FM) size for fully connected and LSTM layers, achieving up to 10.2× speedup with 16 cores over a single extended RI5CY core for single LSTM layers and a speedup of 13.8× for a single fully connected layer. On the full RRM benchmark suite, we achieve an average overall speedup of 16.4×, 25.2×, 31.9×, and 38.8× on two, four, eight, and 16 cores, respectively, compared to our single-core RV32IMC baseline implementation.