KEYNOTE SPEAKERS

Dynamic sparsity is intrinsic to biological computing and is key to its extreme power efficiency. Edge computing systems can improve their energy efficiency and reduce response latency by exploiting this neuromorphic principle. In edge computing, the neuromorphic approach for extracting features replaces conventional ADC and DSP with biological-inspired filters and event generators implemented in mixed-signal circuits. The resulting sparse feature events drive inference in dynamic-sparsity-aware neural network accelerators to reduce computational load and memory access. The demonstration of an ASIC combining the feature extractor and postprocessing network in an edge audio task shows the dynamic savings in power. Exploiting dynamic sparsity at all levels will benefit the design of intelligent devices for edge, wearable and biomedical applications.

In-memory computing (IMC) can significantly enhance tasks such as matrix-vector multiplication in edge AI algorithms. Despite its promises, IMC faces substantial challenges due to device and circuit non-idealities. The state-of-the-art spans a wide range of device technologies and diverse digital, analog, and mixed design approaches. Non-volatile and emerging memories are pivotal for integrating IMC into various products and edge AI applications, addressing energy and compute density constraints.

With the advent of large language models (LLMs) like GPT-4, only some 3D stacked technologies hold the potential to support the required scale. At the edge, traditional Von Neumann architectures, near-memory, and in-memory computing approaches each offer unique benefits and challenges. Energy efficiency varies significantly between SRAM IMC and purely digital architectures, with future process and technology advancements potentially narrowing this gap. Crosscutting issues such as on-device learning, quantization, security, and error correction highlight the complexity of IMC technologies. Beyond machine learning, IMC shows promise in hyperdimensional computing, associative memories, and probabilistic computing. Additionally, the ability of software mapping tools to efficiently adapt high-level algorithmic representations for edge AI is a crucial component of the overall picture.

While emerging memory AI solutions through IMC are not yet competitive with SRAM IMC, emerging optimizations and technologies hold disruptive potential. Edge AI algorithms continue to evolve, and hybrid approaches remain essential for edge AI hardware acceleration.