Abstract Edge artificial intelligence (AI) chips represent a transformative class of semiconductor devices engineered to execute AI workloads directly on endpoint devices, eliminating persistent cloud dependency while satisfying stringent constraints on power consumption, latency, and data privacy. This... Edge artificial intelligence (AI) chips represent a transformative class of semiconductor devices engineered to execute AI workloads directly on endpoint devices, eliminating persistent cloud dependency while satisfying stringent constraints on power consumption, latency, and data privacy. This paper presents a comprehensive hierarchical architecture framework comprising four interdependent layers compute fabric, memory subsystem, interconnect network, and system integration that collectively address the multidimensional optimization challenges inherent in on-device intelligence. We demonstrate that edge AI architectures fundamentally diverge from conventional processors through specialized neural processing units (NPUs), near data processing memory hierarchies, and milliwatt to watt power profiles enabling always on operation. Critically, performance requirements span three orders of magnitude across deployment domains: automotive systems demand  10 TOPS within ~10 W envelopes for real time inference (<30 ms), smartphones operate under  1 W thermal constraints, and always on wearables target sub-100 mW operation with maximal data compression mandating application specific architectural specialization rather than monolithic designs. A foundational insight driving innovation is recognition that data movement, not computation, constitutes the dominant energy bottleneck; consequently, leading architectures implement compression aware memory subsystems that reduce bandwidth requirements by 5–30× while simultaneously increasing throughput and reducing power consumption by 30–50%. Furthermore, neuromorphic co-design principles inspired by biological neural systems, encompassing synaptic plasticity circuits, spike-based asynchronous computation, and in memory crossbar arrays, enable computational efficiencies approaching the brain's remarkable 20 W energy envelope. We establish that edge AI does not supplant cloud infrastructure but rather forms a symbiotic continuum in which edge devices resolve >95% of inference requests locally, while cloud resources handle model training and federated learning aggregation. This holistic co-design philosophy, spanning algorithms, circuits, and systems, positions edge AI architectures as foundational enablers of scalable, responsive, and privacy aware intelligent systems for the next generation of cyberphysical applications. Read More Read Less

Abstract This study introduces a comprehensive entropy characterization and complexity-aware benchmarking framework for residual data generated in learned image and video compression pipelines. Analyzing a large corpus of raw, content agnostic residual symbols at byte level granularity, we... This study introduces a comprehensive entropy characterization and complexity-aware benchmarking framework for residual data generated in learned image and video compression pipelines. Analyzing a large corpus of raw, content agnostic residual symbols at byte level granularity, we uncover critical statistical properties that challenge conventional entropy modeling assumptions. Empirical analysis reveals substantial heterogeneity across residuals, with entropy values spanning 3.5-5.0 bits per symbol (mean ±1 bit standard deviation) significantly below the theoretical maximum of 8 bits indicating rich statistical redundancy exploitable for compression. Symbol distributions exhibit pronounced heavy tailed characteristics, with only 10-15 dominant symbols accounting for 50-70% of occurrences, invalidating Gaussian or Laplacian assumptions pervasive in traditional codecs. Bit plane analysis further demonstrates extreme sparsity (95- 98%) in higher order planes, confirming residuals concentrate tightly around zero a signature of effective prediction. Crucially, entropy exhibits strong non stationarity across content instances, naturally stratifying into low, medium, and high complexity regimes. Rate distortion complexity (RDC) evaluation across hyperprior, autoregressive, and transformer based entropy models quantifies fundamental trade offs: while transformer architectures achieve up to 1.36 dB gains over hyperpriors at low bitrates, they impose >12× computational overhead, revealing diminishing returns beyond moderate complexity budgets. The autoregressive approach delivers optimal balance (0.78 dB gain at 5.2× complexity). Our dataset and analysis framework provide a content agnostic benchmark for evaluating entropy modeling efficacy independent of codec specific preprocessing, establishing that adaptive, complexity aware strategies not static global models are essential for efficient learned compression deployment across diverse hardware constraints. Read More Read Less

Abstract This paper presents a systematic comparative analysis of entropy modeling strategies in learned image compression (LIC), evaluating hyperprior (HP), autoregressive (AR), and transformer based (TR) approaches under a controlled experimental framework. Entropy modeling critically determines compression efficiency by... This paper presents a systematic comparative analysis of entropy modeling strategies in learned image compression (LIC), evaluating hyperprior (HP), autoregressive (AR), and transformer based (TR) approaches under a controlled experimental framework. Entropy modeling critically determines compression efficiency by estimating the probability distribution of latent representations, directly influencing the rate term in rate distortion optimization. To isolate the impact of entropy modeling, all architectures share identical encoder decoder backbones, latent dimensionality, and quantization schemes, with entropy modeling as the sole variable. Results reveal a clear hierarchy in entropy modeling accuracy, quantified by cross entropy gap: hyperprior models exhibit the largest gap due to limited spatial dependency capture; autoregressive models substantially reduce this gap by leveraging causal local context; and transformer based models achieve the smallest gap by exploiting long range global dependencies, particularly benefiting high complexity content. However, improved accuracy entails significant computational trade offs. Context utilization efficiency analysis shows autoregressive models excel with small contexts but face diminishing returns with larger ones. Crucially, decoder centric complexity emerges as a decisive practical constraint. Hyperprior models enable parallel decoding with minimal latency and linear scaling, making them ideal for latency sensitive applications. Autoregressive models suffer from strictly sequential decoding, resulting in super linear latency growth with resolution rendering them impractical for real time or high resolution scenarios. Transformerbased models offer superior compression gains but incur high memory demands and quadratic complexity in global attention configurations; however, configurable attention mechanisms enable controllable performance complexity trade-offs. Rate distortion complexity Pareto analysis confirms no single approach dominates universally: hyperpriors excel in low complexity regimes, transformers lead in high quality compression, and autoregressive models occupy an intermediate position. The study concludes that entropy modeling selection must balance compression efficiency against decoder feasibility, with scalable context utilization being critical for realworld deployment. Read More Read Less