arXiv:2603.13042v2 Announce Type: replace Abstract: Digital Compute-in-Memory (DCiM) accelerates neural networks by reducing data movement. Approximate DCiM can further improve power-performance-area (PPA), but demands accuracy-constrained co-optimization across coupled architecture and transistor-level choices. Building on OpenYield, we introduce Accuracy-Constrained Co-Optimization (ACCO) and present OpenACMv2, an open framework that operationalizes ACCO via two-level optimization: (1) accuracy-constrained architecture search of compressor combinations and SRAM macro parameters, driven by a
The increasing computational demands of AI, especially neural networks, are pushing the limits of traditional Von Neumann architectures, necessitating innovations like Compute-in-Memory to improve power-performance-area.
Advanced Compute-in-Memory frameworks are critical for increasing the efficiency and scalability of AI hardware, directly impacting the capabilities and cost of future AI systems and potentially reducing the energy footprint of AI.
This framework offers a more optimized path to deploy Approximate Digital Compute-in-Memory, promising significant PPA improvements by systematically balancing accuracy and hardware constraints.
- · AI hardware developers
- · Hyperscalers
- · AI accelerator manufacturers
- · Semiconductor foundries
- · Traditional CPU/GPU hardware architectures for AI
- · Less energy-efficient AI hardware designs
More energy-efficient and powerful AI chips become readily available for specialized tasks, particularly neural network acceleration.
The cost of deploying large-scale AI models may decrease, enabling wider adoption and development of more complex AI applications.
Increased compute efficiency could mitigate some of the energy bottleneck concerns facing the continued growth of advanced AI, impacting the sustainability of the AI industry.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG