Composable Heterogeneous Architecture for Scalable Machine Learning (CHASM) Workshop @ MICRO’26

Composable Heterogeneous Architecture for Scalable Machine Learning (CHASM) Workshop
Co-located with MICRO 2026
https://sites.google.com/view/chasm-micro26/home
November 1, 2026

The scale and complexity of modern artificial intelligence workloads, particularly the rise of agentic AI and multi-modal foundational models, have outpaced the capabilities of monolithic compute nodes.  As the industry scales, the future of AI infrastructure relies on the effective disaggregation of workloads across heterogeneous hardware.

The CHASM workshop aims to bring together researchers and practitioners from computer architecture, systems software, and machine learning to address the critical bottlenecks in disaggregated AI.  We seek to explore the full stack of challenges from hardware-level interconnects and memory pooling to the systems software required to compile, orchestrate, and verify highly dynamic execution graphs across heterogeneous accelerators.

• Systems Software and Compilation
  Compilation frameworks and intermediate representations for disaggregated, heterogeneous environments
  Compilation, scheduling, and runtime orchestration of agentic AI workloads as dynamic execution graphs
  Formal verification, debugging, and equivalence checking for distributed tensor algebra and parallel execution
  Online adaptation, test-time training, and adaptive inference compute in deployed agentic AI systems on disaggregated infrastructure
• Architecture and Hardware
  Architectures for scalable memory disaggregation and pooling in AI clusters
  High-performance interconnects and networking protocols tailored for heterogeneous accelerator communication
  Hardware-software co-design for composable AI systems
  
• Characterization and Evaluation
  Characterization of agentic, multi-modal, and tool-augmented Ai workloads on heterogeneous systems
  Marginal cost-efficiency analysis and performance benchmarking of contemporary AI accelerators in distributed topologies
  Novel simulation and evaluation methodologies for disaggregated datacenter architectures
  Telemetry, profiling, and bottleneck analysis in large-scale heterogeneous AI deployments

Submission Guidelines
We solicit both full papers (8-10 pages) and short/position papers (4-6 pages). Submissions are double-blinded. The page limit includes figures, tables, and appendices, but excludes references. Please use standard LaTeX or Word ACM templates. All submissions will need to be made via EasyChair (link) . Each submission will be reviewed by at least three reviewers from the program committee. Papers will be reviewed for novelty, quality, technical strength, and relevance to the workshop.