Tsinghua's CHASE framework delivers 6.2x speedups for AI workloads

Researchers from Tsinghua University have published a paper on arXiv detailing CHASE, an application-driven architecture exploration framework for cross-layer heterogeneous systems (XHS). As AI and HPC workloads increasingly combine dense tensor computation, sparse kernels, large memory footprints, and communication-intensive collectives, hardware design must coordinate choices across accelerators, memory tiers, scale-up fabrics, and cluster networks. The XHS design space is immense and constrained by power, switch radix, cabling, and cost, making brute-force search intractable.

CHASE addresses this with a decoupled two-level loop: an inner mapper translates hardware-independent workload DAGs into topology-aware event traces, a calibrated event-driven simulator evaluates each mapping, and an outer telemetry-guided optimizer evolves the hardware graph. Empirical results show the mapper stays within 6.06% of exhaustive optima while cutting mapping time by 60.5%. Compute-model errors average 4.4–7.5%, and communication validation reproduces real platform trends. The outer search converges to near-global optima in 64 iterations. Case studies reveal that sparse workloads favor criticality-aware heterogeneous pods, while LLM inference prefers scale-up islands, yielding 6.20× and 2.12× geomean speedups, respectively, with reduced cost and power.