Fumiya Shigemitsu

Rethinking Platform Abstractions for Heterogeneous Compute

Abstract

Mission-critical and mixed-criticality embedded systems — in automotive systems, robotics, and industrial automation — now execute end-to-end safety-relevant behaviors across CPUs and accelerators such as GPUs, NPUs, FPGAs, DSPs, and domain-specific engines. Hardware has responded with heterogeneous SoCs and is moving toward chiplet-based packages.

This hardware shift, however, is only partially reflected in the software stack. Many layers still rely on traditional CPU-centric operating system abstractions. This leaves the compute paths that determine real system behavior — spanning accelerator queues, data movement, and bus or chiplet interconnects — insufficiently visible to software as objects for reasoning, optimization, and control. The result is not just lost performance, but weaker analyzability: latency, contention, and responsibility boundaries become harder to locate.

This talk explores what platform abstraction such hardware requires. The platform needs three fundamental shifts: make accelerators first-class citizens, separate critical execution with multi-kernel boundaries, and make the compute path visible through system-wide observability. These moves provide the software foundation needed to “hack” heterogeneous computers within a mission-critical context: expose the paths that matter, harness performance, preserve analyzability and predictability, and keep the platform evolvable over time.

Biography

Fumiya Shigemitsu is a Technology Director at eSOL, where he works on embedded software products and embedded systems engineering. Since joining eSOL in 2019, he has contributed across roles from software engineer to technical lead, driving the development and evolution of real-time OS solutions and related tools across automotive and other embedded systems domains.