Heterogeneous Computing

last updated 2026-08-31

Physics / mechanism

Heterogeneous computing describes systems that combine multiple distinct classes of compute unit within a single machine or fabric, so that each workload segment is executed on the substrate best matched to it. The canonical case in current research is a general-purpose CPU paired with a specialised accelerator: for machine learning inference, a Computing-in-Memory (CIM) accelerator executes matrix-vector multiplications directly in memory, while the CPU remains available as a complementary compute resource for the parts of the graph that map poorly onto the accelerator.

The central engineering problem is partitioning: deciding which operators run where, in what order, and with what data movement. For CPU-CIM systems the constraints are set by the memory technology itself. Resistive Random Access Memory (RRAM) arrays impose limited capacity, high write latency and limited write endurance, which means a partitioner must treat mapping decisions as expensive and semi-permanent rather than freely re-assignable. An Integer Linear Programming formulation has been proposed to minimise end-to-end inference latency under these RRAM constraints while capturing parallelism and low-level architectural effects, combining empirical profiling with analytical modelling rather than relying on either alone.

A second mechanism-level concern is state sharing across the heterogeneous fabric. In the 6G radio access network context, a proposed unified memory paradigm maps biological memory hierarchies onto heterogeneous computing fabrics, using emerging coherent interconnects so that microsecond-level reflexes, millisecond-level reasoning and long-term evolution share state across time scales. The argument is that message passing between disaggregated components forces the physical layer to compress high-dimensional states into low-dimensional metrics, creating a semantic bottleneck; replacing message passing with zero-copy observability removes that compression step.

Heterogeneity also shifts risk into the compilation and deployment toolchain. Software that adapts trained models to target devices is described as complex and fragile, with models that appear valid at a high level failing during deployment because of hidden resource limits, data layout requirements and platform-specific transformations. These failures can silently alter outputs rather than crash, which makes them difficult to detect and diagnose.

Competitive landscape

The sources frame heterogeneous computing not as a competitor to specific accelerator types but as the integration layer above them. CIM/RRAM accelerators are positioned as compelling for ML because they perform MVMs in memory, yet the same source argues that prior partitioning work is deficient precisely because it neglects the CPU as a complementary resource, implying that accelerator-only designs underperform mixed designs on end-to-end latency. Adjacent exotic substrates covered elsewhere in this knowledge base include analog, neuromorphic and optical computing, with mixed-signal analog-digital ICs discussed as an edge AI play ref.

The differentiating axis between approaches is therefore the interconnect and memory model rather than the arithmetic unit. Architectures that rely on interface-bound disaggregation trade away high-dimensional state visibility for modularity; memory-centric architectures built on coherent interconnects trade modularity for a shared cognitive continuum.

Evidence base

Frontier (open questions)

Synthesised 2026-08-31 from 7 KB sources by the resynth pipeline; citations are KB source slugs.

Recent mentions

Frontier questions