A partial-ISA MPSoC offers a way to expand system capability without treating every processor core as if it must support every instruction extension. In heterogeneous systems, some workloads make little use of features such as floating-point (FP), SIMD, or cryptographic instructions. Implementing their hardware on every core can consume area and power that might otherwise support additional cores or application-specific accelerators.
The design approach is selective: retain full-ISA cores for tasks that need the complete instruction set, while simplifying other cores by removing support for chosen extensions. A scheduler can place workloads on a suitable core as their execution requirements change. The result is a larger design space for balancing performance, power, and silicon area.
Why instruction-set support affects MPSoC design
Heterogeneous multiprocessor systems-on-chip combine general-purpose processors with specialized hardware. This mix helps a system serve varied workloads, from general application processing to tasks such as image processing, video decoding, and neural-network inference. Yet adding more processing resources is constrained by chip area and power budgets.
Instruction-set architecture (ISA) extensions add capabilities beyond a processor’s base instruction set. Their implementation can require additional datapath hardware. When all cores provide the same complete ISA support, that hardware is replicated across the system—even if some workloads rarely use it. The source study illustrates the potential cost with its evaluated out-of-order processor: FP support accounted for about 37% of processor area and 53% of power. These figures describe that specific modeled design, not a universal processor-level ratio.
Extension use also varies by workload. In the study’s RISC-V-based examples, FP instructions represented 0% of committed instructions for graph traversal, 2.5% for image processing, and 15% for linear algebra. Where extension use is low, a core dedicated to supporting that extension may provide limited value during much of the workload.
How a partial-ISA MPSoC works
In a partial-ISA design, selected cores omit microarchitectural support for one or more ISA extensions. The case study focused on removing FP support, including the associated floating-point datapath, from some cores. Simplifying those cores can free area and reduce their power requirements, creating room within the system budget for other resources.
Those resources might include additional in-order cores for task parallelism or application-specific hardware for defined workloads. The architectural choice depends on the intended use of the MPSoC: more general-purpose cores can help when independent tasks can run concurrently, while an accelerator can target a particular computational task.
Partial support does not mean every core has to abandon the extension. Full-ISA cores remain in the system so software that requires the omitted instructions can still execute. Workloads can run on partial-ISA cores during phases that do not need the missing capability and move to full-ISA cores when they do. This arrangement is intended to preserve binary compatibility while allowing the system to use simpler cores where appropriate.
Scheduling across full and partial cores
Core specialization makes task placement a system-management concern. The scheduler must account for whether a workload can run on a partial-ISA core, whether it needs an accelerator, and whether performance or energy is the priority. The study proposes two scheduling policies: performance-oriented and energy-oriented.
At a high level, a scheduling policy needs to match each task’s requirements to available processing resources. A task that uses an unsupported extension cannot remain on a partial-ISA core for that work; it must be assigned to a full-ISA core. Tasks that do not require the extension may use partial cores, leaving full cores available for workloads that depend on their broader capabilities. Where an appropriate accelerator is present, the system can also map suitable work to that hardware.
This approach makes the software and operating-system layer important to the architecture. Designers need to consider task migration, workload dependencies, and how the scheduler identifies extension requirements. The hardware savings alone do not determine the outcome: the system must coordinate resources effectively across execution phases.
Reported results and design trade-offs
The evaluation modeled different MPSoC configurations and workload scenarios, using RISC-V processors and FP instructions as its case study. It examined configurations with additional in-order cores and with accelerators for selected tasks. Under the performance-oriented policy, the study reported speedups of 2.81× for road-sign detection, 2.58× for video streaming, 1.76× for an edge-computing workload, and 1.2× for a task-parallel scenario. It also reported lower energy consumption in those cases, with reductions varying by scenario.
Under the energy-oriented policy, the study reported a 61% reduction in geometric-mean energy consumption compared with its highly efficient baseline, alongside increased performance. Across almost all evaluated scenarios, partial-ISA configurations also produced a better energy-delay product (EDP), a metric that combines execution time and energy.
These results should be interpreted within the study’s modeled configurations and workloads. They demonstrate the potential of reallocating hardware budget, rather than guaranteeing the same gains for other processors, applications, or implementation constraints. Any design decision still requires workload analysis and evaluation against the system’s particular area, power, and performance targets.
Practical considerations for system architects
For engineering teams exploring heterogeneous MPSoC design, partial-ISA is most relevant when extension demand differs substantially across workloads or execution phases. Before changing the core mix, designers should establish:
- Which extensions the target workloads use, and how frequently they are used.
- Which tasks can run on simplified cores and which require full-ISA support.
- Whether freed area and power are better invested in additional cores or dedicated accelerators.
- How scheduling and migration will direct work to suitable resources.
- How candidate configurations compare across performance, energy, area, and EDP.
Profiling is particularly important: an extension that is rarely used by one workload may be essential to another. A system therefore needs enough full-ISA capacity for tasks that depend on those instructions, while the partial cores and accelerators should reflect the expected workload mix.
Conclusion
Partial-ISA MPSoCs turn uniform extension support into an architectural design choice. By removing selected datapath hardware from some cores, designers can create budget for additional processing resources and direct workloads according to their instruction requirements. The practical benefit depends on workload characteristics, core and accelerator choices, and effective scheduling. For constrained designs, evaluating this approach can reveal alternatives that a conventional, fully uniform core configuration does not offer.