Benchmark the Complete Workload Path
Benchmark the Complete Workload Path
Hardware Accelerators: Select by Workload Evidence
A computational kernel is the specific computation that consumes time or energy, such as a filter applied to sensor samples. A processor can use an accelerator for a costly task. First measure the real workload and find the slow or power-hungry step. Then compare time, energy per task, cost, volume, and the tools needed to maintain each choice. Include data transfer time; a fast kernel alone may not make the whole system faster.
Fixed-point arithmetic represents numbers with a fixed scale; floating-point uses a changing scale to cover a wider range. Both can round results, so compare their error against the task limit. Contention occurs when processors compete for shared memory or another shared resource and must wait. Verification coverage records which design behaviors have been checked by tests. Fabrication yield is the share of manufactured chips that pass their checks. Errata are known faults or limits documented after release. Heterogeneous computing combines different types of processor for different tasks. A GPU handles many parallel operations; an NPU is designed for neural-network calculations. A DSP is a processor built for repeated signal math. An FPGA is a chip whose logic can be configured; a bitstream is the file that sets that logic. An ASIC is a chip designed for a fixed job; its mask version identifies the pattern used to make that chip. HDL is the language used to describe digital hardware. NRE means one-time design and engineering cost. Timing closure means showing the hardware design meets its timing targets.
| Platform class | Strong fit | Main cost or limit | Evidence before selection |
|---|---|---|---|
| DSP | Repeated multiply-accumulate work, filtering, audio, communications | Specialised memory/compiler model; weaker fit for irregular control | Kernel profile, fixed/floating-point error, worst-case latency, energy/task, compiler/library version |
| FPGA | Deterministic parallel pipelines, custom interfaces, reconfigurable timing | HDL/tool skill, verification effort, static power, bitstream lifecycle | Timing closure, resource use, power at workload, interface tests, bitstream provenance and update/rollback |
| ASIC | Stable high-volume function with strict power/area/latency target | Highest NRE and longest change cycle; faults are expensive after fabrication | Frozen requirement, verification/coverage, process and volume case, yield, lifecycle supply and errata plan |
| Heterogeneous SoC | Control on CPU/MCU plus selected DSP/GPU/NPU/FPGA engines | Data movement, shared memory, scheduler, driver, and observability complexity | End-to-end trace including transfers, contention, fallback path, power states, driver/runtime compatibility |
Use a staged selection:
Save inputs and expected outputs for the kernel as a test fixture. Deliberately change the calculation and check that the fixture catches the wrong result; this makes the fixture mutation-sensitive.
First, Measure the representative software path on the simplest viable processor. Next, Isolate the kernel that misses a deadline or energy budget; retain its inputs and expected outputs as a mutation-sensitive fixture. Then, Benchmark the complete path on each candidate, including data conversion, transfers, queueing, wake-up, and fallback—not just accelerator execution. After that, Price NRE, tools, verification, unit volume, supply, field updates, and the years the evidence must remain reproducible. Finally, Keep a processor-only safe or degraded path where the product risk requires it, and state which requirements that path cannot meet.
Record the workload and deadline so another team can repeat the test. State the allowed numeric error and compare each result with it. Include time and energy for the whole path, including memory and data transfers: a fast calculation can still lose time moving data.
Record the tool and implementation version, such as an FPGA bitstream or ASIC mask, so the result can be tied to the design tested. Add the one-time cost, expected volume, rejected options, fallback and update path, owner, and the measurement that would reopen the choice. Peak operations per second alone cannot show whether the complete system meets its goal.