Axelera Metis Delivers Measured Edge-AI Power Results

Mario Bergeron (Hackster username: AlbertaBeef), an embedded-vision and edge-AI specialist, publishes Part 4 of his "Edge AI Power Benchmarking" series on Hackster.io, applying an independent power-measurement approach to the Axelera Metis edge accelerator. The post references vendor-published throughput figures from Axelera's axelera-ai-hub/voyager-sdk repository, citing ResNet-50 v1.5 throughput of 1756 FPS for the Metis M.2 board and 1946 FPS for the Metis PCIe board. Bergeron documents running tests on an AMD Ryzen AI MAX+ 395 PC and encountering an OpenCL runtime error; the post includes the error message and the suggested workaround --disable-opencl, noting an alternate AMD GPU support fix will be covered later in the writeup.

The article covers installing the Voyager SDK (in a Python virtual environment or Docker) and selecting ResNet-50 as the benchmark model due to its minimal post-processing overhead. Benchmark values are paired with instructions for downloading ImageNet test data and reproducing inference runs. The captured runtime error reads: "terminate called after throwing an instance of 'std::runtime_error' what(): No functional OpenCL platform of type '' found" - and the post records the practical --disable-opencl flag used to bypass the failure on the AMD host.

Axelera's community forum shows Bergeron posting in May 2026 while working toward the 1756 FPS M.2 target, with an intermediate result of 716 FPS after resolving the initial OpenCL issue. The FPS figures cited in this article are vendor-published benchmark targets from the Voyager SDK model zoo, not independently confirmed measured outputs from this series installment. Readers evaluating these numbers should reproduce them in their specific configuration, as host hardware, drivers, thermal conditions, and measurement method all affect outcomes.

This is Part 4 of Bergeron's "Edge AI Power Benchmarking" series, which previously covered Hailo-8 as a reference methodology (Part 1), external power insertion with ElmorLabs (Part 2), and INA228-based measurement (Part 3). Upcoming parts cover DeepX M1 (Part 5) and MemryX MX3 (Part 6), enabling cross-accelerator power comparisons under the same independent methodology. Independent reproducible benchmarking artifacts - with published commands, error logs, and dataset workflows - reduce uncertainty when evaluating hardware for power- or latency-constrained edge applications. The documented AMD OpenCL issue illustrates a class of integration friction practitioners frequently encounter when pairing vendor SDKs with non-reference host platforms.

• •Parts 5 and 6 of this series (DeepX M1, MemryX MX3) for cross-accelerator power comparisons under the same methodology.
• •Updates to the Voyager SDK or Axelera documentation addressing AMD OpenCL compatibility on AMD Ryzen AI MAX+ platforms.
• •Independent measurement repositories publishing raw power traces and scripts that complement this workflow.