Skip to content
Shop
Search icon
Shop
AI Software

Metis® AIPU Benchmarks

The Metis AIPU is a high-performance accelerator for inference acceleration, delivering up to 5x faster throughput for vision tasks. More than just raw speed, the Voyager® toolchain optimizes the entire AI pipeline to ensure real-world application efficiency.

.

Performance Results: Metis vs. Competition

When compared to other AI accelerators, Metis consistently outperforms in key benchmarks. The chart and table below show the frames per second (FPS) processed by Metis, compared to the throughput of other AI accelerators.

 

    Inference-only performance (FPS)
Model Name Model Input Resolution Metisi Competitorii
EfficientNet-B4 224x224 576 No data
SSD-Mobilenet v2 300x300 2024 784
YOLOv5m 640x640 456 156
YOLOv5l 640x640 295 No data
YOLOv7 640x640 212 100
YOLOv8s 640x640 610 491
YOLOv8m 640x640 235 149
YOLOv8l 640x640 181 64

i PCIe Throughput is measured at the host-side on a system with Intel Core i9-13900K CPU
ii Competitor data obtained from public sources as of August 2026

We've tested numerous benchmarks and offer over 50 models in our Model Zoo for immediate use. At Axelera® AI, software is a top priority, and we continuously enhance our models and capabilities to simplify AI development and integration. Performance matters only when users can trust inference accuracy. Thanks to Metis' mixed precision architecture and our SDK's quantization, we achieve state-of-the-art accuracy.

The table below compares accuracy for various models running on full numerical precision (FP32) versus Metis after quantization with Voyager SDK. As shown, accuracy reduction is minimal in many cases. Our software team remains committed to ongoing optimizations in future updates.

Model Name Dataset FP32 Accuracy (%) Metis Accuracy (%)
EfficientNet-B4 ImageNet 79.27 78.63 data

Model Name Dataset FP32 Accuracy (mAP) Metis Accuracy (mAP)
SSD-Mobilenet v2 COCO 19.25 18.42
YOLOv5m COCO 44.94 44.13
YOLOv5l COCO 48.67 47.73
YOLOv7 COCO 51.02 50.46
YOLOv8s COCO 44.80 43.80
YOLOv8m COCO 50.16 48.83
YOLOv8l COCO 52.83 50.66

Performance, Efficiency, Accuracy: Verified by Third-Party Testing

HotTech Vision & Analysis, put leading AI accelerators to the test. Axelera Metis came out on top with the fastest performance, best efficiency, and highest detection accuracy across real-world computer vision workloads.

Download the Full Report
Performance, Efficiency, Accuracy: Verified by Third-Party Testing

Voyager® Toolchain

AI hardware is only as good as its software. That’s why we built the Voyager™ Toolchain, enabling developers to maximize our high-performance hardware. With a simple, high-level YAML-based language, developers can build computer vision pipelines that integrate multiple neural networks and complex image processing tasks. And with Voyager Wingman, developers can leverage our entire documentation and knowledge base to build their AI pipelines, optimize their applications or troubleshoot their implementations, all backed by their favorite LLM!

Voyager automatically compiles, optimizes, and deploys pipelines, running neural networks on all Axelera hardware while offloading preprocessing and post-processing to the host CPU, GPU, or media accelerator. Thanks to our flexible architecture, developers can allocate D-IMC cores as needed—whether running multiple models in parallel or dedicating cores to a single, compute-heavy model.

Application-level performance

Application-level performance

Running a Computer Vision application is much more than just running inference. At axelera.live we believe it's important to understand what the realized performance is – meaning, how long does it take to get the answer a user is looking for, that's the full end-to-end measurement. The axelera.live Voyager SDK helps optimize the entire data pipeline, including the parts that run on the host CPU or embedded GPU. Why does this matter? This means that both the developer and the users will have a better experience because the SDK will handle the work for the developer, and the user gets faster results.


Model Name Inference-only Performance (FPS) End-to-end Performance (FPS)i
EfficientNet-B4 576 563
SSD-Mobilenet v2 2024 1844
YOLOv5m 456 452
YOLOv5l 205 296
YOLOv7 212 212
YOLOv8s 610 617
YOLOv8m 235 236
YOLOv8l 181 181

As can be appreciated in the table, Voyager SDK manages to deliver the raw inference performance to the end-to-end application: by optimizing the execution of non-neural operations in the computer vision pipeline we ensure that the application can take full advantage of the unmatched capabilities of Metis.

The Voyager SDK is compatible with a variety of host architectures and platforms to accommodate different application environments. Additionally, the SDK allows embedding a pipeline into an inference service, providing various preconfigured solutions for use cases ranging from fully embedded applications to distributed processing of multiple 4K streams.


State-of-the-Art Digital In-Memory Computing

Why is Metis so powerful? One of the key innovations that sets Metis apart from its competition is its use of Digital In-Memory Computing (D-IMC) technology. D-IMC allows for the simultaneous processing and storage of data within memory cells, allowing extremely high throughput and power efficient matrix-vector-multiplication. This approach is particularly beneficial for AI workloads, which require high-speed data access and intensive computation, and all with an average power consumption below 10 watts!