Version: v1.7

Voyager SDK release notes v1.7

Voyager SDK release notes v1.7.0

Voyager SDK v1.7.0 expands the pipeline builder capabilities, adds new computer vision models, and broadens hardware and platform support.

Support for Metis 1-chip PCIe Rev2 2 GB and 8 GB DDR variants.
New computer vision models including YOLOv4, YOLOv4-CSP-Leaky, and YOLO26 semantic segmentation family.
We introduce capabilities in the Pipeline Builder API such as support for streaming video, cascaded model pipelines and a scheduling API.
DMA support in virtual machines (Linux and Windows).
New documentation portal launched at docs.axelera.ai, covering Voyager SDK and all Axelera AI hardware products.

Release Qualification

This is a production-ready release of Voyager SDK. Software components and features that are in development are marked with one of the following maturity labels:

Experimental: May change or be removed without notice; no support guarantees.
Alpha: Usable but incomplete; breaking changes possible.
Beta: Feature-complete but not fully stable; committed to developing this further in future releases.

Installation and Compatibility

Installation

pip install as root now supported. Installation via pip with superuser privileges is now fixed.
axelera-devkit now supports PyTorch versions 2.7–2.12 (previously 2.7–2.10).
axelera-devkit now supports NumPy versions 1.x and 2.x. Note: axelera-devkit[all] still pins NumPy to < 2.0.0.

Release Compatibility Matrix

The following compatibility matrix describes the recommended and supported versions of firmware and driver per Voyager SDK release. Consult it before installing or upgrading a card or Voyager SDK release.

Tip! axversion outputs the SDK version and axversion --driver outputs the driver version. axdevice outputs the firmware and board controller firmware version.

Recommended = version shipped with the SDK release.

Supported = versions tested with the SDK release.

Note: Other versions may work but are not actively tested. An upgrade of the card's flashed firmware and board controller firmware to a compatible version using axdevice interactive_flash_update script is advised.

Release	Board controller (Recommended)	Board controller (Supported)	Flashed Firmware (Recommended)	Flashed Firmware (Supported)	PCIe driver - Linux (Recommended)	PCIe driver - Linux (Supported)	PCIe driver - Windows (Recommended)	PCIe driver - Windows (Supported)
v1.6.0	7.4	7.0	1.6.0	1.5.0 1.4.0	1.4.16	1.4.10 1.4.4	1.3.4	1.3.1 1.3.0
v1.6.1	7.4	7.0	1.6.0	1.5.0 1.4.0	1.4.17	1.4.16 1.4.10 1.4.4	1.3.11	1.3.4 1.3.1 1.3.0
v1.7.0	7.4	7.0	1.7.0	1.6.0 1.5.0 1.4.0	1.5.5		1.3.11	1.3.5 1.3.4 1.3.1 1.3.0

Metis M.2 Max upgrade notes

Metis M.2 Max is required to be updated to recommended board controller and firmware versions. The axdevice interactive_flash_update script handles board variant selection automatically. Enable the average power controller if deploying on hosts with limited power delivery. See the Power Management Guide.

New Features / Support

New Axelera AI Cards and Systems

Support for Metis 1-chip PCIe Rev 2.0 cards with 2GB and 8GB DDR memory configurations.
Default device power limit on Metis M.2 Max lowered from 11.0 W to 8.5 W for broader host compatibility. When paired with hosts which are compliant with PCIe SIG M.2 Specification Revision 4.0 power rating or higher, increasing the power limit setting results in potential performance gains. Users are recommended to experiment in steps of 0.1 W up to 11 W. On the other hand, configuring the power limit setting to a lower value enables M.2 Max in power-constrained hosts by trading off performance e.g. a power limit setting of 4 W when paired with embedded SBCs.

Host Platform Support

Validated hardware platforms

OnLogic K801 (12th Gen Intel Core-i)

The full list of Validated Host Systems for Axelera Metis Cards is available here.

Operating Systems

Windows IoT native support for Metis on x86 host machines.
The standalone axelera-rt wheel with ManyLinux support remains the recommended installation path for Yocto images. Refer to the meta-axelera Yocto layer for the recommended recipe structure.

Virtualization support

Production ready for Metis PCIe passthrough in VMs, with the entire runtime stack including the driver running inside the guest. The PCIe driver now supports DMA when running inside a virtual machine. This enables PCIe passthrough for Metis even in VMs that have not been configured with multi-MSI support.

New Networks Supported

Voyager SDK model zoo includes computer vision tasks and LLMs. For a full list of supported models and data about their performance and accuracy see here.

Models that are supported but not included in the model zoo are documented here.

For convenience, pre-compiled models are available to download by running axdownloadmodel in the parent folder of Voyager SDK.

The release includes new YAML files for all new models offered in our model zoo in this release (see tables below).

New models for Image Classification

Model Name	Resolution	Format
MobileNetV3-small	224x224	PyTorch
MobileNetV3-large	224x224	PyTorch

New models for Object Detection

Model Name	Resolution	Format
YOLOv4	416x416	Darknet
YOLOv4-CSP-Leaky	640x640	Darknet

The YAML model deployment and Pipeline Builder now support the Darknet model format.

New models for Semantic Segmentation

Model Name	Resolution	Format
YOLO26n-Seg	1024x1024	ONNX
YOLO26s-Seg	1024x1024	ONNX
YOLO26m-Seg	1024x1024	ONNX
YOLO26l-Seg	1024x1024	ONNX
YOLO26x-Seg	1024x1024	ONNX

Trained on the Cityscapes dataset. Note: model zoo FP32 accuracy uses a 1024×1024 input dimension aligned with Cityscapes, which differs from the Ultralytics default of 1024×2048; reported accuracy numbers will differ accordingly.

The YAML Pipeline Builder supports both Accuracy mode and Performance mode for semantic segmentation models.

AI Pipeline Builder

[Alpha] Pipeline Builder API

The Python-native Pipeline Builder API that was introduced in v1.6 now gains a streaming runtime and many new capabilities.

Streaming and scheduling

a new Scheduler owns model instances and connections, caches loaded models, and distributes work across the available AIPU cores.
pipeline.stream pipelines frames across cores automatically. Per-model priority can be tuned via the core_allocation argument to op.load. A single-core option runs each frame end-to-end in the calling thread for easy operator stepping/debugging.
pipeline.batch allows optimised scheduling of multiple inputs without having to use asynchronous APIs.

Video decode and sources

A built-in video decoder binding for ffmpeg and OpenCV with a zero-copy streaming API. cv.create_source enables accelerated decoding into a new Image class that enables zero-copy whilst facilitating access via a to_numpy() read only view.

New task types

Depth estimation
Oriented Bounding Boxes (OBB)
Re-identification

Other

Multi-level cascade pipelines now track coordinates correctly through each stage, extended to OBB and pose keypoints, so per-object crops map back to the original frame.
New and expanded API documentation, getting-started and model-compilation tutorials, and a coordinate-system tutorial, as well as a new set of examples in examples/pipeline_builder/.

Demo Scripts

New standalone runnable demo scripts are self-contained and depend only on axelera.runtime, providing a ready starting point for building custom inference pipelines.

Demo Script	Description	Model
`classification.py`	ImageNet top-5 classification with standard ImageNet preprocessing.	`squeezenet1.0-imagenet.axm`
`detection.py`	YOLOv8 object detection on COCO (80 classes) with NMS.	`yolov8n-coco.axm`
`detection_vary_confidence.py`	Object detection with runtime confidence threshold variation every 60 frames.	`yolov8n-coco.axm`
`pose_detection.py`	YOLOv8 human pose estimation with 17 COCO keypoints.	`yolov8npose-coco.axm`
`segmentation.py`	YOLOv8 instance segmentation with prototype-based mask prediction.	`yolov8nseg-coco.axm`
`depth_estimation.py`	Monocular depth estimation with FastDepth (NYU Depth V2).	`fastdepth-nyudepthv2-onnx.axm`
`obb.py`	YOLO11n oriented-bounding-box detection on DOTA (15 classes).	`yolo11n-obb-dotav1-onnx.axm`
`tracking.py`	Multi-object tracking with state lifecycle, detection correlation, and class filtering.	`yolov8n-coco.axm`
`tracking_with_classification.py`	Detection → filtering → tracking → per-track classification cascade.	`yolov8n-coco.axm`, `squeezenet1.0-imagenet.axm`

YAML Pipeline Builder

Per-stream crop: crop configuration can be set independently per stream via the usage ./inference.py yolov8n-coco crop[left,top,width,height]:rtsp://...
Robust ONNX preamble: NHWC support, new transform operators, output validation, and an optimizer fix, so extracted preprocessing matches the original model more reliably.
Decoder improvements: faster fused semantic-segmentation decoder, additionally improves accuracy over the decoder in 1.6.
fp16 (half-precision) support added to OpenCL kernels where hardware support is available.
DMA-buf import/export via a Khronos extension cl_khr_external_memory for copy-free OpenCL interoperability on supported platforms.
Multiplanar image support enabled for OpenCL.
OpenCL 3.0 compatibility.

[Beta] Model Compiler

Framework and dependency support expanded:
- PyTorch 2.10–2.12 support added across the compiler toolchain; supported range is now 2.7–2.12.
- NumPy 2.x compatibility across axelera-tvm, qtools, and onnx2torch.
- Python 3.13 wheel validation and aarch64 wheel builds.
Unique .bin file paths during compilation, preventing collisions when multiple models are compiled into the same output directory.
Quantization accuracy improvement: Hardswish LUT updated to v4 (uniform 16-bin) for better activation-quantization accuracy.

Tools

axWinGrantLargePages.exe (Windows): New utility that grants the Large Pages privilege to improve DMA performance. libaxldev will warn at runtime if this has not been run.
axmonitor improvements:
- Per-sensor power statistics now available for Metis 4-chip PCIe card and Metis Compute Board.
- MVM utilization and stack usage now reported per AI core.

Breaking Changes

None

Known Issues and Limitations

IMPORTANT - Memory leak in GStreamer software video decode (`gst-libav` 1.24.2)

Pipelines that perform software H.264 decoding via GStreamer's avdec_h264 element (gst-libav 1.24.2, as shipped with GStreamer 1.24.2 on Ubuntu 24) leak approximately 104 bytes per decoded video frame, per decoder. This is a defect in gst-libav itself (avdec/avviddec does not free the per-frame AVPacket), not in the Voyager SDK, and the SDK cannot workaround it.

The leak grows linearly with runtime and scales with the number of decoded streams, so it is most noticeable in long running, multi-stream deployments. For example, 16 streams at 30 FPS leak roughly 50 KiB/s in aggregate.

Impact: affects only pipelines that use GStreamer software video decode. Hardware-accelerated decode paths are not affected.

Mitigation: the only fix is to upgrade GStreamer / gst-libav to 1.28.3 or later, which resolves the underlying AVPacket leak. Or alternatively where possible, prefer hardware-accelerated decode over software avdec_h264. Restarting long-running pipelines periodically bounds the resident memory growth.

Other

Performance variability is observed on certain hosts. Inference-only performance (FPS) drops of up to 5-10% is observed on yolox-x-crowdhuman-onnx and yolo11l-obb-dotav1-onnx compared to SDK Release v1.5.
Numpy compatibility: axelera-devkit supports only NumPy 1.x APIs and is not compatible with NumPy 2.x. axelera-rt supports both NumPy 1.x and 2.x but constrains the numpy dependency to prevent breaking axelera-devkit on Linux. The Windows runtime environment supports NumPy 2.x.
Python 3.13 incompatibility with wheel installer on Ubuntu 24.04. Not reproducible with Python 3.12 (default for Ubuntu 24.04).
Metis Compute Board video output rendering is choppy (1–2 FPS). This impacts rendering to display only, inference performance is not impacted.
YOLO26* models (e.g. yolo26x-obb-dotav1-onnx) sometimes fail to deploy unexpectedly.
MobileNetV3 may yield degraded accuracy on some combinations of hosts and cards.
Device monitoring with AxMonitor is not supported on single-MSI hosts. For some systems with single-MSI hosts, device monitoring with AxMonitor does not display any data. An example of a host with this issue is Arduino Portenta X8 Mini.

System Requirement

Development Environment

For model compiling purposes, these are the host requirements:

Requirement	Detail
OS	Linux Ubuntu 22.04, Ubuntu 24.04, Docker (on Windows or Linux), Windows + WSL/Ubuntu
CPU architecture	ARM64, x86, x86_64
Recommended CPU	Intel Core-i5 or equivalent
Minimum System Memory	16 GB (large models may require swap partition)
Recommended System Memory	32 GB

Runtime Environment

This release is expected to work with Intel (x86), AMD (x86) and Arm64 host CPUs. See here for a list of validated host systems for Axelera Metis AI Accelerator Cards.

Further Support

For blog posts, projects and technical support please visit Axelera AI Customer Portal.
For technical documents and guides please visit docs.axelera.ai.