Voyager SDK release notes v1.8

Voyager SDK release notes v1.8.0​

• Metis 4-chip PCIe with 8GB RAM and Metis chip-down designs are supported in this release.
• Pipeline builder API [Alpha] has a more finalized API surface with video streaming and model cascading support as well as new task types and operators. The Pipeline Builder is still "alpha" with no performance guarantees; performance optimizations will be delivered in future releases.
• Pre-compiled kernel binaries are now mandatory. The AxModel format moves to major version 5 and the RISC-V toolchain is no longer part of a runtime installation. Existing models must be re-deployed or re-downloaded. See Breaking Changes.
• New: standalone Python model recipes (Alpha). Selected Model Zoo models now ship a readable, self-contained build script under model_recipes/ with no zoo runtime dependency, so users can copy and adapt it directly.
• Linux kernel support extended down to 5.4 (Yocto dunfell) and up to 6.17.

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.

Product naming​

Axelera AI hardware products were renamed in August 2026. The Voyager SDK release notes and command output still use the previous Metis product names, so the mapping is below.

The silicon is unchanged, and Metis remains the architecture name. A name such as Axelera Edge 130p reads as optimization lane, architecture generation, relative performance and form factor. See Understanding Axelera product names for the full scheme, and the Hardware documentation for the product pages.

Installation and Compatibility​

IMPORTANT - AxModel version 5: models must be rebuilt​

v1.8.0 bumps the AxModel major version from 4 to 5. Every AxModel now contains a pre-compiled kernel, which allows the runtime to execute inference without compiling kernel source at load time - and therefore allows the RISC-V toolchain to be dropped from the runtime environment entirely.

Consequence: AxModels produced by v1.7 or earlier cannot be run or inspected by v1.8. axrunmodel, inference.py and axmodeltool will fail with:

Unsupported model version: 4.0, expected at least 5.0
Please re-deploy or re-download the model.

Action required: re-deploy models with the v1.8 toolchain, or re-download the pre-compiled models with axdownloadmodel.

IMPORTANT - Driver upgrade is mandatory​

For both Linux and Windows, the minimum supported driver version in v1.8.0 is equal to the shipped version, so no earlier driver is accepted:

• Linux: install metis-dkms 1.6.2
• Windows: install MetisDriver-1.3.14. This driver version is certified by Microsoft.

Installation​

• Windows now requires only three packages: Axelera Device Package, Axelera Runtime, and Axelera Services. The axelera-win-toolchain-deps-installer.exe is no longer built or shipped, and "RISC-V Toolchain and Dependencies for Axelera" will no longer appear in Add/Remove Programs.
• The RISC-V toolchain has moved out of the runtime install and into the development kit.
• numpy is no longer pinned to < 2 by the installer configurations; resolution is left to the SDK wheels. The NumPy < 2 constraint documented for axelera-devkit in v1.7 therefore no longer applies at install time, and NumPy 2.x incompatibilities in the detection and segmentation evaluators have been fixed.
• New system packages are required for the Vulkan renderer: glslang-tools and libvulkan1 (installed by install-dependencies.sh).

Deprecation notice: the classic SDK Installer is deprecated as of v1.8.0 and will be removed in a future release. Installation via pip is the recommended path for both new deployments and existing environments.

Release Compatibility Matrix​

Recommended and supported board controller, flashed firmware and PCIe driver versions for each Voyager SDK release are maintained in the Release Compatibility Matrix. Consult it before installing or upgrading a card or a Voyager SDK release.

For v1.8.0 the minimum supported driver equals the shipped version - see IMPORTANT - Driver upgrade is mandatory above.

New Features / Support​

New Axelera AI Cards and Systems​

• Metis 4-chip PCIe with 8GB RAM is supported.
• Chip-down Metis support, in two variants: with a board controller, and without a board controller. For chip-down designs the LPDDR geometry (size, ranks, bus width) is read directly from GPIO pins at boot, enabling support for new Metis boards without changes to Metis firmware.

Host Platform Support​

Operating systems and kernels​

• Linux kernel support now spans 5.4 to 6.17. Kernel 5.4 support enables Yocto dunfell embedded targets. A kernel warning on 6.17 has been eliminated.
• The 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.

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

[Alpha] Model recipes for the Model Zoo​

Selected Model Zoo models now ship a standalone Python recipe under model_recipes/: a self-contained, readable script that quantizes and compiles the model, with no dependency on zoo tooling at runtime. Users can copy a recipe, point it at their own weights and calibration data, and run it directly with python. See model_recipes/README.md.

Recipes are authored and regenerated from a new Python configuration system, axzoo (shipped as the axelera-zoo wheel). This is additive - the existing YAML model zoo, deploy.py and inference.py are unchanged and remain fully supported. Most application developers can work directly from a recipe; axzoo can be used for regenerating recipes, reproducing builds, and managing model catalogs over time.

axzoo list                                   # packaged configs, grouped by task
axzoo info detection/yolov8n_coco            # display name, license, metric
axzoo build detection/yolov8n_coco           # quantize + compile to .axm
axzoo fork detection/yolov8n_coco --script   # get the standalone editable recipe
axzoo benchmark detection/yolov8n_coco --video input.mp4

Command groups: Discovery (list, check, show-config, info), Build (build, export, drift-check, check-assets), Run (benchmark, eval), Tools (fork, completion). Shell tab completion is available via axzoo completion install.

Recipes build through the axrelay backend (--backend axrelay), the established quantization and compilation path, which ingests both .onnx and .pt2.

Other notes:

• axzoo benchmark reports a latency distribution (min/mean/p50/p95/p99/p99.9/max/stddev/jitter), a per-operator breakdown, per-frame device-versus-host split, throughput with --cores N, and host CPU and peak-memory usage. --json is supported.
• .pt2 (torch.export ExportedProgram) is a new first-class source format alongside .onnx. Bare .pt state dictionaries are explicitly rejected.
• Model provenance (upstream framework, version and SPDX license) is now recorded in the .axm manifest.json.
• Named pre-compile passes replace ad-hoc YAML flags: replace_focus_layer, gemm_to_conv, disable_ceil_mode_in_avgpool, yolo26_seg_insert_proto_identity.
• Model exports run in isolated per-framework virtual environments so the SDK environment is never mutated.

The following models already present in v1.7 are now also available as recipes (built via axzoo): resnet18-imagenet, resnet50-imagenet, yolov8n-coco, yolo11l-coco, yolox-s-coco, yolo11n-seg-coco, yolo11n-pose-coco.

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 YAML model zoo is unchanged in this release. New models in v1.8 are delivered through axzoo.

[Alpha] Pipeline Builder API​

Users are advised not to use this API for workloads where performance is critical, such as benchmarks or production workloads. Pipeline Builder will reach top performance and become the recommended API for running models on Axelera hardware in the next Voyager SDK release(s).

• OpenCL (GPU) fast paths are now used automatically for op.resize, op.letterbox and op.color_convert, with transparent fallback to the CPU implementation if the GPU path is unavailable.
• The Scheduler no longer connects to hardware at construction. Connection is established lazily on first use, and new Scheduler.is_connected and Scheduler.opencl_context properties are available. A pipeline with no op.load node never touches the device - previously, merely referencing the scheduler claimed every AI core.
• Optimizer policy control: Seq.optimized() and LoadedSeq.optimized() accept a policy argument ('auto' or 'none'), with new context managers optimize_policy(policy) and optimization_disabled().
• OperatorProfiler now times leaf operators only, so nested containers no longer double-count their children.
• New ResourceMonitor context manager reports process CPU percentage, core count and peak RSS.
• Nested for_each geometry is now correctly scoped, so to_image_space inside a for_each after a resize maps coordinates back correctly.
• Image additions: decoded_timestamp, color_range and color_matrix.
• Reusable pipeline factories are available under axelera.zoo.pipelines for the YOLO detection, pose and segmentation families, YOLOX and classifiers, and a benchmarking helper under axelera.zoo.bench.

New task types and operators​

• Semantic segmentation is now a first-class task: new op.SemanticSegmentation result type, op.ax_semantic_segmentation() and op.decode_semantic_segmentation(num_classes=…), with palette-based rendering.
• Optimized int8 semantic segmentation decode using Intel AVX2 and Arm NEON with a fallback on scalar if neither is available.
• Face detection and recognition: op.decode_retina_face(), op.Embedding with op.ax_embedding(), and op.Recognition with op.match_gallery() for cosine-similarity gallery matching.
• Licence plate recognition: op.decode_lprnet() performs greedy CTC decoding to plate strings.
• op.sort_by(by=…, descending=…, top=…) orders results by a named numeric attribute, supported by new BBox.area, BBox.center_distance, Object.area and Object.center_distance properties.
• op.observer() and op.dumper() run a side-effect callback, or write each input array to disk, and pass inputs through unchanged.

Video decode and sources​

• Frame pacing: a trailing @\<fps\> or @auto on an input specification paces a file as if it were a live camera, for example ./inference.py yolov8n-coco video.mp4@30.
• New colour formats end to end: Y444 (planar YUV 4:4:4) and Y42B (planar 4:2:2), plus corrected GRAY8, RGBA, BGRA and NV16 paths.
• Colour range and colour matrix metadata are now carried per frame from the decoder, so the correct YUV-to-RGB conversion matrix is selected automatically.
• decoded_timestamp is stamped by the decoder on each frame and shares a clock epoch with Python's time.monotonic_ns() on Linux.
• The maximum number of inference children was raised from 4 to 8.

Display and rendering​

• New Surface.rectangle() layer type, App.start_thread() and App.run(interval=…).
• display.Options gained show_bounding_boxes, show_keypoints, show_segmentation, show_trajectory and show_tiles, and the label format {track_id}.
• Screen resolution is now detected via CoreGraphics on macOS rather than only xrandr.

YAML Pipeline Builder​

• New pillow_bilinear interpolation mode for Resize, matching PIL/Pillow anti-aliased downscaling. The OpenCL resize element gained a corresponding interpolation option, and the composite OpenCL preprocessing operators now use it.

• [Alpha] New Vulkan renderer, selectable with --display vulkan. Shaders are compiled on the fly and cached, and headless operation is supported with a suitable GPU. Set AXELERA_VULKAN_DEBUG=1 or 2 for validation-layer diagnostics (requires vulkan-validationlayers).

  Note: the Vulkan renderer is new in this release and is not yet at parity with --display opengl. Not yet supported: multi-resolution layout, frame style settings, the grayscale option, image-overlay and heatmap metadata types, the startup logo and progress bar, and window resizing.

Demo scripts and examples​

All examples under examples/pipeline_builder/ gained a shared --backend {ffmpeg,opencv} option.

Note: semantic_segmentation.py requires yolo26n-sem.axm, which is not in the public pre-compiled catalog. Build it locally with download_axm.py --deploy-missing.

RISC-V toolchain unification​

• LLVM 20 bare-metal toolchain for RISC-V replaces the previous GNU newlib toolchain. As a result of this the wheel shrinks from ~200 MiB to ~20 MiB (~900 MiB less on disk), with up to ~10% inference performance improvement, especially on small, control-bound networks.

Tools​

• axddr (new) is a high-level tool for LPDDR verification, usable both on Axelera boards and by chip-down customers to facilitate and speed up LPDDR design validation. It performs LPDDR memory validation over PCIe, with no JTAG or UART required; it loads the stage0 firmware on demand. Tests available: sequential, random, aliasing, databus. axddr -i prints device name, DDR size, rank, bus width and firmware versions without running a test.
• axrunmodel: reworked with a low-latency executor, outputs verification and statistics; resources are auto-selected from the model's L2 footprint; new transient --set-power-limit and --set-mvm-limit overrides.
• axmonitor:
  • New post-processing mode: axmonitor --post-proc <file.jsonl> generates per-device SVG plots (temperature, power, KPS, DDR bandwidth, PCIe bandwidth, CPU utilisation), a temperature-versus-KPS plot, an average-power-versus-KPS plot, and a Markdown summary.
  • Power-during-inference metrics are included in the post-processing report.
• axcmd: new --get-mvm-limits \<aicore\> reports the user-set, power-control-loop, service-monitor and effective MVM limits separately, so it is clear why MVM utilisation is being capped. --get-ddr-size now works on chip-down boards.
• Power limit configuration now persists across device resets.
• tools/tile_config.py: actionable install hints when wx or cv2 are missing, and corrected control sizing.

Packaging​

Newly shipped wheels:

Removed: axelera-riscv-gnu-newlib-toolchain, the axelera-riscv-openocd Debian package, and axelera-win-toolchain-deps-installer.exe.

Breaking Changes​

1. AxModel major version 4 to 5 - models must be rebuilt. AxModels built with v1.7 or earlier cannot be run or inspected. See above.
2. Minimum supported PCIe driver raised to the shipped version: Linux 1.6.2 and Windows 1.3.14. Older drivers are rejected.
3. Linux device node names changed from colons to hyphens, for example metis-0:1:0 becomes metis-0-1-0. udev backwards-compatibility symlinks preserve the old form, but scripts that glob or parse the colon form should be reviewed.
4. Metis Compute Board (AISBC) running BSP 1.3.3 requires a patch for SDK v1.8.0. This consists of updating the Metis kernel module to 1.6.2 and applying a .deb package that symlinks the device node name change described above. Download both packages from the Axelera software portal: the Metis kernel module 1.6.2 package and the name-change symlink package. To apply:

su root                                                            # root password for the AISBC BSP image
mount -o remount,rw /
dpkg -i kernel-module-metis-6.1.148-rockchip-standard_v1.8.0-r0_arm64.deb   # Metis kernel upgrade
dpkg -i axelera-container_1.0-r0_arm64.deb                                  # device name change symlink
mount -o remount,ro /                                              # if this fails, simply run `reboot`
reboot

Fixed Issues Since v1.7.0​

• Fixed a chip inference abort on models with hand-written hardsigmoid or hardswish, for example MobileNetV3-Small, which previously failed with "fewer padding configurations than tensors".
• Fixed an intermittent crash when streaming with more than one core, caused by a detached decoder thread using freed contexts.
• Fixed color_convert() crashing on odd-width of YUV input formats.
• Fixed the last few frames of long videos being dropped, and made the progress display respect --frames.
• Fixed op.top_k crashing on standard classification head shapes.
• Fixed NumPy 2.x incompatibilities in the detection and segmentation evaluators.
• Fixed a Darknet model's .cfg path not being expanded from its portable form.
• Fixed the Level Zero out-of-memory message, which now names the memory region and reports human-readable sizes and totals.
• Fixed kernel 5.4 build failures and a kernel warning on 6.17.

Known Issues and Limitations​

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​

• The axzoo documentation is currently shipped with the axelera-zoo package rather than being part of the documentation portal.
• Performance variability is observed on certain hosts. Inference-only performance (FPS) drops of up to 5-10% are observed on yolox-x-crowdhuman-onnx and yolo11l-obb-dotav1-onnx compared to SDK Release v1.5.
• Python 3.13 incompatibility with the 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.
• Axrunmodel now benchmarks with random rather than constant input data (needed for output verification). Realistic data draws more power, so reported throughput on power-limited devices is lower than previous releases; device performance is unchanged.

System Requirement​

Development Environment​

For model compiling purposes, these are the host requirements:

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 the Axelera AI Community.
• For technical documents and guides please visit docs.axelera.ai.