- Jira: AK-7557
- Testers: Ali Kayyam, Douglas McLelland (requested by Tharak Krishnan)
- Built by: Tejaswini Dasika, Kevin Tsiknos
- Repo:
brainchipSW/akida_docker/tree/main/metatf_public_images - Scope: Docker images with DevHub pre-installed, plus the README and user guide
- Validate the container install and startup experience.
- Verify DevHub is available and works inside the container.
- Run representative Akida development workflows.
- Review the README and user guide for completeness and accuracy.
- Find usability issues, bugs, missing dependencies and documentation gaps.
- Send consolidated feedback to the dev team and REQ-TOOL-005 stakeholders.
| # | Criterion | Status |
|---|---|---|
| 1 | Image pulled and launched successfully | Not started |
| 2 | DevHub accessible and functional | Not started |
| 3 | Akida workflow runs with no blocking issues | Not started |
| 4 | Documentation reviewed | Not started |
| 5 | Defects and enhancement requests documented | Not started |
| 6 | Feedback shared with REQ-TOOL-005 stakeholders | Not started |
Checked read-only on 2026-10-14:
| Item | Value |
|---|---|
| GPUs | 2x NVIDIA A30 (24 GB each), idle |
| Driver / CUDA | 595.71.05 / CUDA 13.2 |
| Docker | 28.1.1, nvidia runtime registered, default runtime runc |
| NVIDIA container toolkit | nvidia-ctk installed |
| Disk / RAM | 276 GB free / 251 GB |
| Akida device | None seen on this host (to confirm) |
Use --gpus all explicitly, because the default runtime is runc.
Needed for the PCIe passthrough and hardware inference tests. Until one is available, the hardware criteria are reported as not tested, not as passed.
- Registry:
bender:5000(internal in-house registry). Not on Docker Hub yet. - Images:
bender:5000/akida-devhub:<tag>, with tagsubuntu{22.04,24.04}-{gpu,cpu}-x86andubuntu{22.04,24.04}-cpu-aarch64. - Driver installer:
bender:5000/akida-pcie-driver-installer:ubuntu{22.04,24.04}. This is a one-time install per host and needs Secure Boot disabled. - DevHub:
brainchip_devhubis baked into the image at/workspace/brainchip_devhub, pinned to commit8be7c232. The images were re-pushed on 2026-10-14 under the same tags, and PR #45 was merged the same day. Make sure the pulled image is the new one. - Docs:
metatf_public_images/akida_dev_images/README.md(build) andmetatf_public_images/USER_GUIDE.md(use), both in the internal repogit.corp.brainchipinc.com/brainchipSW/akida_docker. - Stated beta scope: AKD1500 PCIe/M.2 and the Akida 1 model zoo examples only. Akida 2 is not in the beta scope.
- Where to report: notebook issues go in subtask AK-7601. Everything else goes in AK-7557.
- Known issues, so don't report them as new:
- The eDMA driver is built from a private branch, and AK-7565 is still In Progress.
- Secure Boot must be disabled.
- Alpine was rejected, because TensorFlow has no musl wheels.
- Image sizes are about 4.7 to 6 GB for CPU and 10.5 GB for GPU. Tharak raised size as a concern.
- Earlier GPU validation: Tejaswini tested GPU training on an RTX 2080 ("Hawkeye"). The A30s here (Ampere) are new coverage.
- Documented commands:
docker pull bender:5000/akida-devhub:ubuntu24.04-gpu-x86 docker run -it --gpus all --privileged -v /dev:/dev \ -p 8888:8888 -v "$(pwd)/workspace:/workspace" \ bender:5000/akida-devhub:ubuntu24.04-gpu-x86 bash
- Registry reachability (checked 2026-10-14):
benderdoes not resolve from this host. Docker already trusts172.17.0.39:5000as an insecure registry. It answers, and it holdsakida-devhubandakida-pcie-driver-installerwith every tag listed in the Jira, plus2.19.3-prefixed copies of each (those are internal CI images, out of scope). Likely bender, but unconfirmed. - Confirm that
172.17.0.39:5000is bender (ask Tejaswini or Kevin). If it is, use it in place ofbender:5000in all commands. -
git.corp.brainchipinc.comresolves and answers (HTTPS 302 to login). Reading it needs credentials, which are present on this host, so Ali must approve their use. Otherwise Ali exports the README and USER_GUIDE. - Akida device: none on this host (no PCI device, no kernel module). Secure Boot is already disabled here. The beta scope is AKD1500, so the hardware phases need a PCIe or M.2 card on another machine.
- Which part Douglas is covering.
- Report format and destination: Jira comment on AK-7557, or markdown or Word.
- Jira's input spec says "Ubuntu 20.04/22.04", but the images are 22.04/24.04. Kevin said 20.04 support was dropped, so the spec text is stale.
- The documented
docker runuses--privileged -v /dev:/dev. This is broad, so check whether narrower--deviceflags work and whether the guide explains why. - The PR table lists five validated combinations, but six tags are on
bender. Check thatubuntu22.04-gpu-x86was validated.
- Read AK-7557 and its comments for the image name, tag and any known issues.
- Read the README and user guide in the repo path above.
- Confirm host details:
docker info,nvidia-smi,lspci, disk and RAM.
Separate host setup problems from image problems.
- Run
nvidia-smiin a small public CUDA image with--gpus all. - Confirm both A30s are visible, and that
--gpus device=0and--gpus device=1each select one GPU.
- Follow the USER_GUIDE exactly as written. Log every command and every deviation.
- Pull
bender:5000/akida-devhub:ubuntu24.04-gpu-x86first, thenubuntu22.04-gpu-x86. Record the image size, pull time and image ID, and confirm that/workspace/brainchip_devhubis at commit8be7c232. - Launch the container and check the entrypoint, default user, volume mounts and port mapping.
- Launch with
--gpus alland check thatnvidia-smiworks inside it.
- Check that DevHub is installed, starts and is reachable.
- Check login or setup, if the flow has one.
- Exercise the functions the guide describes, such as model and dataset access.
- Record any missing dependencies or unclear first-run steps.
- Record versions: Akida, MetaTF, cnn2snn, quantizeml, TensorFlow/Keras, Python and CUDA/cuDNN.
- Check that TensorFlow sees the GPUs with
tf.config.list_physical_devices('GPU'). Compare the CUDA and cuDNN versions in the image with the host (CUDA 13.2), because a mismatch is a likely finding. - Run a representative workflow: build or load a model, quantize, convert and run inference on the virtual device.
- Run a short GPU training or quantization job. Watch for errors and memory use.
- Run the shipped examples and tutorials. Note failures and missing dependencies.
- Repeat on GPU 0, GPU 1 and both together.
- Follow the README device instructions:
--device, privileges, drivers and udev rules. - Check that the container enumerates the device.
- Run the same inference workflow on hardware and compare it with the virtual device.
- Restart the container and check that data in mounted volumes persists.
- Run a longer job and watch CPU, RAM and GPU memory for leaks or crashes.
- Record image size, startup time and any slow steps.
- Cross-platform (Windows, macOS): mark not tested unless a machine is available.
Compare the README and guide with what actually happened:
- Steps that failed or were missing
- Wrong versions or commands
- Unstated prerequisites (host driver, Docker version, GPU runtime, ports)
- Unclear wording
- Record everything in report.md: the acceptance table, coverage, findings, documentation review and test log. Update it as testing goes, not only at the end.
- Draft the feedback for AK-7557 and the stakeholders from report.md.
- Review the draft with Ali before anything is posted or sent.
Record each defect as:
ID: D-001
Severity: Blocker | Major | Minor | Cosmetic
Area: Startup | DevHub | SDK | Examples | Docs | GPU | Hardware
Environment: host, Docker version, GPU/device, image tag
Steps: 1. ... 2. ...
Expected:
Actual: (include log or error text)
Suggested fix: (optional)
Enhancement requests use the same format with Severity: Enhancement.
| Area | Result | Environment | Notes |
|---|---|---|---|
| Startup | |||
| DevHub | |||
| SDK availability | |||
| Examples and tutorials | |||
| GPU | |||
| Akida hardware | |||
| Documentation | |||
| Stability / performance | |||
| Cross-platform |
- Read-only checks run freely. Pulling images and running containers on this host needs Ali's go-ahead.
- Nothing is posted to Jira or sent by email without Ali's explicit approval.
- Anything not tested is reported as not tested, never as passed.