NVIDIA DGX Spark
NVIDIA DGX Spark uses the GB10 Grace Blackwell platform. It is an ARM64 UEFI/SBSA
system, not a Jetson/L4T device. Build its Kairos image with the
nvidia-dgx-spark model.
Prerequisites​
- An NVIDIA DGX Spark.
- Docker with
buildx. - A
kairos-initrelease that includes DGX Spark support. - An ARM64 build host or QEMU support for ARM64 cross-builds.
Build the image​
Create this Dockerfile:
FROM quay.io/kairos/kairos-init:v0.17.2 AS kairos-init
FROM ubuntu:24.04
ARG VERSION=1.0.0
RUN \
/kairos-init --version "${VERSION}" --model nvidia-dgx-spark
Build and push the ARM64 image:
docker buildx build --platform linux/arm64 \
-t my-registry.example.com/kairos-dgx-spark:v1.0.0 \
--push .
On an ARM64 host, you can use docker build without the --platform option.
What the model configures​
The nvidia-dgx-spark model configures these platform components:
- The NVIDIA HWE kernel for Ubuntu 24.04.
- The open NVIDIA 580 kernel modules and headless driver.
- NVIDIA DGX BaseOS and Spark package repositories.
- NVIDIA and Mellanox firmware and platform tools.
- The serial console and kernel arguments that DGX Spark requires.
fwupdfor UEFI capsule updates through ESRT.
The model does not add the CUDA repository. Add NVIDIA's ARM64 SBSA CUDA repository in a derived image if you need host CUDA packages.
Create installable artifacts​
Create an output directory:
mkdir -p build
Run AuroraBoot with the image that you pushed:
docker run --rm -v "$PWD/build:/output" \
quay.io/kairos/auroraboot:v0.26.2 \
build-iso --output /output/ \
oci:my-registry.example.com/kairos-dgx-spark:v1.0.0
AuroraBoot writes the installable ISO and its checksum to the build directory.
See the AuroraBoot reference for other artifact
types and configuration options.
DGX Spark uses the standard Kairos installation and upgrade flow because it boots with UEFI.
You can use providers, Trusted Boot, and stage extensions with this model. See the Kairos Factory reference for the available options.