ubuntu-gb10
Ubuntu 24.04 + NVIDIA Stack setup guide for GB10 / DGX Spark systems. Documents full GPU server configuration on ARM64 Grace Blackwell architecture.
Ficha actualizada 22 sept 2026, 21:24 · upstream 22 sept 2026, 21:15 · release sin release detectada
Qué es ubuntu-gb10
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
This repository documents the complete setup of an NVIDIA GB10 Grace Blackwell system as a clean Ubuntu 24.04 GPU server, modeled after DGX OS 7 but running standard Ubuntu Server without NVIDIA's customized DGX OS image. The guide applies to all GB10-based partner systems including ASUS Ascent GX10, NVIDIA DGX Spark, Lenovo ThinkStation GB10, Dell GB10 workstations, and other GB10 OEM units.
The GB10 is an ARM64 platform featuring the NVIDIA Blackwell Superchip SoC with 5th-gen Tensor Cores supporting FP4 (1 petaFLOP AI), 128 GB unified LPDDR5x memory shared between CPU and GPU via NVLink-C2C, a ConnectX-7 NIC, and 10 GbE LAN. All packages, ISOs, and repository URLs must use the arm64 variant.
The guide covers Ubuntu Server 24.04 LTS installation on ARM64, NVIDIA driver stack including Open kernel modules, CUDA Toolkit 13.0, HWE kernel 6.17, Docker CE with NVIDIA Container Toolkit, optional ConnectX-7 DOCA-OFED drivers, verification steps, performance tuning, and dual-node NCCL interconnect configuration.
- Run AI inference workloads on the GB10's FP4 Tensor Cores
- Deploy machine learning models using CUDA-enabled frameworks
- Build and serve web applications with Next.js on ARM64
- Process scientific data with high-performance GPU computing
- Run containerized microservices via Docker CE + NVIDIA Container Toolkit
- Develop edge AI applications leveraging the 10 GbE ConnectX-7 NIC
- ARM64 architecture (Grace Blackwell) with unified memory shared between CPU and GPU
- NVIDIA CUDA Toolkit 13.0 with Open kernel modules for full GPU access
- ConnectX-7 NIC supports DOCA-OFED drivers for high-performance networking
- Supports both single-node and dual-node NCCL interconnect configurations
- Clean Ubuntu base image avoids NVIDIA's proprietary DGX OS customization
- 128 GB unified memory (CPU+GPU shared) via NVLink-C2C
- Requires ARM64 hardware; x86_64 systems are not supported
- ConnectX-7 NIC requires DOCA-OFED drivers for full performance
- Ubuntu Server 24.04 LTS may have limited GPU driver support compared to DGX OS
- 128 GB unified memory is shared between CPU and GPU, limiting per-GPU resources
- Dual-node NCCL setup requires additional configuration beyond the base guide
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Install Ubuntu Server 24.04 (arm64) first, then add NVIDIA repos, install HWE kernel 6.17 with NVIDIA modules, configure CUDA Toolkit 13.0, and install Docker CE + NVIDIA Container Toolkit. For ConnectX-7 support, install DOCA-OFED drivers. Verify via `nvidia-smi` and `docker run --gpus all`. For dual-node NCCL, use the provided playbook roles or manual configuration of the NCCL driver. Always reboot after kernel changes to apply Secure Boot MOK enrollment.
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.