DGX Station GB300: current configuration
This is the living software and firmware snapshot of the deskside Station, not a tok/s post. The numbers below were read off the box after tonight's maintenance reboot. Recipes live elsewhere: DSF Vision-Exp Recipe v2, GLM-5.3-Flash round 3, model testing summary.
Kernel
7.0.0-1019-nvidia-64k (#19~24.04.2, Sep 6 2026). Matching open modules installed before reboot.
Driver
595.91.07 from Ubuntu noble-updates. CUDA 13.2 reported by nvidia-smi. Was 595.84.
UEFI
AMI ES429NMS.103, dated May 22, 2026. Not flashed tonight.
CDMM
NVreg_CoherentGPUMemoryMode=driver. MemTotal 518,524,160 kB (494.5 GiB). That is the expected post-CDMM host view, not a missing-RAM fault.
What the box is
| Item | As measured September 16, 2026 |
|---|---|
| Chassis | Exxact-built MSI XpertStation WS300 — DMI product XpertStation WS300 - Station GB300, serial Q6T0005673, hostname sn4622130625 |
| Superchip | NVIDIA GB300. nvidia-smi reports 256,703 MiB HBM (250.7 GiB). Spec is 288 GB HBM3e. Do not round the MiB reading to a 256-gigabyte headline. |
| Display GPU | RTX PRO 4000 Blackwell SFF, 24,467 MiB, VBIOS 98.03.AA.00.02. Console only. Do not add a 6000 for inference — it statically caps Superchip power. |
| GB300 VBIOS | 97.10.77.00.0B |
| Grace host RAM | 494.5 GiB visible with CDMM=driver. Without CDMM this reads closer to ~780 GB and serving math goes wrong. |
| OS | Ubuntu 24.04.5 LTS, arm64. Apt-updated DGX OS 7 lineage. No /etc/dgx-release on this image. |
| OS volume | md0 RAID1, 1.9 TB, 473 G used (27%) |
| Model cache | md1 RAID0 at /models, 15 T as df reports it, 2.8 T used (19%). How that array was built: empty CX8 M.2s. |
| BMC (IPMI) | ipmitool mc info Firmware Revision 0.08, IPMI 2.0, manufacturer NVIDIA. That is the IPMI field, not a Galaxy-GB300-WS package string. |
Tonight's update
Ubuntu had 162 pending packages. A plain apt-get upgrade took the language packs and GNOME bits and left the kernel and NVIDIA driver on the old versions. Phased updates. The packages that actually matter had to be named:
sudo apt-get install nvidia-driver-595-open linux-nvidia-64k-hwe-24.04That pulled driver 595.91.07, kernel 7.0.0-1019-nvidia-64k, and — the trap from day one — linux-modules-nvidia-595-open-7.0.0-1019-nvidia-64k (7.0.0-1019.19~24.04.2+1). We checked that package was ii before rebooting. SSH was back in about four minutes. nvidia-smi on both GPUs, matching driver string, CDMM still driver-mode.
595.91.07 (August 3, 2026) is the current 595-branch data-center driver. Release notes call out Fabric Manager, gRPC, CUDA-graph Xid 32, and channel cleanup. This Station is one B300, not an 8-GPU HGX, so most of that is hygiene, not a rescue. The reason to take it is staying on the apt track so the next 7.x kernel does not land on a stale 595.84.
The kernel-module trap is still the one that bricks first boot. After any apt that moves linux-image-nvidia-64k-hwe-24.04, confirm linux-modules-nvidia-595-open-$(uname -r) exists for the next kernel before you reboot. Day one booted 1031 without 1031 modules. Tonight we refused to repeat that.
Firmware we did not touch
nvfwupd is not on this box and is not in Ubuntu apt. Station firmware (UEFI, BMC, HMC, GPU VBIOS, CX8) ships as Galaxy-GB300-WS .fwpkg bundles and needs a power cycle or Redfish activation. We read versions only:
| Component | Source | Value | Action |
|---|---|---|---|
| UEFI | dmidecode | AMI ES429NMS.103 (May 22, 2026) | Leave |
| GB300 VBIOS | nvidia-smi | 97.10.77.00.0B | Leave |
| RTX PRO 4000 VBIOS | nvidia-smi | 98.03.AA.00.02 | Leave |
| BMC | IPMI mc info | Firmware Revision 0.08 | Leave |
No thermal, link, or BMC symptom to justify a flash. A later window can install nvfwupd, run show_version against a current Galaxy-GB300-WS package, and decide. Until then, do not treat the IPMI 0.08 field as an NVIDIA package revision.
DGX OS 7.6.0 exists (September 14, 2026). We did not reimage. Apt is the path on this Station. DGX OS 8 is documented with GB300 limitations; not tonight.
What the reboot did to serving
The running experiment container glmf-w8-moe-trtllm had been up 25 hours on :30001 with Docker restart policy no. The reboot stopped it. It did not come back, which is the correct default. This page is not a restore authorization. Production family traffic is still the :30003 DSF contract when that lane is lit; tonight the box is idle GPU after maintenance.
Standing rules this snapshot is written against
- Stage checkpoints on
/models(local RAID0). Do not serve from CIFS. - HBM-only serving on this Station: pick models that fit 250.7 GiB visible HBM plus KV. Grace is staging, not the daily offload path, unless a campaign says otherwise.
- Display card stays the 4000 SFF.
- Unscheduled swaps on family
:30003are forbidden. A scheduled reboot is not a restore.
Related: day one, M.2 /models, GB300 topic index. I am a wannabe inference engineer. This post is the configuration, not a benchmark.