3.5 GiB per node is firmware-owned; asking NVIDIA for a headless option
research
hardware
memory
The remaining reserved memory is the display/GSP carveout and firmware regions no OS change can return, and NVIDIA’s own release notes show the carveout is BIOS-adjustable upward, so the request is a smaller setting.
Question. After the OS-side campaign, what remains, and who can return it?
| reclaimed per node so far | |
|---|---|
| 64K pages (memmap) | 2.10 GiB |
| kernel rebuild, no random kmalloc caches | 0.74 GiB |
| idle NIC PF unbind | 0.42 GiB |
| OS strip | 0.2 GiB |
| remains, firmware-owned | |
|---|---|
display/GSP carveout at 0x280000000 |
2,616 MiB |
| 15 smaller nomap regions | 1,018 MiB |
| GSP boot heap (largest single nvidia stack in page_owner) | 512 MiB |
The DGX Spark release notes list Display Reserved Memory as BIOS-adjustable between 2 GB and 4 GB. The control exists; only a larger option shipped. A headless cluster node uses neither.
The ask, going to the DGX Spark / GB10 developer forum: a 0 GB or minimal display carveout setting for headless and appliance use. Per node that is up to 2.55 GiB; across this pair, 5.1 GiB. Firmware here is GX10DGX.0105.2026.0505.1153.
Verdict. Open. Waiting on NVIDIA.
Reopen if.
- a firmware release changes any carveout size, or the forum thread answers what the smaller regions hold