OLARESOS ON SPARK.
THE ROUTE TO A
WORKING RESULT.
A real installation, a specific driver failure, an assisted recovery, and application-owned AI workloads. The record keeps the successful outputs and the unfinished parts together.
I tried OlaresOS because I already knew the workspace on other hardware. Spark offered a new place to explore it. The useful outcome was a working path through the actual platform, with enough evidence to explain how we got there.
One machine. Successive installations.
The same physical DGX Spark first ran its factory DGX OS, then a clean OlaresOS ARM64 installation. The recorded Olares stack used kernel 6.14.0-1015-nvidia and NVIDIA 595.84 open ARM64 modules.
The DGX OS and Olares runs used different models, software versions and image workflows. The results answer different acceptance questions. They do not show that one OS is faster than the other.
Secure Boot → incomplete driver → missing NVML.
The kernel rejected an unsigned NVIDIA module. Driver installation aborted, NVML was absent, and Olares waited at InstallPlugin → CheckGPUState. The logs described a failed installation path; they did not establish that GB10 was unsupported.
I disabled Secure Boot in UEFI for this lab experiment. The operator completed matching driver modules and userspace libraries, checked CUDA/NVML and used the supported GPU-enable path. Disabling Secure Boot by itself would not have restored missing libraries.
The incomplete Olares installation phase was subsequently reset and recreated without reinstalling the base OS. Activation included a transient TLS error that later cleared; its root cause was not established. These are recorded lab actions, not a general recovery prescription for an existing machine with data.
A working answer. A working image.
A standalone CUDA test Pod was rejected because it had no application owner. That policy was retained. A normal Olares Market-owned Ollama application later received GPU access and ran inference.
| Workload | Observed outcome | Boundary |
|---|---|---|
| Qwen2.5 0.5B Q4_K_M | 25/25 layers on GPU; warm 42-token answer; OpenAI-compatible endpoint worked | One short acceptance scenario; five bullets returned when three were requested |
| ComfyUI / SD 1.5 FP16 | 512 × 512 image, 20 steps; 9.41 s application time | Core workflow passed; colours/materials did not fully follow the prompt |
| Optional node packages | Twelve import failures | The tested core graph did not depend on them |
| Inspected FreeToken package | Its manifest/container was amd64-only | Result applies to that package, not every possible ARM implementation |
A 120B model was demonstrated during the earlier DGX OS phase. It was not benchmarked under Olares in this run.
Assisted operation, with visible boundaries.
I handled physical setup, firmware, activation and approvals. The AI operator inspected logs, helped complete repairs and checked application outputs. Olares browser Desktop and a Windows ARM64 guest supplied useful interfaces; that is not a claim of NVIDIA GPU passthrough to Windows.
The most useful product feedback follows the failure path: check signing and ARM64 application support earlier; expose the failing dependency; explain what retry/reset preserves; distinguish management, shell and actual desktop access.
KEEP THE METHOD AND THE MISSING PIECES VISIBLE.
The Olares acceptance package retains responses, selected workload metadata, image history and output. Its manifest covered 63 files, all verified in the publication package; the manifest itself is the 64th file.
Some earlier DGX OS raw timing bundles are not included; their results remain report-level evidence. No controlled same-model comparison, long-term reliability result or multi-user delegation test is claimed.
Full earlier page, including all workload figures and related notes →