Summary
Requesting support for the NVIDIA GB10 (Grace Blackwell Superchip, as found in the DGX Spark) to be added to SUPPORTED_GPUS in api/gpu.py, with any necessary adjustments to GraVal verification to accommodate unified memory architecture.
This is similar in nature to #34 (A10G RT core mismatch) — a hardware variant that falls outside the current validation parameters, not a fundamentally unsupported class of GPU.
Hardware
Device: NVIDIA DGX Spark (GB10 Grace Blackwell Superchip)
Memory: 128GB unified LPDDR5X — CPU and GPU share a single pool, no discrete VRAM
CPU: 20-core NVIDIA Grace (aarch64)
CUDA: 12.8+
OS: Ubuntu 22.04 (aarch64)
Blockers
- GB10 not in SUPPORTED_GPUS
Not listed in api/gpu.py. First gate.
- GraVal VRAM check fails on unified memory
GraVal requires 95% of VRAM free, seeded by discrete device info. The GB10 has no discrete VRAM — the full 128GB is a unified pool. nvidia-smi reports this differently than a traditional discrete GPU, so the availability check likely returns 0 or errors out entirely.
The fix would need to either detect unified memory architecture and skip/adapt the VRAM check, or define "95% available" in terms of the full 128GB pool when no processes are running.
- Ansible/K3s assumes x86_64
The provisioning stack would need aarch64-aware defaults for the GB10's ARM CPU.
Why It's Worth Adding
The B200 and B300 are already on your supported list — same Blackwell generation. The GB10 is the consumer-class sibling of that line. Its 128GB unified memory lets it serve 70B+ inference endpoints that would otherwise require dual A100s or H100s, which would meaningfully increase chute diversity on the subnet.
The DGX Spark shipped widely to developers in early 2026. There's likely a pool of GB10 owners blocked by this that would add capacity if the path existed.
Questions
Is GB10 support on the roadmap given B200/B300 are already there?
Is GraVal open source? Happy to submit a PR for unified memory detection if so.
Is there a testnet path to validate GB10 compatibility before mainnet registration?
Happy to provide a test node if the team wants one.
Summary
Requesting support for the NVIDIA GB10 (Grace Blackwell Superchip, as found in the DGX Spark) to be added to SUPPORTED_GPUS in api/gpu.py, with any necessary adjustments to GraVal verification to accommodate unified memory architecture.
This is similar in nature to #34 (A10G RT core mismatch) — a hardware variant that falls outside the current validation parameters, not a fundamentally unsupported class of GPU.
Hardware
Device: NVIDIA DGX Spark (GB10 Grace Blackwell Superchip)
Memory: 128GB unified LPDDR5X — CPU and GPU share a single pool, no discrete VRAM
CPU: 20-core NVIDIA Grace (aarch64)
CUDA: 12.8+
OS: Ubuntu 22.04 (aarch64)
Blockers
Not listed in api/gpu.py. First gate.
GraVal requires 95% of VRAM free, seeded by discrete device info. The GB10 has no discrete VRAM — the full 128GB is a unified pool. nvidia-smi reports this differently than a traditional discrete GPU, so the availability check likely returns 0 or errors out entirely.
The fix would need to either detect unified memory architecture and skip/adapt the VRAM check, or define "95% available" in terms of the full 128GB pool when no processes are running.
The provisioning stack would need aarch64-aware defaults for the GB10's ARM CPU.
Why It's Worth Adding
The B200 and B300 are already on your supported list — same Blackwell generation. The GB10 is the consumer-class sibling of that line. Its 128GB unified memory lets it serve 70B+ inference endpoints that would otherwise require dual A100s or H100s, which would meaningfully increase chute diversity on the subnet.
The DGX Spark shipped widely to developers in early 2026. There's likely a pool of GB10 owners blocked by this that would add capacity if the path existed.
Questions
Is GB10 support on the roadmap given B200/B300 are already there?
Is GraVal open source? Happy to submit a PR for unified memory detection if so.
Is there a testnet path to validate GB10 compatibility before mainnet registration?
Happy to provide a test node if the team wants one.