// TECHNICAL NOTE · GB10 THERMAL BEHAVIOR
The NVIDIA DGX Spark uses the GB10 Superchip — Blackwell architecture, compute capability sm_121 (or sm_90 in some earlier CUDA toolchains). Under sustained inference or training workloads, GB10 sits in the low-80s °C indefinitely. The reason is not a hardware defect; it is a design choice. The cooling levers a typical NVIDIA GPU exposes are absent on this silicon.
Both power-limit and fan control return [N/A] from
nvidia-smi:
$ nvidia-smi --query-gpu=power.limit,power.max_limit,power.min_limit,fan.speed --format=csv,noheader [N/A], [N/A], [N/A], [N/A]
These are not "hidden" or gated behind sudo; they are not implemented on the GB10 driver interface. The firmware manages both dimensions internally based on thermal and workload state. No amount of userspace configuration changes them.
The one lever that does respond is the graphics clock ceiling.
Even though nvidia-smi --query-supported-clocks=graphics
returns [N/A] on GB10, the driver still accepts arbitrary
integer MHz values in the silicon's operating range through
--lock-gpu-clocks:
$ sudo nvidia-smi -lgc 1500,2000 -i 0 GPU clocks set to "(gpuClkMin 1500, gpuClkMax 2000)" for GPU 0000000F:01:00.0 All done.
The reference upper clock on our production hardware is ~3003 MHz. The
minimum accepted seems to be around 900 MHz. Steps between are honored.
-rgc resets to stock. Both commands take effect immediately.
Because the firmware alone decides when to clock down, and it does so conservatively (avoiding SW-slowdown, not chasing an arbitrary temperature target), a DGX Spark serving Ollama at 90+% utilization will settle into the low-80s °C and stay there indefinitely. The user cannot ask the firmware to be more aggressive. The only remaining actuator is the graphics clock ceiling.
We built a small daemon that samples GPU temperature every 30 seconds
and steps the clock ceiling by 150 MHz using three-band hysteresis:
step down when the temperature enters a warning band, hold
in the target band, step up after three consecutive cool samples.
On our production GB10 (spark-23), sustained temperature dropped by 11 °C
at the same 94% utilization. The full mechanism and the license terms are
at the product page: thermal.zctechnologies.org. Live
telemetry: /live-telemetry.html.
Reference on the -lgc flag's actual behavior on GB10:
/nvidia-smi-lgc-on-gb10.html.
Published by C.R. Burrell LLC. Written for engineers running DGX Sparks in production. Corrections welcome: [email protected].