- Nvidia’s new GB10‑based edge AI boxes still give developers zero insight into CPU power consumption.
The study audited an ASUS Ascent GX10 desktop built on Nvidia’s GB10 SoC. The platform provides only instantaneous GPU power via NVML; there is no CPU energy counter, no INA rail monitor, no IPMI/BMC access, and no SCMI power‑cap protocol exposed to software. Nvidia’s own firmware does compute per‑rail energy through an undocumented ACPI interface called SPBM, but the company has said it has “no plans to expose CPU rail information.” Because of this gap, the per‑process energy attribution that tools like RAPL provide on x86 cannot be reproduced on the GB10 edge devices.
The omission matters because agentic AI workloads—multi‑step tasks that call tools, retry, and recover from failures—use far more energy than linear models, with the research citing 4.33× higher energy per successful goal and up to 7.63× more for complex reasoning. CPU‑side processing alone accounts for about 44% of dynamic energy and over 90% of latency in these workloads. Without visibility into that slice of power draw, developers cannot measure, optimize, or certify low‑carbon operation.
The authors propose a temporary calibration bridge that works on platforms like the Acer Veriton GN100, where CPU energy accumulators are exposed, and point to the SCMI power‑cap standard as a long‑term solution. Until hardware vendors make CPU telemetry a first‑class feature, low‑carbon AI on the edge will remain a blind spot.
The situation echoes earlier desktop GPU‑only monitoring tools, which forced engineers to estimate system power by subtracting a fixed CPU baseline. Nvidia’s refusal to open the CPU rail data suggests the same legacy workaround will persist for edge AI boxes, at least for now.