Its been two weeks of non-stop problems with the Asus GX10 and now I am just probably sending it back to the store. One last attempt… I am seeing serious periodic slow downs every 30 minutes or so that last for 5-10 minutes and then performance returns. I have the latest patches and firmware updates. I do not have anything else plugged into the system. With AI assistance, I created a short monitoring process and it found the below. Is there any known solution to this problem (seems somewhat related to the 30w cap bug, but its intermittent).
Issue summary: ASUS Ascent GX10 / NVIDIA GB10 intermittent GPU low-power state during inference
System: ASUS Ascent GX10, NVIDIA GB10 platform
Workload: Local LLM inference / token generation
Symptom: Approximately every ~30 minutes, token generation slows dramatically, down to roughly 0.1 tokens/sec. The slow period lasts from tens of seconds to several minutes, then performance recovers without reboot.
This does not appear to be normal thermal throttling or a user-space inference/runtime issue. During the slow windows, the GPU remains heavily utilized, but GPU power and graphics clocks collapse to a low fixed range.
Primary observed failure signature
During the slow windows, repeated nvidia-smi telemetry shows:
GPU utilization: ~94-96%
GPU power draw: ~15-18 W
GPU graphics clock: ~630-650 MHz
GPU temperature: ~44-52 C
Performance state: P0
SW Power Cap: Not Active
HW Slowdown: Not Active
HW Thermal Slowdown: Not Active
HW Power Brake Slowdown: Not Active
SW Thermal Slowdown: Not Active
SW Power Capping: 0 us
HW Thermal Slowdown: 0 us
Representative log sample from 2026-05-24T17:18:34-05:00:
power.draw: 17.69 W
clocks.current.graphics: 650 MHz
utilization.gpu: 96%
temperature.gpu: 44 C
pstate: P0
SW Power Cap: Not Active
HW Slowdown: Not Active
HW Thermal Slowdown: Not Active
HW Power Brake Slowdown: Not Active
SW Thermal Slowdown: Not Active
Average Power Draw: 17.57 W
Graphics current clock: 643 MHz
Max/available graphics refs: 2418 MHz / 3003 MHz
The same sample shows no memory, CPU, I/O, or kernel fault pressure at the time: memory available was ~29 GiB, swap was essentially idle, CPU pressure was 0.00, I/O pressure was 0.00, and recent kernel signals were empty.
Representative repeated slow-window samples
The condition persists across multiple samples in the same window:
2026-05-24 17:18:41
power.draw: 17.56 W
graphics clock: 637 MHz
GPU utilization: 96%
temperature: 44 C
pstate: P0
slowdown flags: Not Active / 0 us
2026-05-24 17:18:47
power.draw: 17.11 W
graphics clock: 630 MHz
GPU utilization: 96%
temperature: 44 C
pstate: P0
slowdown flags: Not Active / 0 us
2026-05-24 17:18:59
power.draw: 17.30 W
graphics clock: 630 MHz
GPU utilization: 96%
temperature: 46 C
pstate: P0
slowdown flags: Not Active / 0 us
Across these samples, system memory remained healthy at ~29 GiB available, swap did not move, CPU pressure stayed 0.00, I/O pressure stayed 0.00, and the log showed no recent kernel signals.
Transition from normal/idle state to low-power trap
Before the slow condition began, the GPU was idle and reporting clocks around 1.8 GHz at low power:
2026-05-24 16:28:43
power.draw: 7.59 W
graphics clock: 1800 MHz
GPU utilization: 0%
temperature: 43 C
pstate: P0
slowdown flags: Not Active
Similar idle samples show ~1.7-1.8 GHz graphics clocks with no throttling flags. Then, when inference load begins, the GPU enters a high-utilization low-power state instead of boosting normally:
2026-05-24 16:29:44
power.draw: 17.33 W
graphics clock: 637 MHz
GPU utilization: 96%
temperature: 44 C
pstate: P0
slowdown flags: Not Active
This suggests the GPU is not idle or CPU-starved. It is busy but constrained to a very low power/clock envelope.
Suspicious thermal-limit telemetry
Some samples show a suspicious GPU T.Limit Temp below the current GPU temperature while all slowdown flags remain inactive. For example:
GPU Current Temp: 52 C
GPU T.Limit Temp: 45 C
SW Power Cap: Not Active
HW Slowdown: Not Active
HW Thermal Slowdown: Not Active
HW Power Brake Slowdown: Not Active
SW Thermal Slowdown: Not Active
SW Power Capping: 0 us
HW Thermal Slowdown: 0 us
Average Power Draw: ~17 W
This looks inconsistent: if the thermal limit is truly 45 C and the GPU is at 52 C, a normal thermal slowdown flag would be expected, but none is reported.
What appears ruled out
Based on the captured data, this does not look like:
-
Normal GPU thermal throttling
GPU temperature during the slow state is only ~44-52 C, and NVIDIA reports all thermal slowdown counters as inactive /0 us. -
NVIDIA-visible software power cap
SW Power Capis reported asNot Active, andSW Power Cappingremains0 us. -
Hardware power-brake slowdown
HW Power Brake Slowdownis reported asNot Active. -
System memory pressure / swap pressure
Representative slow-window sample shows ~29 GiB available memory, minimal swap use, and memory pressureavg10=0.00. -
CPU or I/O starvation
CPU and I/O pressure both showavg10=0.00in representative slow-window samples, and recent kernel signal logs are empty.
Observed behavior appears to be an intermittent GB10/GX10 platform-level low-power or low-clock latch:
High GPU load + P0 state + low GPU power + low graphics clock + no NVIDIA-reported slowdown reason
This seems related to, but not identical to, the known GX10 “30 W cap” reports. In this case the issue is transient and the GPU appears to be constrained around 15-18 W / ~630-650 MHz, then later recovers. The logs suggest the limiter may be outside normal nvidia-smi throttle accounting, possibly in platform firmware, EC/PD power policy, or a GB10 driver/firmware reporting path.
GX10 / GB10 platform firmware, EC, PD controller, or NVIDIA firmware can enter a transient low-power state (?) where:
GPU utilization remains ~95-96%
GPU stays in P0
GPU power is limited to ~15-18 W
graphics clock is limited to ~630-650 MHz
all NVIDIA-visible throttle / slowdown flags remain inactive
Meaning of GPU T.Limit Temp on this platform, especially cases where GPU Current Temp exceeds GPU T.Limit Temp while no thermal slowdown flag is active?
