Not running at full speed on DGX Spark?

[jerkstorecaller]

I'm playing around with a DGX Spark, Asus GB10.

• First llama-bench test: llama-bench -m noctrex_NVIDIA-Nemotron-3-Nano-30B-A3B-MXFP4_MOE.gguf --repetitions 1 -p 950000 . Although nvtop is showing 96% GPU usage, wattage is quite low.
• For the first 15 minutes, POW mainly stayed around 15-20W. It sometimes goes up to 40W for a few seconds, then back down to 15-20W.
• Then for X amount of minutes, it was consistently 31-36W
• Then I noticed it alternating between the above two states
• The whole time, GPU=96%, around 2405 MHz
• The whole time, TEMP=58-62C
• VRAM usage is 87-90/114GB
• New test: GPU burn: I figured I needed a proper stress test to see if it's thermal throttling. I used wilicc/gpu-burn. I see the GPU pulls around 94W at full load (TEMP goes up to 80), which is closer to what's expected. (An Nvidia rep had posted on their forums: "Spark GPU can draw a max of 120W and the rest of the system including CPU, CX7 and peripherals can draw 100W. The power supply is rated for 240 Watt of peak draw.")
• Repeat llama-bench test: I reran the same llama-bench command. POW quickly stabilized around 40W for the first 15mins, whereas previously it was 15-20W.

I'm a novice when it comes to hardware, so I don't understand what's going on.

1. What can explain such a difference in power draw between 2 identical runs of llama-bench?
2. Why would power draw be so low in general? Is the GPU underutilized?

Tbh all I'm interested in is maximizing the speed of my inference by using the hardware to its fullest. So what would that take? I already tried -t 20 and -b 2048, they produced identical power draw as the 2nd test.