Can You Lower CMP 170HX Power Without Losing Speed?

A power limit sets the maximum power the GPU² may draw. Lowering it keeps the GPU within a smaller power budget; when a workload reaches the cap, clock speed or throughput may fall. Power use varies with the model, input length, and request count, so compare using the requests you expect to run.

GPU-reported power and whole-system power measured at the wall cover different boundaries. The four-GPU figure below is the sum for four cards; it excludes CPU³, memory, fans, and power-supply losses.

The public GLM-5.3-Flash test used four CMP 170HX cards at Tensor Parallel⁴ 4 and varied the per-card power limit. The single-user column is output speed for one structured request; the eight-user column is aggregate throughput across eight requests. The table separates solo speed, concurrent throughput, and the four-GPU power sum.

Conditions: GLM-5.3-Flash TP4 on four CMP 170HX cards, all at 1,728 MHz HBM, with limits applied live. GPU power is the four-card sum, not wall power. PixelML source record

Reducing the per-card cap from 165 W to 140 W lowered the four-GPU sum from 650 W to 537 W, a 113 W reduction. Single-user speed fell about 1.4%, from 418.6 to 412.7 tok/s; aggregate throughput for eight users fell about 1%, from 822.4 to 814.2 tok/s. Both workloads stayed close in this four-card GLM setup, but solo and eight-user results cannot be substituted for one another.

A separate one-card Qwen3.8-27B test completed at 180 W on a 64 GiB CMP 170HX with 74 SMs, 1728 MHz HBM, PCIe Gen2 x16, and forced airflow. The researcher began a 250 W run but stopped within a minute when memory temperature reached 80°C. Thus 180 W was the completed test condition, and 80°C was the stop point in the 250 W run. It is not a universal manufacturer maximum.

This result differs from the four-card GLM table in model, card count, workload, and cooling, so the two cannot be compared directly. The researcher recorded them in separate public run reports. Qwen one-card test conditions and logs

On Linux, run nvidia-smi -q -d POWER,TEMPERATURE,CLOCK and save the output. First record each card’s current limit, supported minimum and maximum, temperature, and clocks. Do not apply an arbitrary -pl value before checking the range reported by the card. The CMP 170HX has a passive heatsink, so also confirm that strong forced airflow is working.

Use the example only if 180 W is within card 0’s supported range. The command changes card 0 only. Repeat the same request after applying it and record TTFT⁵, output tok/s, power, and memory temperature. Stop the workload if temperature keeps rising or an error appears.

To restore card 0’s original limit, replace OLD_LIMIT_W with the number you recorded and run the restore command. For example, if your note says 165 W, replace the placeholder with 165. After saving the test results, read the setting again and confirm it matches the recorded value.

If the power difference is 100 W at the same measurement point and the system runs 8 hours a day for 30 days, the calculation is 100÷1000×8×30=24 kWh. This is an arithmetic example, not a CMP 170HX wall-power measurement. Do not treat a GPU reading as if it covered the whole system.

To estimate electricity cost, measure whole-system power at the wall and enter actual operating hours and your tariff in the site’s electricity tool. Use only a limit that meets your normal speed and temperature requirements.