Sensor and Performance

The MSI Clutch GM41 Lightweight is equipped with the PixArt PMW3389. According to specifications, the 3389 is capable of up to 16,000 CPI, as well as a maximum tracking speed of 400 IPS, which equals 10.16 m/s. Out of the box, five pre-defined CPI steps are available: 400, 800, 1600, 3200, and 6400.

All testing was done on the latest firmware (1.03.00). As such, results obtained on earlier firmware versions may differ from those presented hereafter.

CPI Accuracy

"CPI" (short for counts per inch) describes the number of counts registered by the mouse if it is moved exactly an inch. There are several factors (firmware, mounting height of the sensor not meeting specifications, mouse feet thickness, mousing surface, among others) which may contribute to nominal CPI not matching actual CPI. It is impossible to always achieve a perfect match, but ideally, nominal and actual CPI should differ as little as possible. In this test, I'm determining whether this is the case or not. However, please keep in mind that said variance will still differ from unit to unit, so your mileage may vary.


I've restricted my testing to the four most common CPI steps, which are 400, 800, 1600, and 3200. As you can see, there is no deviation at all, which is a perfect result.

Motion Delay

"Motion delay" encompasses all kinds of sensor lag. Any further sources of input delay will not be recorded in this test. The main thing I'll be looking for in this test is sensor smoothing, which describes an averaging of motion data across several capture frames in order to reduce jitter at higher CPI values, increasing motion delay along with it. The goal here is to have as little smoothing as possible. As there is no way to accurately measure motion delay absolutely, it can only be done by comparison with a control subject that has been determined to have the lowest possible motion delay. In this case, the control subject is a G403, whose 3366 has no visible smoothing across the entire CPI range. Note that the G403 is moved first and thus receives a slight head start.


First, I'm looking at two xCounts plots—generated at 1600 and 16,000 CPI—to quickly gauge whether there is any smoothing, which would be indicated by any visible "kinks." Gauging smoothing on the GM41 is complicated by the fact that tracking is not consistent. Whereas in most sensor implementations, the way plots look is typically dictated by the speed a swipe is performed with, things are different on the GM41. Basically, every swipe yields a different result, sometimes looking decent as the first one above, yet horrible at other times. I haven't been able to ascertain what could be causing this, but it is fair to say that something is seriously wrong with SPI timing.


In order to determine motion delay, I'm looking at xSum plots generated at 1600, 3200, and 16,000 CPI. The line further to the left denotes the sensor with less motion delay. 1600 CPI shows no motion delay differential. 3200 CPI displays a differential of roughly 3 ms, and 16,000 CPI a differential of roughly 14 ms. This is in line with what to expect from a 3389, which has 32 frames of smoothing at and above 1900 CPI that is then doubled at 6000 and 11,300 CPI.

Speed-related Accuracy Variance (SRAV)

What people typically mean when they talk about "acceleration" is speed-related accuracy variance (or SRAV for short). It's not about the mouse having a set amount of inherent positive or negative acceleration, but about the cursor not traveling the same distance if the mouse is moved the same physical distance at different speeds. The easiest way to test this is by comparison with a control subject that is known to have very low SRAV, which in this case is the G403. As you can see from the plot, no displacement between the two cursor paths can be observed, which confirms that SRAV is very low.

Perfect Control Speed

Perfect Control Speed (or PCS for short) is the maximum speed up to which the mouse and its sensor can be moved without the sensor malfunctioning in any way. I've only managed to hit a measly 4.5 m/s, which is within the proclaimed PCS range and shows no sign of the sensor malfunctioning.

Polling Rate Stability

Of the available polling rate settings (125/250/500/1000 Hz) only 1000 Hz looks and performs fine, whereas all of the other polling rates have periodic off-period polls. Polling stability is unaffected by any RGB lighting effect.

Paint Test

This test is used to indicate any potential issues with angle snapping (non-native straightening of linear motion) and jitter, along with any sensor lens rattle. As you can see, no issues with angle snapping can be observed. No jitter is visible at 1600 and 3200 CPI, the latter of which already has smoothing applied. 16,000 CPI still looks surprisingly well-controlled owing to a ridiculous amount of smoothing. Lastly, no sensor lens movement can be observed.

Lift-off Distance

The Clutch GM41 Lightweight offers two pre-defined LOD levels. At the "Low" setting, the sensor tracks only intermittently at a height of 2 DVDs (1.2 mm<x<2.4 mm, with x being LOD height). Using the "High" setting, this doesn't change one bit, which is why I strongly suspect lift-off distance adjustment being broken in some way. Apparently, "High" is applied even when "Low" is selected. Keep in mind that LOD may vary slightly depending on the mousing surface (pad) it is being used on.

Click Latency

Since mechanical switches are being used for the buttons in most computer mice, debouncing is required in order to avoid unintended double clicks. Debouncing typically adds a delay (along with any potential processing delay), which shall be referred to as click latency. As there is no way to measure said delay directly, it has to be done by comparing it to a control subject, which in this case is the Logitech G203. Click latency has been measured to be roughly +3.6 ms when compared to the SteelSeries Ikari, which is considered as the baseline with 0 ms, with standard deviation being 0.60 ms. Please keep in mind that the measured value is not the absolute click latency. Comparison data comes from this thread as well as my own testing, using qsxcv's program.