We launched the site with CPI data for 10 mice.
Shortcuts
temppal.gg team
YESLAB Inc. researchers
- Dr. Byungjoo Lee CEO, YESLAB Inc.; Associate Professor, Yonsei University.
- Dr. Sunjun Kim CTO, YESLAB Inc.; Associate Professor, DGIST.
- Minhyeok Baek Researcher, YESLAB Inc.; PhD Student, DGIST.
- Namsub Kim Researcher, YESLAB Inc.; PhD Student, Yonsei University.
- Seonho Kim Researcher, YESLAB Inc.; PhD Student, Yonsei University.
VISITORS
------
Gaming Mouse CPI Benchmark
CPI Benchmark Browser
Current CPI Benchmark Leaders
Loading current benchmark leaders.
Test plan: 400/800/1600/3200 CPI, FIRM/GLASS/SOFT pads, 1000 Hz. Leaders use the average across all published test conditions.
New to these numbers? See Method for what CPI bias, CPI jitter, and CPI drift mean.
Browse Measured Mouse Records
Bias, jitter, and drift columns show weighted score terms. Higher is better; the three terms add up to Score.
Click a mouse row to open raw CPI x pad fitted data.
Mouse
Repeated-unit CPI tests will be standardized by model, then grouped into a brand-level build-quality score.
Unit-to-unit spread in fitted base CPI for the same model.
Unit-to-unit spread in residual report variation.
Unit-to-unit spread in motion-dependent CPI drift.
Preview table
Waiting for repeated-unit CPI benchmark data
| Brand | Model | Units | Bias z | Jitter z | Drift z | Model score | Brand score |
|---|---|---|---|---|---|---|---|
| Example Brand A | Model X | 4 | +0.42 | +0.31 | -0.08 | 0.22 | 0.18 |
| Model X Mini | 3 | +0.10 | +0.18 | +0.14 | 0.13 | ||
| Example Brand B | Model Y | 5 | -0.22 | +0.06 | -0.35 | -0.17 | -0.17 |
When the database is ready, this page will rank models and brands by repeated-unit consistency rather than by single-device performance.
Gaming Mouse CPI Benchmark
Compare a Mouse Against Ideal Motion
Choose a measured mouse, then compare its fitted CPI model against an ideal mouse trajectory.
Choose a motion speed and aim style to check drift from an ideal mouse.
Gaming Mouse CPI Benchmark
What We Report
| Reported quantity | Question | Direction |
|---|---|---|
| CPI bias | After model fitting, how far is the fitted base CPI from the selected nominal CPI? | Closer to zero is better |
| CPI jitter | After modeled CPI effects are removed, how consistent is the remaining report stream? | Lower is better |
| CPI drift | How much does effective CPI change with translational speed and angular speed? | Closer to zero is better |
How CPI Benchmarking Works
CPI is not treated as a vendor label. It is treated as a hardware
term in the hand-to-pointer mapping. The benchmark protocol starts
by comparing physical motion at the optical sensor location with
the raw dx,dy counts reported by the mouse, then fits
CPI models to derive the public benchmark records.
Variables, Model, and Reported Quantities
After synchronization, every HID report is treated as one microstep. The variables below are introduced before they are used in the model.
| Symbol | Definition |
|---|---|
| CPInom | The CPI selected in firmware or driver software. |
| i | Index of a single synchronized mouse-report microstep. |
| di = (dxi, dyi) | The raw HID count vector reported by the mouse at microstep i. |
| mi = ||di|| | The observed raw-count magnitude for that report. |
| vi | The translational speed of the optical sensor location, reconstructed from motion capture. |
| ωi | The angular speed of the mouse body around the desk plane. Both speeds are scalar magnitudes. |
| Δti | The time interval represented by microstep i. |
The fitted CPI model says that the effective CPI at a moment can shift with translational speed and angular speed:
mi =
||di|| =
√(dxi2 +
dyi2)
CPIi =
c0 +
c1vi +
c2ωi
m̂i =
CPIi
0.0254
viΔti
ei =
mi -
m̂i
Here c0 is the baseline effective CPI, c1 is the CPI change per 1 m/s of sensor speed, and c2 is the CPI change per 1 rad/s of angular speed. Because one inch is 0.0254 m, the model converts physical sensor travel into predicted raw counts through the division by 0.0254.
The parameters are fit by minimizing the difference between observed and predicted count magnitudes. The public page summarizes the CPI model and reported quantities; quality-control and robustness procedures are applied consistently across benchmark records.
Once the model is fit, the public benchmark reports three derived quantities. CPI bias is computed from the fitted baseline coefficient c0, while CPI drift is computed from the speed and angular-speed coefficients c1 and c2. The published CPI analysis covers sensor speeds up to 1 m/s and angular speeds up to 5 rad/s, a range chosen to include fast esports-style hand motion rather than only slow straight-line travel.
| Reported quantity | Formula | Meaning |
|---|---|---|
| CPI bias (BCPI) | BCPI = 100 × c0 - CPInom CPInom | Base CPI bias after model fitting. Zero means the fitted base CPI c0 matches the selected CPI. |
| CPI jitter (J) |
εi =
0.0254ei
viΔti
J = 100 × S({εi}) CPInom |
Residual CPI-equivalent fluctuation after the fitted model is removed. S is a standardized spread operator applied consistently across records. Lower means less report-to-report jitter. |
| CPI drift (DCPI) |
Dv =
100 ×
c1
CPInom
,
Dω =
100 ×
c2
CPInom
DCPI = Dv + Dω 2 |
Signed average of speed dependence and angular-speed dependence. Dv is percent CPI change per 1 m/s, and Dω is percent CPI change per 1 rad/s. Zero means CPI is flat with respect to movement speed. |
Movement-regime balancing and reference constants are fixed across comparable records, so devices can be compared on the same fitted summary quantities.
How the Aggregate Score Is Combined
The benchmark table also shows a single Score so users can scan the public records quickly. That Score is not a separate measurement. It is a dataset-relative summary made from the three reported CPI quantities above.
zB =
z(|BCPI|),
zJ =
z(J),
zD =
z(|DCPI|)
Score =
-0.5zB
-2zJ
-2zD
4.5
Higher Score is better. CPI bias receives weight 0.5 because a player can often compensate a small base-CPI offset after measurement by adjusting the selected CPI, including through temppal.gg's CPI Match tool. CPI jitter and CPI drift receive weight 2 each because they describe consistency and motion-dependent behavior that cannot be corrected as cleanly with a simple CPI setting. In benchmark tables, the Bias, Jitter, and Drift columns show the weighted score term added by each metric, so those three displayed values sum to the final Score.
Measurement Procedure
The physical measurement procedure is separated from the fitted model. A public benchmark record is only published after the mouse has passed through instrumentation, synchronized acquisition, and model-based summarization.
Instrument and calibrate
Prepare the tracked mouse body and estimate the sensor location.
Acquire and align
Record natural motion and raw reports, then align both timelines.
Fit and report
Fit CPI models and publish comparable fitted quantities.
- Track the mouse body. A lightweight 3D-printed jig with reflective markers is attached to the top of the mouse. The jig is designed so the hand can still grip and move the mouse normally.
- Locate the sensor. A temporary underside marker is attached near the optical sensor opening. This short calibration step estimates the sensor offset relative to the tracked rigid body. The underside marker is removed before normal movement recording.
- Record natural movement. The experimenter moves the mouse freely for 20 seconds while avoiding clutching. The collection software encourages coverage across translational speed and angular speed rather than only straight, rail-like strokes.
- Capture both streams. Motion-capture pose and raw mouse input are recorded together. Raw input is collected as relative counts without OS pointer acceleration or cursor filtering.
- Reconstruct sensor motion. The rigid-body trajectory is converted into the physical trajectory of the sensor location, including the effect of mouse rotation around the desk plane.
- Synchronize timelines. Motion-capture-derived displacement is resampled onto mouse report timestamps. A relative delay is estimated so physical motion and raw reports can be compared.
- Fit CPI models. We fit a constant CPI model and a movement-dependent model using translational speed and angular speed. The fitted parameters produce effective CPI estimates and the reported benchmark quantities.
- Summarize jitter. Residual variability is summarized across the tested movement range, up to 1 m/s translational speed and 5 rad/s angular speed, using a standardized balancing procedure. This prevents one trial's particular motion path from dominating the jitter estimate.
Apparatus and Test Surfaces
The current benchmark apparatus follows the setup described in our CPI measurement paper. The exact hardware matters because the public records compare raw HID counts against an independently tracked physical trajectory.
| Part | Configuration used |
|---|---|
| Workstation | AMD Ryzen 5 5600X 6-Core Processor at 3.70 GHz, NVIDIA GeForce RTX 2060 GPU, 32 GB RAM, 64-bit Windows 10. |
| Display | 24.5-inch BenQ ZOWIE XL2540K, 1920 x 1080 resolution, 240 Hz refresh rate. |
| Motion capture | OptiTrack system with 13 Prime 13 cameras sampling at 240 Hz. Calibration and rigid-body tracking are managed in Motive 3.0.0 Beta 3. |
| Acquisition software | Custom Python acquisition framework with a PyQt5 interface. Raw mouse reports are collected through the Windows Raw Input API; motion-capture poses are streamed through the OptiTrack NatNet SDK. |
| Tested pads | Pulsar eS Jupiter Pro, Pulsar eS Neptune Pro, and Pulsar Superglide 3 Type S, each 420 x 330 mm. In the public benchmark these correspond to Soft, Firm, and Glass test-surface labels. |
| Tracking jig | Lightweight 3D-printed rigid-body marker jig, approximately 3.3 g including attached reflective markers. |
Gaming Mouse CPI Benchmark
Compare Mice on Shared Conditions
Gaming Mouse CPI Benchmark
Switching Mice: What CPI Keeps the Same Feel?
Changing Skates: How Much Should CPI Move?
Enter your current CPI and skate thicknesses. The calculator uses the published surface-distance model to recommend the CPI setting that keeps the effective base CPI closest to your old setup.
Personal Gear Advantage
Choose a Task and Read the Rules
- Use your real play setup Set the CPI and OS pointer/tracking speed to what feels best before you start.
- Push speed and accuracy Move as quickly and as accurately as possible. If the result feels off, tune and try again.
6
difficulty levels
72
analyzed trials
About 1 minute
typical completed run
- Use your real play setup Set the CPI and OS pointer/tracking speed to what feels best before you start.
- Push speed and accuracy Move as quickly and as accurately as possible. If the result feels off, tune and try again.
6
difficulty levels
72
analyzed trials
About 1 minute
typical completed run
Personal Gear Advantage
How Personal Runs Are Analyzed
The personal test is a browser-based pointing and aiming study. It does not replace the hardware CPI benchmark; it records how a player performs with a specific mouse, display, and browser.
Pointing performance
Fit Fitts-style speed-accuracy models from target distance or camera-rotation distance, target width, movement time, and endpoint spread.
Movement strategy
Use model-based analysis to study trajectories, corrections, and player-device interaction.
Submovement structure
Extract correction episodes from the pointer-speed signal with Persistence1D-style peak analysis.
Fitts Law Fitting
Each completed run contains six index-of-difficulty conditions. For each condition, temppal.gg summarizes target-acquisition time and endpoint variability, then fits the relation between task difficulty and movement time.
ID = log2(D / W + 1)
MT = a + bID
TP = 1 / b
D is the center-to-center distance between consecutive targets, not the diameter of the target circle. W is target width. Effective throughput uses the IDxTTAe convention: endpoints are projected onto the target-to-target task axis, then effective amplitude and effective width define the effective ID. The reported fit includes nominal-width and effective-width variants, the fitted line, throughput, and R2. This follows the effective-throughput method recommended in Accuracy Measures for Evaluating Computer Pointing Devices.
Computational Rationality Fitting
Fitts law is compact, but it compresses the whole movement into a line. We also plan to fit richer computational-rationality models inspired by point-and-click simulation work from our team. That line of work models point-and-click behavior as a controlled process with perception, limb movement, click decisions, device effects, and intermittent corrections rather than only final movement time.
The reference model is described in A Simulation Model of Intermittently Controlled Point-and-Click Behaviour from CHI 2021. On temppal.gg, this analysis will be used as a research layer for explaining how a player's movement strategy changes across mice and settings.
Persistence1D Submovement Analysis
Pointer trajectories often contain several corrections inside one target acquisition. The submovement analysis converts each movement into a speed profile, smooths it to a standardized analysis rate, and detects persistent local peaks. Those peaks are treated as candidate submovements.
| Output | What it means |
|---|---|
| Submovement count | How many persistent movement pulses appear in one target-acquisition trial. |
| Correction timing | Where major corrections occur between movement start and click. |
| Sampling quality | Whether the browser delivered enough pointer samples for reliable kinematic analysis. |
These analyses are shown next to the saved run summary so players can compare not only whether one setting was faster, but how the movement was produced.
Gear Meta and Personal Aggregation
Players differ in skill, display size, practice history, and testing context. Gear Meta therefore does not rank mice by raw average scores alone. It treats each player as their own control, then asks whether a mouse improves that player's throughput relative to the same player's other mice.
| Step | What is computed |
|---|---|
| Best run per mouse | For the same player, task, and mouse, only that player's best valid run is used. Repeating the same mouse can improve your own best, but it does not create extra copies of the same mouse in Gear Meta. |
| Player baseline | Within each player and task, temppal.gg averages the best runs from the mice that player tested. Each mouse is then expressed as a percent improvement over that same-player, same-task baseline. |
| Mouse effect | For each mouse, the public Gear Meta chart averages those same-player improvement values across players. Error bars show the variation of those user-level mouse effects, not raw trial noise. |
| My Data | Your private My Data page can show all saved runs or only the best run per mouse, so your personal view can match the public Gear Meta logic when you want it to. |
A positive Gear Meta value means a mouse performed better than the same players' own baseline on throughput. Runs are included only when the quality checks pass.
Personal Gear Advantage
Choose a Task and Record What You Use
What You're Using for This Run
XP Rules
Log in and identify a mouse to preview this run.
XP is awarded only after a valid completed run.
Personal Gear Advantage
Analyze Your Saved Runs
Loading your data analysis.
Personal Gear Advantage
Gear Meta
Muscle Memory Tester
Find Your Familiar CPI
Blind CPI Tuning
You say you know your CPI. Can you find it blind? Each fullscreen trial starts from a random hidden virtual CPI; trim it by feel, then press Enter when it feels like yours.
This test requires a desktop browser with pointer lock.
Review Your Muscle-Memory Data
How the Muscle-Memory Test Works
This test does not ask whether a mouse is objectively fast. It asks a narrower question: when the visible number is hidden, can you tune a virtual cursor back to the sensitivity that feels familiar?
1. Random hidden start
Each trial begins with a random virtual CPI multiplier. The number is hidden so the adjustment is based on feel, not memory of a setting menu.
2. Same target task
You use either the 2D point-and-click task or the 3D aim-and-shoot task. The target loop continues until you press Esc.
3. Blind adjustment
Press W or S to raise or lower virtual sensitivity. Press Enter only when the cursor feels like your usual CPI.
4. Consistency score
After at least 10 trials, temppal.gg summarizes how tightly your selected virtual gains cluster.
What Is Recorded
For each accepted trial, temppal.gg records the hidden random multiplier and the adjustment multiplier you selected. Their product is the virtual gain you settled on for that trial. Logged-in results are saved to your account; logged-out results are shown only temporarily in the browser.
Gselected =
Grandom × Muser
CV =
SD(Gselected) /
mean(Gselected)
How to Read the Result
| Quantity | Meaning |
|---|---|
| Mean selected gain | The center of the sensitivity you repeatedly chose during that session. |
| Standard deviation | How much the chosen virtual gain changed from trial to trial. |
| CV | Standard deviation divided by the mean. Lower CV means the familiar feel was reproduced more consistently. |
| Population percentile | Where your best saved CV sits among other saved temppal.gg muscle-memory tests. Lower is better. |
Important Assumption
Keep your mouse hardware CPI and OS mouse speed setting unchanged during a session. The test measures consistency under one physical setup; changing the real setup mid-session changes what the hidden virtual gain means.
Team and Contact
People Behind temppal.gg
| Name | Affiliation |
|---|---|
| Dr. Byungjoo Lee | CEO, YESLAB Inc.; Associate Professor, Yonsei University. |
| Dr. Sunjun Kim | CTO, YESLAB Inc.; Associate Professor, DGIST. |
| Minhyeok Baek | Researcher, YESLAB Inc.; PhD Student, DGIST. |
| Namsub Kim | Researcher, YESLAB Inc.; PhD Student, Yonsei University. |
| Seonho Kim | Researcher, YESLAB Inc.; PhD Student, Yonsei University. |
Prior Work From Our Team
These are papers authored by temppal.gg team members and close collaborators. The list is filtered to work directly related to mouse hardware, pointing, jitter, FPS aiming, and input behavior.
- Journal of Electronic Gaming and Esports 2026Profiling Rhythm Game Performance Using Multi-Lane Moving-Target Acquisition Modelforthcoming
- CHI 2025Hardware-Embedded Pointing Transfer Function Capable of Canceling OS Gainsread
- International Journal of Human-Computer Studies 2025Modeling Visually-Guided Aim-and-Shoot Behavior in First-Person Shootersread
- CHI 2024User Performance in Consecutive Temporal Pointing: An Exploratory Studyread
- CHI 2024Quantifying Wrist-Aiming Habits with A Dual-Sensor Mouse: Implications for Player Performance and Workloadread
- Journal of Electronic Gaming and Esports 2024Operationalizing General Mechanical Skill in Time-Pressure Action Esportsread
- UIST 2024Effects of Computer Mouse Lift-off Distance Settings in Mouse Lifting Actionread
- CHI 2022Quantifying Proactive and Reactive Button Inputread
- UIST 2021Do We Need a Faster Mouse? Empirical Evaluation of Asynchronicity-Induced Jitterread
- CHI 2021A Simulation Model of Intermittently Controlled Point-and-Click Behaviourread
- CHI 2021Secrets of Gosu: Understanding Physical Combat Skills of Professional Players in First-Person Shootersread
- CHI 2020Button Simulation and Design via FDVV Modelsread
- CHI 2020AutoGain: Gain Function Adaptation with Submovement Efficiency Optimizationread
- CHI 2020Optimal Sensor Position for a Computer Mouseread
- CHI 2018Impact Activation Improves Rapid Button Pressingread
- CHI 2018Moving Target Selection: A Cue Integration Modelread
- CHI 2018Neuromechanics of a Button Pressread
- CHI 2016Modelling Error Rates in Temporal Pointingread
- Human-Computer Interaction 2015A Mouse With Two Optical Sensors That Eliminates Coordinate Disturbance During Skilled Strokesread
- Ergonomics 2013A Kinematic Analysis of Directional Effects on Mouse Controlread
Contact
For collaboration, commercial measurement requests, manufacturer inquiries, dataset licensing, protocol review, or esports-team research partnerships, email yeslab@temppal.gg.
Please include your organization, the mouse or product involved, relevant settings such as CPI, polling rate, pad, feet, connection mode, and what you want to measure, validate, license, or discuss.
Contact
For collaboration, commercial measurement requests, manufacturer inquiries, dataset licensing, protocol review, or esports-team research partnerships, email yeslab@temppal.gg.
Please include your organization, the mouse or product involved, relevant settings such as CPI, polling rate, pad, feet, connection mode, and what you want to measure, validate, license, or discuss.