Home/Hardware

SG90 and MG996R: the two servos most often bought wrong

The default actuation of first robots. Cheap is their virtue; inflated specs are their norm. This entry takes the numbers apart.

SG90
9 g, plastic gear, 1.8 kg·cm claimed (4.8 V)
MG996R
55 g, metal gear, 9–11 kg·cm claimed (4.8–6 V)
Signal
Standard PWM (50 Hz, 500–2500 µs)
Price
SG90 about USD 1–2; MG996R about USD 3–6
Feedback
None (position and load not readable)
Counterfeit risk
High (same model, threefold price spread)
The spec, and what it measures
Both figures are stall torque at nominal voltage — instantaneous. MG996R's 9–11 kg·cm means 'at 6 V with the output shaft locked, for a moment'. It is not a sustained rating; continuous load should be taken at a third to a quarter, so 2–3 kg·cm as a design value. The SG90's number means the same, but its plastic gears fail before its torque limit does — the failure mode is stripped teeth, not stall, and that difference decides where each belongs.
Price tiers
USD 1–2 (SG90, 9 g, light loads) → 3–6 (MG996R, 55 g, legs) → 8–12 (DS3218, bearinged output, stiffer case; common in quadrupeds) → 15+ (Feetech STS / Dynamixel: bus, position feedback, configurable). A pound for a pound mostly holds in this category — counterfeits being the one place where price and quality detach.
How to pick
Three moves: ① buy versions with a bearinged output shaft — several times the service life; ② size by continuous load, not stall (legs get MG996R or better; bench trinkets get SG90); ③ buy spares in the same batch — stripped gears are a probability, not a defect. The day you need feedback or parameters, jump straight to bus servos; there is no path there through PWM.
Known pitfalls
① Inflated specs: the same model from different factories varies up to 30% in real torque; ② a plastic-geared SG90 on a leg is a scheduled failure; ③ PWM deadband and jitter — worse at the low end, visible as the 'fine trembling' that makes gait robots look unstable at stance; ④ counterfeit detection by weight is the most direct test: SG90 about 9 g, MG996R about 55 g — light means wrong.
What we checked
We verified: the torque-voltage definitions across multiple datasheets; that 'using stall figures as continuous ratings' is the category's most common misuse; and the difference in failure modes between plastic- and metal-geared versions. We did not verify: per-brand counterfeit rates or lifetime statistics — those need batch-level data, and community reports give orders of magnitude only. The grade stays accordingly.
This site's call

Entry-level trial parts with a clear division of labour: SG90 for unloaded small motion, MG996R / DS3218 for legs. They teach you every limitation of PWM servos — no feedback, deadband, stall damage — and those limitations are precisely what the premium on bus servos buys away. Step straight to bus servos if the budget allows; start here if it does not. These parts survive being learned on.

The rated torque is not the usable torque

For the same SG90, published stall torque ranges from 1.2 to 1.8 kg·cm depending on the source. That spread is not a typo — it is a difference in convention (test method, voltage, batch). And it captures the defining trait of this class: the rating is a stall peak, not a value you can use continuously.

Discount it in three steps: (1) stall is a peak, and sustained current heats the winding and derates it, so keep at least a 2× margin; (2) ratings are usually quoted at 6.0 V while real supplies sit at 4.8–5 V, and dropping from 6 V to 4.8 V costs roughly 20% of torque; (3) gearbox efficiency is about 60–85% for plastic gears and 75–90% for metal. Stack those and a "1.8 kg·cm" SG90 may hold less than half of that continuously.

The place that kills SG90s is the gripper

A gripper is the one use that sits near stall for long periods — holding an object is not motion, so the winding stays energised with nothing to cool it. When an SG90 dies in a gripper, this is almost always why, not because the torque was too low. For gripping, move to the metal-gear MG90S: same footprint, same wiring, gears that survive stall.

Power: never use the Arduino 5 V pin

ModelStall torque (4.8 V)Stall current (approx.)Gears
SG901.2–1.8 kg·cm~0.7 APlastic
MG90S2.0–2.2 kg·cm~1.2 AMetal
MG996R9–11 kg·cm1.5–2.5 AMetal

An MG996R can pull over 2 A at stall — two to three times an SG90 — and wiring that to an Arduino 5 V pin will brown out or cook the board. For more than two or three servos, or any metal-gear servo, use a dedicated 5–6 V supply (or a BEC) and tie all grounds together.

Frequently asked

How strong is an SG90 servo, really?

The rated stall torque is 1.8 kg·cm at 4.8 V — but sources quote anywhere from 1.2 to 1.8, so even the convention is not settled. And that is a stall peak: after a 2× margin, voltage drop and gearbox efficiency, the continuously usable figure is usually under half of it. Size by usable torque, not by the label.

What is the difference between SG90 and MG90S?

Same size and same wiring; they are controlled identically. Two differences only: the MG90S has metal gears (it survives stall instead of stripping), and slightly more torque (about 2.0–2.2 kg·cm versus 1.2–1.8). So the call is simple — use the MG90S when a joint carries weight or takes knocks; the SG90 is fine for light, unloaded motion.

How much voltage does a micro servo need?

The standard operating range is 4.8–6.0 V. Running at 6 V buys roughly 20% more torque and speed than 4.8 V, at the cost of faster heating and a slightly shorter life. Above 6 V — and most standard servos hit their limit around 7.4 V — you shorten the life sharply or kill it outright, unless the model is explicitly rated for high voltage.

How much current does an SG90 draw?

Tens to about a hundred milliamps unloaded, and roughly 0.7 A at stall; the MG90S about 1.2 A; the MG996R 1.5–2.5 A at stall. One MG996R already exceeds what an Arduino's on-board 5 V pin can safely supply — use a dedicated supply for multiple servos and tie the grounds together.

Sources

  • Multiple vendor datasheets (torque definitions, voltage, weight)
  • Community usage reports (failure modes, counterfeit identification)

Objects in this entry

Last checked 2026-09-28