Supplying a constant current to a variable, low-resistance load, such as a thermoelectric cooler, can be a very inefficient endeavor when using the linear approach. Even powering the linear current source from an efficiently downconverted version of the supply voltage will incur unnecessary dissipation if the load resistance is variable. On the other hand, a switch-mode current source provides the most efficient means of supplying such a load. A buck regulator, sense resistor, gain stage, and boost generator give the desired result.

The circuit in the figure supplies a constant 3.5 A to a load resistance (R_{O}), which varies from 0.4 to 1 Ω, with a supply voltage (V_{IN}) that ranges from 8 to 18 V. At R_{O} = 0.4 Ω, the efficiency is 80% and the ripple current is 10 mA p-p.

The LT1374 is a buck-regulator chip that operates at 500 kHz. In this application, the output voltage is too low to supply an adequate, efficiency-enhancing boost for the chip. So, a simple 5-V boost generator consisting of Q1, D3, and the 47-kΩ resistor has been included. Boost power of 70 mW reduces chip dissipation by about 0.7 W.

To promote efficiency and keep waste heat low, the sense resistor (R_{S}) is chosen to be much lower than the lowest load resistance. The sense voltage developed across R_{S} is 105 mV. Hence the gain stage, consisting of the LM2904 and associated resistors, is included to amplify this sense voltage to a level consistent with the regulator chip's control voltage. For the LT1374, this level is 2.4 V.

When ensuring the stability of the current source, the presence of the LM2904 in the feedback loop must be considered. The open-loop gain of the LM2904 rolls off at about 20 dB/decade—from a maximum of about 110 dB and crossing zero at about 1 MHz. Measurements indicate that the LM2904's phase shift begins at about 1 kHz and is −20° at 25 kHz, −40° at 50 kHz, −60° at 75 kHz, −80° at 150 kHz, and −160° at 1 MHz. Once the effect of the LM2904 is accounted for, stabilizing the control loop can proceed according to the principles set forth in Linear Technology's Application Note 76, which can be found online at *www.linear.com*.

The expected output ripple current can be predicted by dividing the ripple voltage on the filter capacitor (C1) by the impedance of the load branch. This impedance consists of the load resistance, sense resistance, and branch inductance. At these frequencies, this inductance can have a significant impact in reducing the current ripple. It may also have some impact on what must be done to stabilize the loop.

For a buck regulator in continuous conduction, if (1−D)T < 2R_{C}C1 (which is the usual case), the peak of the C1 ripple voltage occurs at t = DT and the amplitude is given by I_{P}(R_{C} + DT/4C1). Here, I_{P} is the peak inductor current, R_{C} is the ESR of C1 measured at f = 1/T, f is the operating frequency of the control IC, and D is the duty cycle. This formulation assumes that the inductive reactance of the filter branch can be ignored.