Roundoff error analysis of a pipelined adaptive differential pulse code modulation (ADPCM) coder is presented. The pipelined coder has been developed by employing the relaxed look-ahead technique. It is shown that the precision of the quantized prediction error and those of the predictor coefficients are critical and analytical expressions for lower bounds on these precisions are presented. In addition, an expression for the excess mean-squared prediction error due to finite-precision is also derived. Based on these expressions, a systematic method for assigning precision to various signals in the coder is developed. Simulation results with real image data confirm the results of this analysis.