Critical dimension (CD) errors are traditionally specified and characterized without reference to their spatial frequency spectra. However, a given amplitude of CD variation can have very different consequences depending on its spectrum. CD errors whose variation is over a few micrometers can be much more serious than those of the same magnitude that extend over several chips. Existing CD metrology tools, such as scanning electron microscopy or electrical resistance measurements, are seldom used to characterize these short-range CD variations, particularly those with spatial wavelengths below 100 mu m, because of the large amount of data required and the difficulty of collecting data in such a dense grid. We report a new method of measuring CD variations using static random-access memory (SRAM) circuits in which direct measurements of bit-line currents reveal the individual transistor gate length variations within each memory cell. With the compactness and regularity of the SRAM layout we can measure CD variations with spatial periodicities down to 6 mu m. By repeatedly measuring each cell in a memory chip and recording the corresponding currents we can achieve sufficient data to minimize noise, and through two-dimensional bandpass filtering 0.2 nm CD variations can be detected. Two designs of 4 Mbit SRAMs fabricated using 250 nm design rules were studied. The resulting CD variations yielded spectra that were dominated by peaks whose origins included uncorrected electron beam and optical proximity effects. Pattern-independent variations ascribable to the reticle generator itself appeared to contribute only a small fraction of the total error observed.