The CARMENES radial velocity (RV) survey is observing 324 M dwarfs to search for any orbiting planets. In this paper, we present the survey sample by publishing one CARMENES spectrum for each M dwarf. These spectra cover the wavelength range 520--1710nm at a resolution of at least $R > 80,000$, and we measure its RV, H$\alpha$ emission, and projected rotation velocity. We present an atlas of high-resolution M-dwarf spectra and compare the spectra to atmospheric models. To quantify the RV precision that can be achieved in low-mass stars over the CARMENES wavelength range, we analyze our empirical information on the RV precision from more than 6500 observations. We compare our high-resolution M-dwarf spectra to atmospheric models where we determine the spectroscopic RV information content, $Q$, and signal-to-noise ratio. We find that for all M-type dwarfs, the highest RV precision can be reached in the wavelength range 700--900nm. Observations at longer wavelengths are equally precise only at the very latest spectral types (M8 and M9). We demonstrate that in this spectroscopic range, the large amount of absorption features compensates for the intrinsic faintness of an M7 star. To reach an RV precision of 1ms$^{-1}$ in very low mass M dwarfs at longer wavelengths likely requires the use of a 10m class telescope. For spectral types M6 and earlier, the combination of a red visual and a near-infrared spectrograph is ideal to search for low-mass planets and to distinguish between planets and stellar variability. At a 4m class telescope, an instrument like CARMENES has the potential to push the RV precision well below the typical jitter level of 3-4ms$^{-1}$.

We discuss the effect of atmospheric dispersion on the performance of a mid-infrared adaptive optics assisted instrument on an extremely large telescope (ELT). Dispersion and atmospheric chromaticity is generally considered to be negligible in this wavelength regime. It is shown here, however, that with the much-reduced diffraction limit size on an ELT and the need for diffraction-limited performance, refractivity phenomena should be carefully considered in the design and operation of such an instrument. We include an overview of the theory of refractivity, and the influence of infrared resonances caused by the presence of water vapour and other constituents in the atmosphere. `Traditional' atmospheric dispersion is likely to cause a loss of Strehl only at the shortest wavelengths (L-band). A more likely source of error is the difference in wavelengths at which the wavefront is sensed and corrected, leading to pointing offsets between wavefront sensor and science instrument that evolve with time over a long exposure. Infrared radiation is also subject to additional turbulence caused by the presence of water vapour in the atmosphere not seen by visible wavefront sensors, whose effect is poorly understood. We make use of information obtained at radio wavelengths to make a first-order estimate of its effect on the performance of a mid-IR ground-based instrument. The calculations in this paper are performed using parameters from two different sites, one `standard good site' and one `high and dry site' to illustrate the importance of the choice of site for an ELT.