1. Introduction. Testing methodology.
2. The best, or anything above 75 dB SNR.
3. The midrange, or 60 to 75 dB SNR.
4. The worst, or anything below 60 dB SNR.
5. Summary.
 

Some day in the past I've started to test different kinds of sound cards and collect that information. It has been necessary personally for me to know what a particular sound card is worth of, and so I've been able to make suggestions of its use to certain people who appreciate my opinion. Therefore, I've decided to get those figures and screen shots all together and turn them into some order while adding new still. The last two ones have been tested in January of 2005, and I have had frozen the project since.
 
(1-Jan-2007, update): All screen shots have used to be 960x720 32-colour GIFs for a long time until they've got converted to PNGs to reduce their size by approximately 40%.

All sound cards have been tested in the same environment, I mean inside the same machine which I eventually happen to use for the last 5 years (since 1999) as a personal workstation, so results are fully comparable. The testing suite is fairly simple: analogue loopback, i. e. the primary linear output is connected directly to the primary linear input of a particular card, some test tones are played and recorded, further processed with a FFT analyser (SpectraLAB) using 16K FFT size to get some useful information. No hard feelings or preconceived attitude, I've tried to benchmark every card to the best.
 
The tests include flat-weighted SNR (Signal to Noise Ratio), THD (Total Harmonic Distortion), THD+N (Total Harmonic Distortion + Noise) and IMD (InterModulation Distortion). For the first three, a 1kHz sine waveform with -1dB amplitude has served as a source signal, for the last one — a complex sine waveform of 250Hz -3dB and 8020Hz -15dB. All waveforms are 44100Hz 16-bit stereo.
 
SNR (also known as S/N) is used often to demonstrate how much noise a particular unit carries. To obtain, it's usually enough to set volume controls to certain power level (for example, +4dBu), make sure there is no signal applied (but all necessary equipment is connected and powered), measure a noise level and divide one by another. Expressed in decibels. Should be flat-weighted (in other words, without any noise shaping) on full 22kHz range — 20kHz is considered to be the uppermost frequency recognised by human ear, though some people cannot distinctly hear even beyond 16kHz; on the other hand, some investigations have proved that tones beyond 20kHz are sensed and used by subconscience for space orientation, like bats have but much weaker; additional 2kHz are used mostly by low-pass filters to take effect without attenuating frequencies close to 20kHz. Some manufacturers may prefer A-weighting, i. e. noise shaping adaptive to human ear, but it's just a trick to achieve higher figures for better sales without improving products actually. Of course, higher results are better.
 
THD is a form of non-linear distortion. It takes into account harmonics related to input signal as they appear in output spectrum. In other words, if input signal contains a prominent frequency of X Hz, then frequencies of 2*X Hz, 3*X Hz, 4*X Hz and so on are checked and counted towards distortion. So, rms voltage of these harmonics is divided by rms voltage of a prominent frequency and expressed in percent. Lower results are better.
 
THD+N is very much like THD, but it checks for all harmonics that appear in output spectrum, not only for related to input signal. Thus, it takes into account other noise sources as well. Like THD, lower is better. Note that flat-weighted SNR and THD+N are basically the same, with the only difference that the first is expressed in decibels while the second is in percent.
 
There are two distinct kinds of IMD, by SMPTE (Society of Motion Picture and Television Engineers) and by ITU-R (International Telecommunications Union, Radiocommunications division). The most commonly used one is that of SMPTE, so let's follow it as well. It shows how one frequency within signal given affects another within the same signal. Frequencies are supposed to be clearly low and high, correlating as 4:1 (12dB). Though the frequencies may be any reasonable, SMPTE insists on 60Hz low tone and 7kHz high tone, while DIN assumes 250Hz low and 8kHz high. I've followed with the DIN frequencies. The output signal is checked for modulation of the high frequency by the low one, i. e. an area centred at the high tone is presented as consisting of multiple bands as wide as the low tone, then rms levels of all bands except the central one are summed up, divided by rms level of the high frequency and expressed in percent. Lower is better.
 
Having finished with theory, let's get to actual cards.