Here are some speed results I got on my dual AthlonXP 1600+ for a test song, Green Day - Burnout. I used both the original wavefile and a copy downsampled using SSRC to 22050Hz. Unless otherwise noted, the encoder downsampled the 44khz original on-the-fly.
I used Lame v3.92, Fastenc v1.02, Gogo v3.10, wm8eutil SSRC v1.29.
format: encoder & settings: speed(22khz input), speed(44khz input)
-------------------------------------------------------------------
lame --alt-preset 56: 12.1x, 7.1x
lame --alt-preset cbr 56: 11.2x, 6.8x
lame -b56: 20.0x, 9.3x
lame -b56 --lowpass 10: 22.6x, 9.9x
fastencc -br 56: 72.8x, 45.8x
fastencc -br 56 -hq: 43.2x, 37.2x
gogo -b 56 -cpu 1: 60.5x, 39.6x*
gogo -b 56 -a -cpu 1: 59.6, 41.2*
gogo -b 56 -cpu 2: 102.1x, 69.8x*
oggenc -q0: 12.8X(55kbps), 7.1x(66kbps)*
oggenc -b56: (fails), 5.8x*
wme8util -a_setting 48_32_2: ~15x (44khz input only)
wme8util -a_setting 64_44_2: ~10x (44khz input only)
(*)--These files did not downsample to 22khz
Sound quality comments: Vorbis -q0 at 22khz was by far the closest-sounding to the wavefile, with the highest lowpass and fewest wishy-sounding artifacts. Next came lame --alt-preset, then fastenc. --ap had a higher lowpass and fastenc had a bit less of the wishy-sounding artifacts, but in my opinion --ap was better overall. Then came Gogo, closely by lame -b56. Both Vorbis and Gogo sounded worse than the others at 44khz, except for WMA, which had an incredibly painful metallic nails-on-chalkboard sound to it.
The problem with Vorbis and Gogo is that I can't figure out how to downsample on-the-fly; I found no documentation showing me the switches.
If your friend is willing to downsample the audio sources beforehand, there's a chance his machine might run fast enough to handle Vorbis or Lame --alt-preset. Otherwise, Gogo might be the only available choice.
ff123 has done more comprehensive (and valid) testing of audio quality at 64kbps using many different samples, which you can read
here. However, he didn't post encoding times (since that wasn't his focus). I've started doing the wavrate evaluation of his samples, and so far on newkid and wait, my results correlate fairly closely to his.