diff --git a/stepmania/TODO.glenn b/stepmania/TODO.glenn
index 6ce79b0591..5e44d9bb98 100644
--- a/stepmania/TODO.glenn
+++ b/stepmania/TODO.glenn
@@ -332,3 +332,52 @@ Instead, walk through all combinations of 4; 11110, 11101, 11011,
 
 Fix the hold note strangeness first, though.
 
+********
+
+Texture cache.  Decompressing, resizing and occasionally dithering
+images takes a good amount of time (most of the load time).  Do all
+this and write the postprocessed data to disk; read it and load it
+in a simple block load later on.
+
+
+Compressed textures.  It gives a decent performance improvement.
+Also, if designed correctly, the texture cache can write compressed
+images to disk; then load times are even faster--we load a small
+block of data from disk and load it right into hardware with no
+decompression or compression overhead.  (see
+http://developer.nvidia.com/docs/IO/1327/ATT/ARB_texture_compression.pdf)
+
+Keep track of the available texture compression formats.  If it changes,
+invalidate the texture cache.  1: we might not be able to use the
+compressed images; 2: we might be able to compress better/higher quality.
+
+Also, some mechanism to cache song images during the song scan,
+so we don't spend time doing it in the music wheel or gameplay.
+
+This can give nearly the speed of 16bit textures: with a 16bpp framebuffer
+it gives me about 10fps; with 32bpp, about 20.
+
+Do this after the new OpenGL code is stable.
+
+********
+
+Note that a 16bpp framebuffer bit depth can give extremely high quality
+output; it's usually only worse than 32bpp if you're looking for it.  (It
+becomes more important with heavy filtering; for example, Q3's transparent
+layered smoke puffs.)  Using 32bpp (or compressed) textures and a 16bpp
+framebuffer will dither in hardware, which is much higher quality than
+dithering the texture itself and is free.
+
+Jumping to a 32bpp framebuffer is *much* more expensive than jumping
+to 32bpp textures, and doesn't improve quality nearly as much.
+
+ex. in the graphics menu, 1024x768, vsync off, I get 170 fps in 16 bpp
+textures, 16 bpp framebuffer.  Jumping to 32bpp textures costs 20 fps.
+32 bpp framebuffer costs 62.  Both costs 74.
+
+(Compressed textures is even faster: 177 fps in 16bpp framebuffer.)
+
+We're more fillbound than most apps: we don't draw very many polys,
+but backgrounds can result in multiple fullscreen passes.
+(Okay, this isn't a TODO ...)
+