What do you mean? I thought accumulation buffers are deprecated
It's just a simple color render target, nothing fancy about it, you just blend each lighting passes result into it. (You accumulate lighting results into it, thats why I called it a accumulation buffer.) You don't even need a separate render target for this, you could blend each passes result into the main backbuffer if it's setup right. (I wouldn't recommend it, but it's certainly doable.)
Also, can you explain the next sentence from the article?
Quote
To reduce overdraw and improve performance, we render a depth-only pass first. This standard technique ensures that all subsequent lighting passes occur only on visible pixels.
Does this refer to the shadow-map pass or the main pass?
It refers to the main pass, they basicly render the whole frame with color writes turned off and with the pixel shader set to null, before any lighting is calculated. After this the depth buffer is filled with the scene's "final" depth values, so when later you're rendering your scene, the GPU can reject pixels before the pixel shading pass whose depth value would be bigger than what's in the depth buffer. (Given that it's guaranteed that you won't be tampering with the depth value in the pixel shader, which you can do, but usually is a bad idea.) Because you basicly render the scene twice, it adds overhead quite obviously, and this overhead can be bigger than the gain from not shading and outputing occluded pixels. If you leave out the mentioned prepass, order your objects that you want to render front-to-back, and render them in that order, you can have similar gains, but without rendering everything twice. It's important to mention that all of this only works for opaque objects, handling transparent object are a whole new level of pain.
If all this is still confusing, you can download the contents of the CD that was given with GPU Gems from here:
ftp://download.nvidia.com/developer/GPU_Gems/CD_Image/Index.html
It looks like this chapter of the book had sample code, seeing that would probably make it easier to understand what they were doing.