They are not really saving bandwidth but saving "external bandwidth" (= "external memory bandwidth requirement"), that is, bandwidth on the bus between GPU and main memory, since they are, speaking in a simplified manner, only reading texture data from main memory but render (i.e., write pixels) into the internal eDRAM. The total bandwidth that they consume does not change, it is just split between main memory and eDRAM.
That is nothing new, since it is exactly what the eDRAM/eSRAM's main purpose is. It is small enough to hold render targets ("pixelbuffers") which, for 1080p, take 1920 x 1080 x the information per pixel (in the example you quoted 12bytes) ~= 23MB (or ~11MB at 720p). Despite their relatively small memory footprint, reading and writing into those buffers consumes much bandwidth, especially if you are rendering in multiple passes; which you do in case of deferred rendering. In this technique you gather information in a first pass and store that into a G-Buffer. You then use that information in a second pass to render the actual picture more efficiently. That was actually some sort of a problem for the 360 since a G-Buffer like the one you described didn't quite fit into its 10MB of eDRAM. That is the reason why games like Halo 3 didn't run at true 720p.
Anyway, in summary, you don't save total bandwidth, you save bandwidth between GPU and main memory. Since the bandwidth of DDR3 is limited, mitigating that limit is the main purpose of eSRAM.
