Why CPU rng
A few reasons:
1) The CPU's random number generator can be random, based upon provably random phenomena rather than a pseudo-random number based upon mathematical manipulation.
2) There are some sources of actually-random data in a computer, although they are usually not the same strength as "provably random". An example is the jitter from disk drive events. But these sources are rapidly disappearing as physical devices towards silicon. This is the operational problem with not enough 'entropy' (aka real randomness) being available as a machine starts.
3) It's "too easy" for these actually-random sources of data in a computer to be influenced from outside the computer. Since they are not built as cryptographic devices. Whereas the random instructions within the CPU can include tamper detectors (such as for high EM fields).
4) Timing and other covert channel attacks are simpler against software than against hardware. Those attacks are also simpler against hardware not intended to be cryptographic devices than against hardware designed with covert channels in mind. It is easier in hardware to build a black box where all instances of the instruction take the same time to complete, use the same power, and so on. (As an aside the current issue with CPUs is that the care of design needed to defeat covert channels done for the RDRAND instruction needs to be repeated throughout the CPU design for other instructions.)
These reasons explain the last line of Ted's LKML e-mail: "Note: I trust [Intel's hardware instruction] RDRAND more than I do Jitter Entropy [from the computer's hardware devices]".