I think people underestimate how powerful low-cost DSPs have become.  For example, C. Gouvea was able to verify secp256k1 ECDSA signatures in 700 ms using an ultra-lower power 16-bit microcontroller running at a only 8 MHz (and available as in a tiny 3 x 3 mm package size).  Even very cheap ARM-core processors ($2 - $4) will typically run at ~150 MHz (and with 32-bit registers).  So let's say signature verification for each received TX takes no more than 20 ms, at which point you can relay the transaction to your other peers.  This seems perfectly acceptable to me.  

The only other issue I can think of is RAM, but I don't see why you can't store the unspent outputs in the SD card itself (new SD cards permit reads in excess of 30 MB/s so I think the time it takes to cross check the outputs in a TX with the blockchain unspent outputs stored in flash will be short compared to the signature verification times [20 ms]).  It seems you just need sufficient RAM to pool the unconfirmed transactions (and some overhead to accept new blocks, etc), so upon first glance the 6x8mm 64 MB IC for $1.50 seems sufficient for this purpose.  

I think such a bitcoin node would be useful if it could be made only a few cm in dimensions, cost less than $10-$15 excluding the SD card, and be trivial to set-up.