Unmeasurably small - all it's doing is grabbing the first two bits from each of the 16 sectors, storing them in the page header, and then writing an alternating bit pattern into the two bits in each sector. And the reverse when the page is read again. Almost nothing compared to what a page checksum is doing by reading the whole page and adding the contents to a 'checksum'.
However, torn-page detection does not detect corruption within sectors - only when a write fails some way through. On 2005 onwards you should always have page checksums enabled for this extra benefit - 1-2% CPU shouldn't be an issue (otherwise you've got other problems).
Some blog posts around this: How to tell if the IO subsystem is causing corruptions? and Inside The Storage Engine: Does turning on page checksums discard any torn-page protection?.