test-suite for epiphany

[piotr5]

at his talk andreas said a chip with thousands of cores wont have 100% of them working correctly.
as far as I understood usually chip-developers test chips and sell only the ones which work, for the price of those which don't.
seems adapteva plans to sell all the chips, even if parts of them don't work correctly. (sounds alike to what hdd-industry does.)
and they expect the users (rather os) to check what cores on their chip are useful.
unfortunately, neither software-developers nor linux-os developers know what to test.
there is a nice program called memtest. it tests memory on various ways. and it claims each test must repeat several times.
obviously knowledge from hardware development and fault-analysis did define what tests must be done and how.
but for 1024-core parallella we need to test an actual processor-core as a whole, not just memory!
how is this done? if some test fails, can the core be used for programs which won't use the failing component?

so, I'm asking for a list of things that need to be tested in a power-on-self-test of epiphany.
and I'd like that list to contain explanations which assembly-sequences to avoid if the test failed.
maybe start with this program "memtest" in explaining how to avoid the various faults it is testing for...