> How about turning this on its head and installing a Phi directly on > the mainboard, with lots of memory attached (unless it has some > in-built limits, haven't checked...)
Phi is implemented similar to a GPU: it has a very wide interface to soldered-on gddr5 ram. gddr5 channels are 32b wide, and it has 16x, which for around 300 GB/s works out to a gddr5 performance of 5.860 Gt/s. AFAIK, gddr5 can't tolerate multiple banks (so that's 16 chips, each obviously 4Gb, which is the largest in production.) > only for I/O - for which it should be fast enough. To avoid changing > the architecture too much, attach the whole thing to an Atom CPU - low > power, but fast enough to handle the I/O and to control the > computation on the Phi. no, the atom would be fairly pointless, since the Phi already has plenty of cores. as far as I know, you could plug the Phi into an otherwise hostless PCI backplane - it would mainly just need a BIOS like a normal PC, and it would probably have to do some emulation for that, since I think I remember reading that Phi doesn't implement the same set of modes as a "normal" x86 (good old real mode, for instance.) in short, I think it's a question of pin counts and when higher-density gddr5 chips will arrive. _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
