3 ms·
> The vector multiply-add operations are the most memory bandwidth intensive of any part of BiCGSTAB. Each such operation reads over a billion values from memor
by tomohawk 6y ago
> The vector multiply-add operations are the most memory bandwidth intensive of any part of BiCGSTAB. Each such operation reads over a billion values from memory and writes a half billion values to memory. This requires three memory accesses per scalar multiply-add operation. The CS-1 has full bandwidth to all of its memory: it can load two values and store one value for every multiply-add performed on every data-path lane, on every core, on every machine cycle.
wow
- modeless 6y agoThat's crazy. The entire 18 GB is directly adjacent to ALUs with no cache hierarchy? Also: "the on-wafer interconnect can send out one 32-bit pair of 16-bit words to each of a given core’s four neighbors on every machine cycle and can take in one 32-bit word on the same cycle. [...] supercomputers require a ratio of perhaps one word communicated per 1,000 floating point operations."