[Beowulf] Chinese supercomputers to use ‘homemade’ chips
hahn at mcmaster.ca
Sat Mar 12 11:48:36 PST 2011
> some more details about the chips.
seems like mostly an nth-gen rehash of the info here:
which discusses the provenance of the basic architecture.
40W gets you 8 cores at 1 GHz. each core appears to have 2x 256b FP SIMD,
but no mention is made of their rate. if the rate were 1/cycle, that would
be 32 gflops per unit. the article mentions "performance on those vector
units is 128 Gflops, so I'm not quite sure how the math goes. later in
the article it says 16x 8c chips go into a single 1U and yield 2 Tflops,
therefore 16 gflops per core. so maybe it can start a new instr every other
cycle. in any case, 128 Gflops for 40W.
a similar comparison with a mainstream cluster chip might be closer to 1
gflop/watt (at the socket level, ignoring support chips, fans, PSU efficiency.)
current-gen Nvidia and ATI chips are in the 2-3 gflops/watt range,
though that's not a completely fair comparison.
regards, mark hahn.
More information about the Beowulf