Thinking About AI Power in Parallel
Up to this point, a large portion of the worry has been
about enormous upgrades in performance, here and there as high as 1,000X over
past designs. Those numbers drop altogether as the chips become less specific,
which consistently has been a factor in performance — even back to the times of
mainframes and the Microsoft-Intel (WinTel) duopoly at the beginning of PCs and
commodity servers. Performance is as yet the key measurement in this world, and
there is so much disseminated energy as heat that fluid cooling is just about
guaranteed.
windows admin |
In any case, power is as yet a worry, and it's turning into
a greater concern now that these AI/ML chips have been appeared to work. While
the initial usage of training was done on arrays of modest GPUs, those are
being replaced by more modified equipment that can do the MAC calculations
utilizing less power.
All of the big cloud providers are building up their own
SoCs for training (just as some inferencing), and performance per watt is
thought in those systems. With that much computer influence, saving power can
indicate massive measures of money, both for powering and cooling the equipment
and for protecting the life expectancy and functionality of these exascale
computer farms.
Read: What does a Work-From-Anywhere windows admin do for you?
Comments
Post a Comment