SANTA CLARA, Calif.--Server design--and by extension microprocessor
and component considerations--are now driven by new-media companies
such as Facebook and Google, whose system specifications render
traditional benchmarks useless.
That was the word this week from panelists at
TechCon here exploring the future of big computing and
its impact on semiconductor design.
"J.P. Morgan is one of the largest consumers of compute on earth,"
said Andrew Feldman, CEO of AMD acquisition Seamicro and now head of
AMD's Data Center Server Solutions. "They're one of the top 10
buyers of CPUs. They're a 130 year-old company."
"In Facebook's fifth year they bought the same amount of compute--in
their fifth year," he added. "It's at that type of company where the
battle is going to rage. That's where price-per-unit-compute,
compute-per-watt will move to the fore. Those are the dimensions the
battle will take place on. And they'll take place...not over
abstract benchmarks, but the exact work that these companies have
The panel took place against the
of AMD allying with ARM to drive into the server market
in the coming years, where power consumption has become an enormous
issues in massive server farms that drive digital commerce. There,
the question arose as to what the role of traditional benchmarks is
in a world where power consumption can make or break a deal.
"Traditional benchmarks are about defining headroom and performance
per dollar," said Karl Freund, vice president of marketing with
ARM-based server company Calxeda. "In the new world, it's not
performance per dollar; it's supporting number of users at a given
service level at the least capital and operationg expense. In that
type of world, there's a smaller role for benchmarks."
"Traditional enterprise guys are not going to look at ARM first," Feldman said. "The
people who will use ARM first...almost every one of them uses
computer to manufacture profit. If you use compute to manufacture
profit, you are always and every day interested in how you can
improve the efficiency of your ability to manufacture profit. That
class of customers will move first. Customers who use compute to do
IT will move last. If you look at those two markets, those who use
compute to generate profit are growing many hundreds of times faster
than the other segments of the market."
Benchmarks will always be used as long as CPUs are sold. The type of benchmarks will change and the traditional compute oriented benchmarks will be replaced with more system and IO throughput ones.
There is a need for a suite of benchmarks that not only measure performance but also power. A spec-power for the new workloads - maybe a cloud-power benchmark suite that measures throughput in transactions/sec/watt.
And can all the self-respecting CPU vendors stop quoting DMIPS?