development in the 1990s, IBM Power Systems served databases. They crunched big
data for big business better than anyone else in the industry. But so these
systems would support the boom of mobile and cloud computing – not to mention social
media and its unstructured data ilk – IBM decided to open Power 8 technology up
to the world via the OpenPOWER
exactly are founding members such as Google and NVIDIA getting access to?
answer is speed. Power 8 is loaded with up to 12 cores running up to 4.15GHz – which
required a 4x increase in memory bandwidth over Power 7 to keep all that muscle
fed with different things to do,” said Fadi Gebara, senior manager and Master
Inventor at IBM Research-Austin.
|Fadi Gebara (middle) at IBM IMPACT Conference 2014|
processor utilization is also improved by a lower cache and memory latency
(wait time). A processor can’t run 100 percent of the time because it has to
“check” its memory for needed data. But 128 MB L4 cache on DRAM and up to 230
GB/s per socket of sustained memory bandwidth (supported by 400 GB/s at the
DRAM chips) means no stalling. By comparison, Intel’s XEON processor only
manages 37.5 MB of cache and 85 GB/s of memory bandwidth – on 15 cores.
Power 8’s greatest advancement is in its Coherent Accelerator Processor
Interface (CAPI) interface. It allows others to build new systems on top of
Power 8. For example, CAPI can be made to quickly access and analyze unstructured
NoSQL data stored in Flash memory.
unstructured social media data stored in the cloud and being transferred across
mobile devices? CAPI treats that Flash, NoSQL-stored information like
traditional, or “slow,” memory. That translates to Power 8’s 192 threads and
24-to-1 memory density advantage over x86 doing more work, faster and at a
lower cost for many cloud-based NoSQLs. But perhaps more importantly than the
raw speed at a lower cost, is what can be done with that information.
“In the past,
we’ve rolled out a new Power system every one to three years. Power 8 lets
anyone develop a new system, and innovative applications, at their speed. This
is the advantage of Power 8 and OpenPOWER,” said Peter Hofstee, Hybrid and
Workload-Optimized Systems engineer and Master Inventor at IBM Research-Austin.
Power 8 open to ideas
OpenPower member and search giant Google showed off its Power 8 motherboard at
the recent IMPACT Conference. Their Senior Director, and OpenPOWER Foundation
Director, Gordon McKean wrote that “we're always looking to deliver the highest quality
of service for our users, and so we built this server to port our software
stack to Power.” This means future applications we can’t imagine our personal
Google machines performing, today.
|Google's Power 8 motherboard|
a server-class processor is prohibitively expensive. OpenPOWER gives more
companies and different industries enterprise-level capability that would
otherwise cost 100s of millions of dollars to develop. Foundation members
NVIDIA is already seeing improvement in their analytics applications; and Mellanox reported a 10x increase in application throughput over their 40GbE
adapters and switches. More than 25 other OpenPOWER members are testing their
part, IBM’s cloud and Watson solutions run Power 8. And over time, the company
hopes that OpenPOWER will create inclusive, creative opportunities for what
were previously only high-end options.
Labels: big data, google, nvidia, openpower, Power8