2012 in supercomputing: Ceepie-geepies, a weak ARM and the need for speed
Plus: How are algorithms like your ex?
HPC blog A lot has happened in HPC over the past year. I would say that the speeding up of development of accelerators and the rising number of hybrid CPU+GPU systems are probably the most noteworthy trends we saw in 2012. Over the next year we’re going to see even more use cases for hybrid systems, and I expect to see much wider use in commercial enterprise computing as well.
We’re also going to have a front-seat view of a truly brutal dogfight between Intel and NVIDIA for accelerator market and mind share. NVIDIA has the edge in performance and price/performance, plus real-world results. But Intel’s Phi has that old, familiar x86 architecture that allows it to run almost every app right out of the box – although optimising the app to take full advantage of Phi might take some effort.
HP’s Marc Hamilton is also of the view that hybrid CPU+GPU systems and development in accelerators were important trends last year. Marc leads the HPC group embedded in HP’s Hyperscale Business Unit, working with HPC engineering, marketing, and sales organisations. In "Top 10 HPC Events that Mattered in 2012", he runs down what he feels are the most significant HPC events of the last year. He placed accelerators on the top four spots on his list, starting with Intel’s Phi introduction at #1 and NVIDIA’s Kepler feature set coming in second, third, and fourth.
For another view on 2012 HPC happenings, take a look at Andrew Jones’s two-part “Review of 2012 in Supercomputing”. Andrew is Vice President for HPC at NAG (Numerical Algorithms Group – an HPC software/services business, not an organisation of exes or old horses). He is a frequent blogger on supercomputing issues, and in this entertaining entry, he not only reviews 2012 but also looks back on his predictions for the year and their accuracy.
It looks to me like he batted about .500 on his 2012 prophecies. He scored hits on predicting that “Big Data” would be the premiere marketing buzzword in 2012 and on saying that the GPU wars would heat up. He also predicted “ARM would become a common part of the architecture debate alongside x86 and accelerators,” which arguably happened, but ARM isn’t quite there yet. I don’t think ARM will be regarded as a true alternative for HPC until we see 64-bit ARM, which is due sometime in 2014.
I think Andrew blew it big time when he predicted “The use of 'cloud computing' as the preferred marketing buzzword used for large swathes of the HPC product space would come to an end.” I still can’t swing a medium-sized cat at any industry event without hitting at least a few people bloviating about clouds, HPC in the cloud, cloud analytics, etc. etc. (Note to self: get a new cat before next industry event - existing cat is pretty much used up.)
Without taking away too much of Andrew’s thunder, he goes on to talk about the trends he missed in 2012 (“exascale race stalling,” “petaflops becoming ordinary,” plus other topics). It’s a quick and punchy read, and well worth the time. ®