IBM Scores $325 Million Power Supercomputing Deals With DOE
November 17, 2014 Timothy Prickett Morgan
The Power Systems ecosystem just got a huge shot in the arm now that IBM has landed a $325 million contract with the U.S. Department of Energy to build two new massive supercomputers, the largest of which, called “Summit,” could scale to as much as 300 petaflops of aggregate number-crunching performance.
The Summit system will be installed at Oak Ridge National Laboratory, home to Titan, one of the largest supercomputers in the U.S. and indeed one of the most powerful machines in the world. The Sierra system will go into Lawrence Livermore National Lab, where IBM sold its first prototype BlueGene system and several subsequent models as it developed and expanded that massively parallel system, based on a modified set of Power processors. About 90 percent of the floating point performance in the systems will come from future generations of graphics coprocessors from Nvidia, code-named “Volta” and packed with huge amount of performance and local memory. They will link to future Power9 processors, and yes this is the first time that IBM has mentioned Power9 publicly that I can think of and so we now know that IBM has a Power9 chip in the works. The Summit machine is slated to come in at 150 petaflops, and could be expanded to 300 petaflops. The machines are expected to be installed in the 2017 to 2018 timeframe, so we also know when to expect Power9 chips as well.
IBM is collaborating with Nvidia to create a special point-to-point interconnect that can link GPU coprocessors to each other with very fast links running at 20 GB/sec, and allowing multiple links between devices. This interconnect is called NVLink, and it will also hook directly into the Power9 processors in a manner that is similar to, and yet distinct from, the Coherent Accelerator Processor Interface, or CAPI for short, that rides atop that PCI transport between the processor and the PCI bus. CAPI is a subset of PCI-Express and NVLink is a superset is the way IBM’s techies have explained it to me. As I have said before, I would not be surprised to see IBM rip the vector engines out of future Power processors and compel customers to use GPU coprocessors to do the math and thereby have room to either crank the clock speed on CPU cores, add more cores, or a little of both.
The new machines will be examples of what IBM is calling “data centric computing,” and I am still puzzling out what this means precisely, but I can make you laugh. IBM is, I think, talking about building a system with a strong central processor with lots of intelligent I/O processors that can help prechew data and manage functions without direct CPU control.
Does any of that sound familiar to any of you? Any of you? Remember the AS/400 before the iSeries convergence? Do you remember that?
Plus Ã§a change, plus c’est la mÃªme chose.
Anyway, I am going to ponder this and what effect it has on the IBM i customer base. The point is that this is good news. So be happy.