2020 Processor Technology Could Unite Power And Mainframe Chips
October 20, 2014 Dan Burger
It happened once. Could it happen again? Many IT Jungle readers recall IBM‘s decision to converge the System i and System p lines, which resulted in Power Systems. A shared Power processor and common system architecture for both platforms was a business choice that made sense at the bean-counter level. And at the technology level, there would not be a lot of blood on the floor. The savings over the long haul would be the salve that healed all wounds.
There was a time when the convergence of those two lines was unimaginable, much like the conventional thinking today about the Power and mainframe lines. Once you get past the “it could never happen” notion and consider the i and p convergence along with the end point of current processor technology, and then factor in the bean-counters, the pieces start coming together for a possible single-minded, single-processor future for IBM.
Preposterous? Well, hold on a second. Let’s ask Ross Mauri, general manager of the System z division and the guy who was the Power Systems general manager at the time of the i and p convergence. Mauri and I had a conversation earlier this month at the IBM Enterprise2014 conference, where IBM’s $1 billion investment in processor research and development was mentioned more than once during the general session appearances of top executives and research scientists.
In the mid-1990s, Mauri was leading the mainframe development when processor technology made a moonshot-type of leap from bipolar chips to CMOS chips, which was made necessary because scientists had reached the threshold for squeezing more performance from bipolar technology. This technology was just too hot and too expensive to be more broadly commercialized, and required water cooling on top of that. But just as bipolar chip technology couldn’t be pushed beyond its limit, we are now approaching a similar ceiling with CMOS, a technology that has served the industry well for right around two decades.
“When we moved from bipolar to CMOS, we knew CMOS would tail off. But it seemed so far in the future,” Mauri recalls. “Now there will have to be investment in the next big thing and so it’s interesting to consider whether the Power chip and the mainframe CMOS chip will merge into one chip for both platforms.”
Mauri is not suggesting this is right around the corner. You won’t see CMOS technology replaced in the next generation of chips, or even the one after that. But the time is coming.
“I don’t know what is next in that 2020 time frame based on the technologies we are looking at today–anything is possible,” he says. “We haven’t backed off yet from CMOS. But if we have to shift off CMOS, it will be a big transition for us and for the industry. It means a new way to build the fundamental building blocks for chips and that will affect the packaging and the interconnect with other pieces.”
Technologies that are getting the most attention include quantum computing, carbon nanotubes, graphene, and light beam computing. But something could come out of left field to trump those leading candidates.
IBM, Mauri says, is investing in the multiple technologies that may or may not lead to the next big breakthrough on performance and density and electrical characteristics. But that’s not the only criteria. Manufacturing costs have to be considered.
“If the manufacturing costs are astronomical, it’s not going to be viable as a business,” he says. “So we are betting on a variety of outcomes because we don’t know which will be the winners. Anyone who makes chips out of CMOS–Intel and the others–are looking at the same problem. We are all being affected by the flattening of the CMOS performance curve.”
The migration to whatever replaces silicon chips does not mandate a one chip for all and all for one chip inevitability. With the current Power and mainframe chips, the architectural the floor plan is very different. In the future, it’s possible the architectures would become homogenized.
“New technologies might make it easier or they might not,” Mauri says. “But if you are considering new system design, you might look at convergence.”
It’s not the technology that drives the idea of converged systems, in this case Power and mainframe. It would only be done for business reasons. That’s just me thinking, but Mauri does not disagree.
“Look back to AS/400 and RS/6000,” he says. “Two great but different architectures, both built by IBM teams. One was built for very technical computing and the other was built as a beautifully integrated, simple to use, simple to run business system. But in the end, for business reasons–economy of scale–we converged the underlying hardware.”
“The challenge with a system convergence is that you have to serve both well,” Mauri says. “And there are always trade-offs.”
And when there are trade-offs, the folks who build the systems, sell the systems, and use the systems on one side or the other believe they got the short end of the stick.