Earlier this summer I wrote about the upcoming Power7 processor, including some of the major features. The Cliff Notes:
- Four, six, or eight cores
- Four computing threads per core, for a total capacity of 32 computing threads per chip
- DDR3 memory
- Due out by mid-2010
Well, at the HotChips conference at Stanford University last week, IBM revealed some more details for the upcoming processor. Though I wasn’t at the show, I was able to talk to Ron Kalla from IBM, who along with Bill Starke was giving a presentation at the show about the processor.
Though Kalla wouldn’t reveal what frequency the chip will run at — which is the detail IBM has been most reluctant to part with — he did say that it will be lower than Power6. This presentation given by IBM indicates Power7 will run at 4.04GHz, but the range is still unclear.
“We did go to a lower frequency to get a better power optimization point,” he said.
As a result, he said, customers running Power7-based systems will have “basically identical power consumption” to those running Power6-based servers. IBM is pushing this talking point a lot, because it wants as many Power6 users as possible to upgrade to Power7, and this could be an incentive.
This marks a clear departure with IBM’s push behind Power6, which was centered around clock speed. Back then, IBM bragged that the chips were hitting close to 5GHz in the labs. But experts at the time questioned whether cycle speed was the ultimate and only attribute of a good processor.
“The industry, Intel, was going to get away from talking about clock speeds and start talking about heavy lifting, like multi-threading,” one analyst, Charles King, said.
“Is a four-gig Power6 processor possible? I’m sure they could do it,” another analyst, Clay Ryder, added. “But I’m not sure the market is leaning toward that.”
Now, power-performance is the new metric, rather than just simple clock speed. In addition, IBM wants to push Power Systems as a good solution for Web-serving applications that can handle many users at the same time. Many cores accommodating many virtual machines may be more beneficial to that end that super-high clock speeds.