When GlobalFoundries bought IBM's semiconductor business two years ago, it signed a number of agreements to continue to manufacture IBM hardware and to provide facilities that would meet IBM's manufacturing requirements. This week, the company announced that it would deploy a new 14nm FinFET process to build IBM's z14 mainframe processors -- and it'll use SOI (silicon-on-insulator) as well.
IBM's z mainframes are quite different from the other CPU architectures we've discussed on ET over the years. These systems are designed to offer as close to 100 percent uptime as human engineering is capable of providing. They emphasize reliability and redundancy, including the ability to switch to backup hardware while the system is hot. Because these systems are bought by the largest customers, cost is typically no object: A z14 CPU (IBM refers to this as a CP) is 696 mm sq, with 6.1 billion transistors. It supports up to 10 cores clocked at 5.2GHz, with 6MB of L2 cache per core and a 128MB L3. The System Control (SC) chip that connects to the various CPs and has a 672MB L4 cache. Yes, this means a z14 mainframe has more cache than plenty of desktops not so very long ago.
It isn't clear if any other companies will use this particular process. AMD used SOI for many years, from K8's initial debut in 2003 to Piledriver and its associated APUs in 2012 - 2013. The company later moved to 28nm bulk silicon for its hardware and has since stated that it will not use custom silicon in the future. This doesn't mean the company has no access whatsoever to tweaks that improve yield or clock speed, but it's not going pay GF to build a specific process line to its own specifications.
There are also fundamental differences in wafer cost. SOI wafers are more expensive than bulk silicon, and while PD-SOI has some benefits for certain transistor designs, no mainstream vendor has adopted it for use in either the x86 or ARM markets. One benefit of using SOI is decreased power consumption, and consequently higher clock speeds, which is one reason why IBM can hit 5GHz+.
The other benefit is more straightforward: Because IBM's CPs are all water-cooled, the company can take advantage of superior cooling to deploy processors with much higher power consumption than the equivalent cores from AMD or Intel. We weren't able to find exact TDPs for z-class mainframes, but some of IBM's POWER CPUs have TDPs of 250W or more. That gives IBM more thermal budget to play with, and explains some of the clock speed differences.
