IBM launches Power8 chip architecture for next-generation servers

April 2014

Recently IBM launched its new Power8 chip architecture for next-generation servers, which it describes as “a sliver of silicon that measures just one square inch…embedded with more than 4 billion microscopic transistors and more than 11 miles of high-speed copper wiring.”

Maybe more significantly, IBM is announcing that the OpenPower Foundation — a group that includes Nvidia and Google — is taking the first steps toward making systems based on Power8 available.

The OpenPower foundation is taking an approach similar to ARM, which powers virtually all of the world’s mobile devices, by making its hardware, software, and intellectual property available for licensing.

But success is far from a given. “It will take significant investments and time to approach the level ARM enjoys today as an open architecture,” said Patrick Moorhead, president and principal analyst at Moor Insights & Strategy.

IBM Power8 Processor HPC Architecture Detailed

The technical details presented at the Hot Chip 2013 conference by IBM showcased the Power8 Processor featuring 12 cores with the new 8-Way simultaneous multithreading (SMT) compared to the 4-Way SMT in Power7 processors. Being the most high-performance chip that IBM has to offer, the components are scattered across a 22nm die featuring silicon-on-insulator technology spread across an area of 650mm²which is quite huge.

The Power8 Processor also features a 16 Execution Pipeline followed with 64K data cache per-core and 32K instruction cache. On the cache front, the new processor has 512 KB SRAM L2 cache per core, 96 MB eDRAM shared L3 cache and also 128 MB of eDRAM L4 cache that’s situated off-die unlike the on-board eDRAM on Haswell processors featuring GT3e iGPUs.

“The entire Power-based ecosystem has lost many supporters in the industry like Freescale, AppliedMicro, Xbox and Playstation, so Power 8 is very important to get right,” said Patrick Moorhead, an analyst at Moor Insights & Strategy. “It will be a real challenge to fight both Intel and AMD in this space given their scale, and very important to IBM given they are banking on Power architecture for the their platform future. We believe IBM will exit the X86 server market and sell their line to Lenovo, making Power 8 critical to IBM’s future success.” IBM

Memory support is offered in transnational memory and support for Crypto & memory expansion. The chip offers a total bandwidth of 230 GB/s. An On-Chip power management micro-controller is used for regulating power and voltage delivery to the chip while each core has its share on integrated voltage regulator modules (VRM) which is yet another similar approach as Intel’s Haswell processors.

Along with its tremendous design, Power8 processor would also use a Coherently Allocated Processor Interface or CAPI for short featuring coherent memory addressing for CPUs and external co-processors such as NVIDIA’s Tesla or the Intel Knight’s Landing and Ferry. There’s no mention of TDP at the moment but IBM has been offering their previous Power series processors with TDPs of around 200W, with the new 22nm architecture in play the TDP should drop down a bit while improving the performance tremendously over Power7.