This article is more than 1 year old
IBM: Why our Power9 CPU is going to make data centers great again
Big Blue starts making performance promises
HPC blog Rani Borkar, IBM’s vice president of OpenPOWER development, gave attendees at the OpenPOWER Foundation Summit in Barcelona at the end of October a sneak peek at the vendor's upcoming Power9 processor. I captured it on video (see below) to show you what to expect.
The first Power9, due in 2017 (probably mid-year but IBM didn’t disclose specific dates), will feature 14nm finFET technology with 24 cores.
The new CPU is designed to be a platform for hardware acceleration – and there are a lot of ways to attach accelerators to it. First, it will have PCIe gen 4, which is not that remarkable since everyone can be expected to have it in this time frame.
Where IBM ups the ante is with the inclusion of its second-generation CAPI 2.0 and OpenCAPI. This allows for high-speed (25Gbps) direct coherent connections to compute accelerators, storage and network adapters.
The chip will also feature the next generation (2.0) of NVIDIA’s NVLink technology, which should surpass the 80Gbps bandwidth of today’s version.
Power9 won’t be a single processor with different packaging options. IBM will be offering versions of the chip: one aimed at scale-out markets like HPC, and the other designed for the scale-up market. The scale-out processor will offer direct attach memory with up to eight DDR4 ports and commodity packaging, meaning it won’t have the CPU parts that are necessary to build a larger than two-way SMP system.
The scale-up chip will offer buffered direct attach memory and some more link lanes, and will have all the logic and transistors needed to handle large SMP configurations.
There will also be different chips for running either Linux or traditional Power ecosystems.
The presentation included some performance comparisons between the Power8 processor and the new Power9. On a core-to-core basis, the Power9 outperforms Power8 by a factor of at least 50 per cent (on floating point), up to 225 per cent more on graph analytics workloads.
This is a huge performance increase for the industry. We’ve grown used to 30-40 per cent generational improvements on x86 based processors, so if Power9 offers 50-100 per cent or more improvement (depending on workload), it’s going to capture a lot of customer attention. ®