IBM Power Systems built with the POWER8 architecture and processors J. J. Cahill T. Nguyen M. Vega D. Baska D. Szerdi H. Pross R. X. Arroyo H. Nguyen M. J. Mueller D. J. Henderson J. Moreira This paper describes architectures and significant implementation features of two systems in the IBM POWER8i processor-based family of servers. Specifically, the scale-out 2-socket rack server and the enterprise scale-up 16-socket rack server are detailed. The description of these systems highlights the increase in memory bandwidth from previous POWER A systems, the enablement of coherent accelerators, the highly-extensible I/O subsystem, and the high-performance directly attached storage subsystem. In addition, reliability, availability, and serviceability features are described. These systems deliver significant increases in core count, memory, and input/output bandwidth over previous POWER systemsVwith reliability and availability enhancements commensurate with the performance improvements. Introduction At the time of the publication of this paper, the IBM POWER8* processor-based family of servers consists of two groups of systemsVthe scale-out Power S812L, Power S822L, Power S824L, Power S814, Power S822, and Power S824 (all one-socket or two-socket rack servers available in 2U and 4U heights) and the enterprise scale-up Power E870 and Power E880 (4-socket to 16-socket servers). For descriptions of servers of this type from IBM_s competitors, consult [1–5]. POWER7* processor-based servers were described in [6]. Both IBM server groups are designed to satisfy demands for increased performance, reliability, availability, and serviceability in smaller power, cooling, and spatial footprints. The two-socket server can have up to 24 cores. It can also be configured with only one socket and as few as six cores in the system. Figures 1 and 2 show an isometric view of the 2U and 4U scale-out Power S8xy server (where x is the number of sockets, and y is the form factor U value, and these indicate the various variants of the scale-out Power S812L, Power S822L, Power S824L, Power S814, Power S822 and Power S824 server) in both 2U and 4U form factors. The 16-socket server (referred to in this paper as Power E8z0 where z is either 7 or 8) is implemented as a scalable server that starts with a minimum of 4 sockets and scales in 4-socket increments up to 16 sockets for a total of 192 cores in the system. Figure 3 shows an isolated view of the four-socket CEC (Central Electronics Complex) drawer and the External I/O drawerVand the rack view of a system configured with four CEC drawers and three External I/O drawers, and seven additional External I/O drawers in an additional rack. Reliability, availability, and serviceability By design, the POWER8 processor delivers a significant increase in performance. This increase in performance also increases the proportion of an enterprise_s computing needs that could reside on a single computer. This demands increasing reliability and decreasing downtime from such a computer, even a scale-out two-socket server. The POWER8 processor builds upon a strong base of RAS (reliability, availability, and serviceability) features found in the predecessor POWER processors. These include substantial soft-error avoidance and recovery techniques for processor core and caches, as well as features to minimize the impact of persistent uncorrectable failures. The systems use a dedicated service processor to orchestrate advanced error detection, fault isolation, and recovery. Overall system reliability is also enhanced in POWER8 processor-based systems by integrating new functions into the processor module such as an on-chip controller for power and thermal monitoring and the flexible I/O subsystem interface. In addition, the new coherent accelerator interface is designed ÓCopyright 2015 by International Business Machines Corporation. Copying in printed form for private use is permitted without payment of royalty provided that (1) each reproduction is done without alteration and (2) the Journal reference and IBM copyright notice are included on the first page. The title and abstract, but no other portions, of this paper may be copied by any means or distributed royalty free without further permission by computer-based and other information-service systems. Permission to republish any other portion of this paper must be obtained from the Editor. J. J. CAHILL ET AL. 4:1 IBM J. RES. & DEV. VOL. 59 NO. 1 PAPER 4 JANUARY/FEBRUARY 2015 0018-8646/15 B 2015 IBM Digital Object Identifier: 10.1147/JRD.2014.2376132