IBM Quintuples Performance with the Power 795
August 17, 2010 Timothy Prickett Morgan
The big, bad box of the Power7 lineup rolls out onto the floor today, and it just might sag a little bit. The Power 795, as the top-of-the-line Power Systems server is called, crams up to 256 cores and 1,024 threads into a single system image backed by 8 TB of shared memory in which applications can frolic. The Power 795 has four times cores and roughly five times the raw processing capacity as the top-end Power 595 machine that dates from the summer of 2008 and that was looking a bit long in the tooth.
Like its Power 595 predecessor and the past several generations of IBM System z mainframes (including the zEnterprise 196 machine that was announced in July), the Power 795 comes in a non-standard 24-inch-wide system frame instead of a standard 19-inch rack. This extra space gives the box more room for processors and memory, but more importantly allows for more air to be sucked over all of the components to cool them.
Unlike the System zEnterprise 196 server, the Power 795 does not have a water-cooled option; according to Jeff Howard, director of marketing for the Power Systems division, IBM looked at this option and decided that customers didn’t want it. There is, however, a 520-volt DC power option as well as a 480-volt AC option for the Power 795 box if you want to take a few steps out of the electricity transformation process running from the data center input over to where it feeds into the servers. Many companies are trying to save energy by keeping the voltage high and stepping it down less frequently.
The basic design of the Power 795 server is the same as the Power 595. You have eight processor books with processors and main memory on them, and you glue them together into a giant symmetric multiprocessing (SMP) machine. But the big difference is that the Power 595 was based on the dual-core, 5 GHz Power6 processor (there never was a Power6+ kicker, even though it was planned) while the Power 795 is based on the six-core and eight-core Power7 chips. Machines using the six-core Power7 chips fill in a gap between the Power 780 and the larger Power 795 using the eight-core chips, just like the 32-core System p590 used to fill in the gap in the System p lineup between the 16-core Power 570 and the 64-core System p595. (There never was a System i variant of the 590 box for reasons that IBM never explained.)
There are two variants of the Power 795 machine. The smallest one is based on the same six-core, 3.7 GHz processors that are used in processor feature cards elsewhere in this year’s Power Systems lineup. Each processor book has four processor sockets, which yields 24 cores per book; you can plunk from one to eight of them in a Power 795 system to create a machine that scales from 24 to 192 cores in a single system image.
The next size up of the Power 795 comes in the same MaxCore and TurboCore configuration that makes a Power 780 different from a Power 770. In MaxCore mode, the Power 795 uses eight-core Power7 processors running at 4 GHz. You put four of these on a book, for 32 cores, and eight of them in a system, for 256 cores. Now, let’s say you have a database workload that might do better with fewer cores running at a higher clock speed and with more L3 cache and main memory dedicated to those cores. Then you can go into the Power 795, flip a setting in the firmware, and reboot the box into TurboCore mode. When it reboots, only 16 cores are turned on in each book, but they now run at 4.25 GHz instead of 4 GHz. And, more importantly, twice as much L3 cache and main memory are dedicated to each core.
Generally speaking, a 4.25 GHz core running in a Power 795 is yielding about 15 to 20 percent more performance than the faster 5 GHz Power6 core running in the Power 595. The reason is that the Power7 processor has 32 GB of embedded DRAM (eDRAM) cache memory right on the chip instead of having the 36 MB of off-chip (but on package) L3 cache that the Power6 chips had. When you add that per-core performance advantage up over 256 cores, you get a factor of five improvement in overall system throughput–and within the same power and thermal envelope.
The Power 795 server supports from 32 GB to 8 TB of main memory and has four GX++ remote I/O peripheral slots per processor book, for a total of 32 slots. That allows up to 32 12X-based PCI-Express drawers to hang off the Power 795, for a total of 3,052 disk drives. The box does not have internal disks or peripherals–everything goes into the remote I/O drawers.
The Power 795 can run the i 6.1.1 and i 7.1 operating systems; AIX 5.3, 6.1, and 7.1; Red Hat Enterprise Linux 5.5; and Novell SUSE Linux Enterprise Server 10 SP3 and 11 SP1. The Power 795 will be available on September 17, as will the new AIX 7.1. Pricing on the Power 795 was not available as we went to press.
Just a reminder: If you want to listen in to Power7 server announcements today, IBM is hosting two Webcasts, as follows:
We will be drilling down into the Power Systems announcements further in next week’s issues of The Four Hundred and Four Hundred Stuff.