Edit: And now I wait for POWER10.
Originally posted April 2014 by IBM Systems Magazine
POWER8 technology created some buzz when it was first discussed at the Hot Chips conference and slides that describe the chips could be found online before today. But now we have more information about the actual systems that will be shipping when they become generally available in June.
When you look at the Power Processor Technology Roadmap since 2004, you can see that we regularly get new, more powerful chips. We are almost spoiled. When IBM says it is going to deliver, it does just that, with both new hardware and new OS releases.
In 2004 we had POWER5 followed by POWER5+. In 2007 we had POWER6, which led to POWER6+. In 2010 we had POWER7 and the most current, POWER7+. In 2014 we have POWER8, and there are already charts that show POWER9 is being planned for the future. IBM has consistently delivered on its roadmaps.
I recently attended an education session for IBMers and business partners that covered information around POWER8 and the new IBM hardware announcements that are being made today. I am going to hit some of the highlights, but additional information will be included in future posts.
The POWER8 Chip
The POWER8 chip is another step up from what has come before. We have gone from four threads to eight threads per core. With simultaneous multithreading (SMT) enabled you can have up to eight threads running on a core, which means you can get more work done per CPU cycle.
The charts that I saw showed a linear increase in the number of transactions that could be completed when you compared SMT1 to SMT2 to SMT4 to SMT8. As you made each transition you could see the number of transactions increase. Obviously some workloads won’t benefit from SMT, but those will be the exception rather than the rule.
I also saw charts that compared I/O bandwidth and memory bandwidth on the new systems compared to older models, and the numbers were impressive. It was a significant increase that I will be discussing further in future articles.
While POWER7 technology had up to eight cores per socket, the POWER8 chip has up to 12 cores per socket. New memory controllers and memory cache on the system improve memory latency and performance.
The way the cores communicate with one another across the SMP interconnect has also improved so it takes less “hops” to go from one core to another in the system. The chip also boasts a direct PCIe Gen3 I/O interface for incredible bandwidth.
There is 512 K L2 cache per core, 96 MB shared L3 cache and up to 128 MB L4 off-chip cache.
Understanding the Models
How comfortable are you with the model numbers of the Power servers. If someone says 720, 770 or 795, do you have a pretty good idea what server they are talking about? With today’s announcement, how many of you were expecting to see 820 and 870 server models? This is not going to be the case. The servers are now named with four- or five-digit combinations of letters and numbers. For this first announcement, the servers all start with the letter “S,” which signifies that they are scale-out servers. As time goes on I would expect to see models that start with the letter “E” for enterprise systems. The second digit indicates that it is running POWER8. The third digit indicates the number of sockets in the server, and the last digit indicates how much rack space it takes up, for now either 4U or 2U.
For example, the S822 is a scale-out server, running POWER8 technology, with two sockets, fitting in 2U of rack space. The S824 is a scale-out POWER8 two socket 4U server. If you see an L in the fifth digit, like the S822L, then that is a Linux-only system, much like today’s 7R1, 7R2 or 7R4 servers.
We need to pay attention to the lettering. The L designates it will only run on Linux. The 2U non-L models can run AIX and Linux. The 4U non-L models can run AIX, IBM i and Linux. At the time of this announcement, you cannot have an I/O drawer with PCIe slots on any of these machines, although a statement of direction indicates that this capability will be available in the future.
Here are the specs for the new servers:
- The two socket 2U servers (S822) can have different configurations depending on whether you populate both sockets. If you have one socket populated, you can have six or 10 cores, with up to 512 GB of memory. There are six PCIe Gen3 low-profile hotplug adapters in this configuration. If you have both sockets populated, you can have 12 or 20 cores, with up to 1 TB of memory. Nine PCIe Gen3 low-profile hotplug adapters are included in this configuration. You can run PowerVM with AIX or Linux, but not IBM i on this server.
- The S822L can have 20 or 24 cores, with up to 1 TB of memory and nine PCIe Gen3 low-profile hotplug adapters. You can run PowerVM or PowerKVM and you can only run Linux on this machine.
- The S814 is a one-socket 4U system that can come in a 4U or tower form factor. It has six or eight cores and 512 GB of memory. You can have seven PCIe Gen3 full-high hotplug adapters, and you can run PowerVM with AIX, IBM i or Linux.
- The S824 is a two-socket 4U server. If you populate one socket you can have the same specs as the S814, but if you populate both sockets you can get 12, 16 or 24 cores with up to 1 TB of memory. You will have 11 PCIe Gen3 full-high hotplug adapters and can run PowerVM with AIX, IBM i or Linux.
Performance
The rPerf and CPW numbers that I saw showed improvements, and I will write more about this in the future as well. IBM asked us not to share the numbers until they are audited and vetted, but I will be surprised if the improvements, especially when compared with competitor’s machines, are not as dramatic as we saw during the training sessions. It was also amazing how these new systems perform when comparing an S824 vs a POWER5+ 595 or a POWER4 690.
Another part of the story is how this improved performance translates into needing fewer cores to do the work that you need your server to do. That means you will need to spend less to buy hardware, and you will receive better performance per dollar spent.
We will be able to perform Live Partition Mobility operations between POWER6, POWER7 and POWER8 machines, assuming we’re using the correct processor mode. We can run the LPARs in POWER6 mode, POWER7 mode or POWER8 mode. This will also make it possible to run OS versions that are not POWER8 aware assuming you are using VIOS for your I/O.
Miscellaneous Information
You can run AIX in POWER8 mode with full I/O support once you get to:
- AIX 6 TL7 SP10
- AIX6 TL8 SP5
- AIX 6 TL9 SP3
- AIX 7 TL1 SP10
- AIX 7 TL2 SP5
- AIX 7 TL3 SP3
POWER8 support for IBM i will be available in IBM i 7.1 TR8 and IBM i 7.2, as well. We will need to be running VIOS 2.2.3.3 for POWER8 support.
There is also a new HMC model 7042-CR8 that will be available later in the year.
I should be getting my hands on some of these models shortly and will be able to share more information once I do.
These are some of the highlights that I found interesting. What are you looking forward to the most with these new systems?