The IBM Sales Pitch For The Power E1080
November 29, 2021 Timothy Prickett Morgan
It has been a long, long time since a lot of server customers needed to upgrade their machinery every time a new processor and a new system using it comes to market. Back in the early days of the AS/400 platform, customers on the cutting edge of modernizing their back office, manufacturing, and distribution operations often upgraded their machines once a year to add capacity as use cases for these mission critical systems expanded faster than Dennard scaling and Moore’s Law increases in CPU performance.
These days, machines are installed and not upgraded or swapped out until one, two, or three generations of replacements have been in the field. And so it is that with the new “Denali” Power E1080 server from IBM, the first of Big Blue’s machines based on the “Cirrus” Power10 processor, that no one expects a large number of Power8 shops to buy these big boxes. But there are a much larger pool of big iron Power7, Power7+, and Power8 machines that have been in the field for many years that are due for replacement – and with machines with enough excess capacity to absorb bigger workloads and newer workloads for years to come.
To recap our coverage on the Power E1080 system thus far, the Denali machine, which scales up to 16 Power10 processors with a maximum of 240 cores and up to 64 TB of main memory in a single system image with a stunning 5.27 million units of aggregate performance on IBM’s Commercial Performance Workload (CPW) online transaction processing benchmark test, was announced on September 8. After the announcement, we drilled down into the customer based that the Denali server was aimed at, and last week we finally got into the price/performance of the Denali machines – just the underlying processors and memory, not the full system cost – compared to prior high-end servers.
This week, we are going to talk about the sales pitch IBM has developed to help its sales reps and its partners peddle the Power E1080 server to customers with existing big Power Systems machinery running the IBM i and AIX platforms.
This chart is extremely useful in lining up the Power8, Power9, and Power10 big iron with their distribution of core counts within the Power processors used in the systems and the node counts across the NUMA clustering that makes it all look like one big memory space and one big virtual CPU to IBM i or AIX. Take a look:
Depending on the clock speeds used in the comparison, the Power10 cores are anywhere from 20 percent to 30 percent faster than the Power9 cores used in the prior Power E980 server and anywhere from 55 percent to 80 percent faster than the cores used in the Power E880C. That increase in per-core performance means that to support a given workload, customers with older machines will be able to do so on a system with far fewer cores, and with so much system, middleware, database, application software used on IBM i and AIX platforms being priced on a per-core basis, this can mean – with the right negotiating tactics with software suppliers – a substantial reduction in software license fees. This systems software stack along tends to be many times more expensive than the hardware, so the software consolidation can often cover the hardware cost over the first one or two years while at the same time improving response times and leaving headroom for growth.
It is interesting to see, for instance, that a two-node Power E1080 using the 12-core Power10 processor has more performance than any Power8-based Power E880C server, and all but the most capacious variants of the Power E980 server. IBM did not sell single-node Power E880C or Power E980 servers, but it is selling single-node Power E1080s, and that means customers with Power 770, Power 770+, Power 780, Power 780+, and Power 795 servers – there are still some of these in the field – will be able to plunk down a single-node Power E1080 and have lots of excess oomph while at the same time being positioned for growth in the years ahead.
A power E1080 with 80 cores running at 3.65 GHz is rated at 1.83 million CPWs, which is a two-node, eight-socket system based on the 10-core Power10 processor. A full-on Power 795 machine with 32 sockets and 256 cores running at 4.0 GHz, by contrast is rated at 1.6 million CPWs, and each unit of performance (including a reasonable amount of memory) cost 5.2X more back in 2010 than the Power E1080 does today. This is a testament that microarchitecture improvements in CPUs, system architecture, system software changes, and other forces for improving performance and bang for the buck are still at work, even if it is getting tougher and tougher.
But as the chart below shows, there are big power, cooling, and space savings that come from upgrading to the Power E1080 as well:
Reducing power and cooling costs is interesting, but these are really but the decorations on top of the icing that is on top of the cake. The cake is having a lot more performance per core, and the icing is needing fewer software licenses to do a given amount of work. What is clear that every piece of these total argument helps to make the case for a move to the Power10 iron, and there is nothing wrong – and everything good – about being more “green” in the datacenter.
Here is what the upgrade paths from the Power8 Power E880C machines to the Power10 Power E1080 machines look like, just as an example of the things you need to consider:
The performance figures shown with each tier of system are expressed using the AIX Relative Performance (rPerf) online transaction processing test, not CPWs, but they are roughly analogous. Basically, IBM wants customers to do a 2:1 node consolidation to deliver the same amount of performance and still get a nominal increase in performance. This Power E1080 upgrade pitch is more about saving money on systems software, maintenance, power, cooling, and space. And that stands to reason here at the end of 2021, when these machines are still ramping. Early next year, customers can add another node and significantly boost capacity.
Customers who are shopping for Power E1080 machines have to watch the clock speeds and single threaded performance of a core. You can get equivalent throughput but have lower single threaded performance per core, and that might mean that some batch jobs take longer. So choose carefully.
As we have pointed out in our coverage of the Power10 processor, Samsung is able to deliver volumes across the SKU stack of Power10 chips better than Globalfoundries was able to do with the Power8 and Power9 chips, and therefore the availability of chips with high core counts is better than it has been in the past – and the prices of the performance in these chips is a little flatter than it was with Power8 and Power9. In the past, if you wanted a high core count chip, it came at a very high premium for the processor card and sometimes the core activations, too.
Speaking very generally, the 10-core Power10 processor running at 3.75 GHz in a single node configuration of the Power E1080 has the lowest entry price to get into the Denali server. IBM’s sales people are leading with the 12-core Power10 processor running at 4.15 GHz because it has the best performance per core and the middle of the line core count. Customers that need a lot more throughput are going to have to sacrifice some single-threaded performance, stepping over to the 15-core Power10 chip running at 4 GHz.
I think in the long run, Samsung will be able to get even better yields than it is currently getting with the Power10 chips and there is a good chance that it will be able to have a 16-core Power10 chip running at 4.15 GHz or maybe even as high as 4.25 GHz, giving Power E1080 customers a healthy slice of cake with lots of icing. But that is just a hunch, nothing more.