Bang For The Buck For Big Power Iron Over Time
November 15, 2021 Timothy Prickett Morgan
With the shipments of the “Denali” Power E1080 server ramping at IBM, we have also been ramping up our analysis of the machine to help the relative few of you IBM i shops – on the order of hundreds – who will need to upgrade your Power 795, Power 770, Power 780, Power E870, Power E880, and even possibly your Power E980 machines to the new top-end Power Systems NUMA server.
But make no mistake. What IBM is doing with the Power E1080 in terms of performance and price/performance bears relevance, indirectly and somewhat more directly, on each and every Power Systems shop in the world, whether they run IBM i, AIX, or Linux – or a mix thereof – on their systems and no matter what size machine they might have or what mix of vintage of software and hardware they run their applications on.
As we explained in The Big Iron Customers That The Power E1080 Is Aimed At, in the wake of the announcement of the Power E1080 machine back on September 8, while there are only somewhere on the order of 13,000 to 14,000 Power Systems shops that have machines with four sockets or more – what IBM calls enterprise-class machinery when talking about the servers and scale-up when talking about the processors – and a shared memory architecture across those processors. But the customers using these machines often have two or three big machines for high availability and disaster recovery, and they account for a large share of the Power Systems revenue stream and an even larger portion of the profit pool left over after costs. So keeping these customers happy, with lots of processing and memory and I/O bandwidth to grow into, is key to Big Blue continuing to offer Power-based servers.
In other words: no big iron, no little iron. That is the somewhat direct effect of the Power E1080 launch on the Power Systems base and IBM i within it, who comprise about 80 percent of the 150,000 shops worldwide who have Power-based servers running their mission critical applications.
Last week, we drilled down into the architecture of the Denali system, and as promised this week we are looking at the long curve of performance and bang for the buck that big NUMA machines based on Power processors have had over time and what the pricing for the Power E1080 might imply for the pricing for the future Power S1014 single-socket server, two-socket Power S1022 and Power S1024 servers (and their L, LC, and H derivatives for Linux and SAP HANA), and four-socket Power E1050 server might look like. This is the indirect effect.
One of the interesting bits of information I have derived from looking at Power E1080 pricing – and something that Steve Sibley, vice president and global offering management for Cognitive Systems at IBM, told me was going to happen with the Power10 – is that IBM is no longer charging a huge premium for the Power chips where the core counts are high relative to the price per unit of capacity it is charging for Power chips in the same generation where there are fewer cores. And that tighter pricing is thus reflected in the overall more consistent cost of what IBM has always called the central electronics complex, or CEC. What is meant by CEC is the processor cards, processor books, or system nodes and their motherboards (depending on how the big NUMA machines are packaged) and their system memory.
There are a lot of different ways to configure these big NUMA machines, and pricing and configuration information is a bit thin, too. So we have gone back through all of the big NUMA announcements since the Power 595 based on Power6 chips was launched in 2008 and configured up two representative CECs for these machines for each generation. One is a top-end configuration and the other backs off the core counts (but usually at a higher clock speed). In the early years on the Power 595 and Power 795, when memory was expensive but nothing like it is relative to maximum system capacity, I staggered the capacity as well as the cores (mainly because that is how I got configuration and pricing data back in 2008 and 2010). I used the middle capacity memory cards, which are not the most dense options and which are not the most expensive ones, either.
Here is what the pricing on the big NUMA Power Systems machines looks like for the configurations show, and the CPW ratings for the entire capacity inherent in the machines running the IBM i operating system. CPW is short for Commercial Performance Workload, which is a variant of the TPC-C online transaction processing benchmark that IBM uses to gauge relative performance of machines running IBM i. Take a look:
On the Power E1080, the Power10 processor feature with 40 cores running at 3.65 GHz (four 10-core processors in the node) costs $135,000. The 48-core feature with the cores running at 3.6 GHz costs $220,000, and the full-on 60-core feature running at 3.55 GHz costs $320,000. That may sound like a big price difference, but the cost of activating these cores is $5,000 a pop, and this adds up and counterbalances the base feature cost differences to a large extent. (But not completely.) The 512 GB DDR4 feature card costs $24,490 and it costs $82 per GB to activate the capacity. The 1 TB memory feature card costs a little more than twice the 512 GB feature card (at $57,951) and the 256 GB memory feature card costs a little less than half at $10,204. At $8,929, the 128 GB memory card costs almost as much as the $10,204 for the 256 GB memory feature card, so don’t buy it. The memory activation costs are the same across the cards, and account for the vast majority of the cost of the memory cards.
The base system nodes in these machines cost somewhere between 2 percent to 4 percent of the total system cost. With the Power 595 from 2008, memory on the configurations shown accounted for 17.9 percent of the total CEC cost for the 256 GB by 64 core configuration and memory accounted for 23.9 percent of the CEC cost on the 512 GB by 64 core configuration. The memory was over $2,400 per GB back in 2008 for the densest memory. And as expensive as main memory is these days, it was just under $150 per GB on the Power8 machines shown, just over $120 per GB for the Power9 machines, and just under $140 per GB for the Power10 machines in the table above.
It is hard to grasp those performance and price/performance changes between 2008 and 2021, so here is a handy dandy chart to help you see it more clearly if you think visually as I do:
The fat blue bars are the performance in CPWs of the configurations in the table. Look at how that performance has grown. That is how and why IBM engineers the Power processors as it does: To drive that high-end big NUMA server performance. Everything else is trickle down to smaller machines with this architecture, and that is because IBM knows where its butter gets bread. (In the money sense of that word.)
You can also see that incremental cost of the highest end performance with each generation of machines, but you can also see that with the Power E1080, the price/performance difference is negligible on the big and biggest configurations. And look at how that cost per CPW has come down over the past 13 years and five generations of machines. It was around $7 to $9 per CPW back in 2008 with the Power 595, and now it is around 50 cents per CPW, give or take. The prices on the machines have not changed much except for the heavy premium that IBM charged for the Power 795 (no doubt because of the more expensive book packaging technology and the far scalability to 32 sockets at the time). Most of that price performance improvement is coming through performance, not through lower costs.
Which is how IBM plays this game.