Why Not Overclock Power Chips For IBM i?
August 28, 2017 Timothy Prickett Morgan
Remember way back in 2000, when Intel, the world’s largest chip manufacturer and now the world’s dominant supplier of processors in the datacenter, said that it would be able to deliver processors that run at 10 GHz by 2011. Well, that was six years ago, and that sure as hell did not happen then and it is not going to happen now. But if anyone can crank the clocks high on a processor, it is IBM with its Power and System z engines, and we think it can push them a little higher and give certain customers who have a need for speed something to get excited about.
The pace of process shrinks on chip manufacturing techniques is slowing, and the jumps between nodes are getting stretched out. A lot of the performance improvements that companies are getting with modern CPUs is because they are wider and more clever in that they squeeze more work out of each clock cycle.
Still, it is a fact of physics that clock speeds on processors either flattens or comes down as the number of cores goes up. This has certainly happened in the Xeon processor line from Intel, and it is what happens when the needs of search engine makers – the few that there are – outweigh the needs of enterprise companies doing traditional OLTP processing on relational databases. But Intel has been schooled by database makers like Microsoft and Oracle, who often get custom processors made by Intel for their iron so they have fewer cores that clock higher and therefore can get more single-threaded or lightly multithreaded work done in a given unit of time.
The current Power8 chip and the impending and future Power9 chip have among the highest clock speeds in the industry. Both of these chips had a target design frequency of around 4 GHz, but they are capable of going faster and sometimes do. And that is with air cooling, not water cooling. We think that a Power-based database and batch processing engine with blazing speed for these workloads, and with as few cores as possible, is something that customers want to buy. Moreover, with the Power9 launched pushed out to early next year and with IBM having nothing new to sell as Intel and AMD have new X86 server processors and Qualcomm and Cavium readying credible ARM server processors, Big Blue needs something to get the IBM i base excited and buying something. Sorting through the bins and finding Power8 chips that can be overclocked to boost single-thread performance would be fun and would also test the idea that maybe wider machines are not the answer for everyone, or at least not for every workload in a Power Systems machine doing OLTP as well as other work.
Back in 2015 when I was thinking about this and the Power8 chips were about a year and a half old, I pointed out that the average IBM i customer has somewhere between 1,000 CPWs and 1,500 CPWs on processors that delivered something around 9,000 of CPWs of oomph. That’s on a single core, and Power8 chips can have up to six or 12 cores on a die and have 12 cores on a processor package. All that IBM does is see how many cores are good and what speed they can run at, and given how many low-core count Power7, Power7+, and Power8 chips it actually sells into the IBM i base, all those sexy feeds and speeds don’t mean jack to these customers. And given how expensive it is to drop IBM i onto a core – thousands to tens of thousands of dollars per core, depending on the software tier – there is pressure to buy as few cores as possible.
I have tried to convince IBM to lower that price and get customers to move more workloads to IBM i and therefore increase the compute demand, but IBM wants the IBM base to behave like the mainframe base and adopt Linux on all of that excess capacity in the server, as has happened to its great joy on the System z mainframe. Linux has saved the mainframe, no question about it, but IBM i shops are not going to be Linux shops. They are allergic to Linux, for the most part, and they have Windows Server as their adjunct platform and they can’t run Windows Server on Power. (But soon, they will be able to run it on ARM servers and they can run it on Intel “Skylake” Xeon SP and AMD “Naples” Epyc processors.) By the way, that does not mean that IBM i shops are allergic to open source software that was largely developed on Linux and that has been ported to run, in one fashion or another, on the IBM i platform. All kinds of open source software, from PHP to Ruby to Node.js, are being embraced by developers.
The point of overclocking is not capacity, as expressed by CPWs on the IBM i platform, but rather reducing latency and cranking through relatively serial work as quickly as possible. Like high frequency traders do. It also results in faster compile times for applications, which is important for programmers but not so much for end users. And, importantly, it is something IBM can get out of the factories and put on the truck to sell between now and next March or April, when the Power9 machines are expected to debut running AIX and IBM i.
IBM does not provide the wattages of its Power8 chips when they are at peak or running workloads, so it is hard to reckon just how far Big Blue can push up the clock speeds. In the four-core Power S822 machine that was launched in October 2016, the clock speeds are 3.42 GHz, 3.89 GHz, and 4.15 GHz. The Power E870C has 4.02 GHz processors and can have as many as 64 of them in a single system image, and the Power E880C – the most capacious machine in the product line – has up to 128 cores that run at 4.35 GHz, up to 160 cores that run at 4.19 GHz, or up to 192 cores that run at 4.02 GHz. These are not the fastest processors IBM has delivered, you will recall. Some Power6 chips used in midrange machines clocked in at 4.7 GHz (the Power 520) and 5 GHz (the Power 550) from way back in April 2009; the Power 570 and Power 595 hit 5 GHz, too, back in the summer of 2008 as well. The Power5 chips were in the range of 2 GHz, and the Power4 chips came in at around 1 GHz with a smidgen extra. Just for comparison.
Let’s have some fun here. I think that IBM can jack up the clock speed on a Power8 chip to at least 5 GHz without too much sweat, and with water cooling on the processor and memory components, which is done on game PCs and supercomputing nodes these days, it should be able to crank a Power8 chip to somewhere just south of 6 GHz. What would such a machine get anyone?
In a balanced system with a very good cache hierarchy and enough main memory and flash memory, a 6 GHz Power8 core would basically perform like two 3 GHz cores in terms of raw capacity, and if IBM kept the per core pricing constant – aha! – then the resulting system would be considerably less expensive for a given amount of capacity. Yes, this beast would throw off a lot of heat, but who cares? The other upshot is that the latency on OLTP transactions could be a lot lower, and other kinds of work would be twice as fast, too.
I am well aware that the DB2 for i database has been tuned to run across a lot of threads, like other databases, but the mere fact that so few customers need this and run on machine with one or two cores tells me that we need to do something else that these customers might value, and then to charge for that and build a business that makes sense for the actual IBM i customers. As a proud owner of a Dodge Challenger and as a man who appreciates the joy of acceleration and big iron, I can tell you that low latency is fun; while I love my old Dodge Dakota Sport truck, there is no point in stomping on the gas. But with transactions, particularly for web-based applications where customers are fickle and particular for batch jobs running in the background, shortening the time to completion is a competitive edge that IBM might be able to provide.
Everybody wins. I love a scenario like that. Customers get speed demons, IBM gets money, and maybe it gets a way to sell Power9 machines into the IBM i base and not just chase Google and Rackspace Hosting and other big data analytics workloads with a super-wide, not-very-fast Power9 chip with 24 cores running at maybe 3.5 GHz to 4 GHz. This will be some pretty iron, to be sure. But it doesn’t do a damned thing for IBM i shops who could have done just fine with a single Power7 chip with a few cores.
So what do we charge for this Power8 speed demon? And how would we configure it up with a Power9?
Let’s start with the Power S812 “Mini” system that was announced earlier this year. Rip out the disk drives, because this baby needs flash and DRAM only. Take out the 1 Gb/sec network adapters because we need at least 10 Gb/sec Ethernet these days. The storage backplane is way too expensive, too. IBM wants $16,111 for a machine with a single core running at 3.03 GHz with 64 GB of memory and eight 283 GHz disks. Forget that. I want a machine with one core running at 6 GHz with water blocks on the CPU and the memory, and I want at least 256 GB of memory and two flash drives weighing in at 1 TB each. And it cannot cost more than $15,000. Then, I want a double up machine – two 6 GHz cores, 512 GB of main memory, and 4 TB of flash – that costs $30,000. This will really test the idea, and I want to see how much faster work can get done.
While I am thinking about it, the software stack costs $19,850 on the Power S812 Mini. Cut it in half. Call it $10,000 per core with five users paid and, here’s the rub, any users after that are free. Remember: we are trying to get people excited here about server processors that are now four and a half years old.
Then, we get crazy with Power9 next year. We have the Power9 cores running at the same 6 GHz, which is about as fast as I think they can go, but now we have faster DDR4 memory (let’s push it up to 3.2 GHz) and PCI-Express 4.0 peripherals with the latency-reducing NVM-Express on the flash. And let’s get even crazier and put some Intel Optane 3D XPoint memory-extending persistent memory in among the DDR4 memory to make a much bigger memory space. Call it 256 GB of DRAM memory and 256 GB of Optane memory, and then double up the capacity of the flash once again. And do it all at the same price points, or maybe with something like a 25 percent premium.
Yeah, I know that sounds crazy. So what? I would rather IBM be crazy and make money than be smart and just sit there, waiting, and not doing something that excites the IBM i base.