HPC Blog So, it looks like IBM has moved its chipworks over to Globalfoundries, as described by Reg hack Tim Worstall here. But what are the implications for IBM's systems business?
I agree that the brutal and implacable economics of the chip business forced IBM’s hand here. Manufacturing microprocessors is a hugely capital intensive undertaking, with massive fixed costs for the buildings and equipment necessary to churn out shiny new CPUs. It also costs a lot of dough to acquire and hold onto the highly specialized people who can design and build cutting edge silicon.
IBM Microelectronics just didn't, or doesn't, have the volume to justify the kind of ongoing investment required to keep running on the process treadmill. Last year, the division booked revenues of around $1.4bn, but netted a loss of $700m – which is a bitter pill when you consider the future investments that will be needed to remain competitive in the microprocessor race.
The $1.5bn IBM will be paying Globalfoundries (in $500m annual installments) to provide Big Blue with processors and other stuff is a bargain compared to the cost of moving from their current 22nm process to 14nm and beyond.
Globalfoundries will book around $10bn over the next year in capital expenditures – which shows you just how expensive it is to compete in the chip building side of the industry. For its part, IBM will be still be investing $3bn in advanced chip research over the next five years, but this will probably end up being a cooperative effort with Globalfoundries.
One of the constant refrains I noticed in the comments thread on Tim Worstall’s story was that this move means IBM is leaving the systems business. Over the next few years, it is supposedly going to drop both its mainframe and Power systems business.
IBM... dumping Systems? Really?
I don’t see it.
IBM’s hardware sales have decreased, that’s for sure. And they’re probably going to drop a bit more, at least in the near term. In the medium term, the mainframe business is going to continue on at pretty much the same pace. They’ll sell plenty of MIPs, but margins will drop over time – which is natural, given that mainframe alternatives are constantly improving. But they won't be capturing a lot of new customers and workloads with the mainframe.
The sale of their x86 business to Lenovo leaves IBM's stable with only one horse they can ride in the system sales derby. That horse is Power8. (I guess the saddle and tack is the system hardware.)
But just because IBM has only one horse left to ride, it doesn’t necessarily make that horse a good mount.
Power vs x86
IBM has tried to use Power to displace x86 systems before. With previous Power boxes, IBM had performance and efficiency advantages vs comparable x86 based systems. But that goodness came at a cost.
First, you’d have to run its commercial UNIX variant (AIX) and use AIX ports of popular apps. If your shop doesn’t already run a fair amount of commercial Unix, you probably wouldn’t be too keen on adding a new o/s to your existing mix.
You’d also pay more in terms of both acquisition and ongoing service costs. IBM would argue that the higher performance, RAS and utilisation rate of a Power server tipped the TCO (total cost of ownership) math in its favour, however it’s a complex argument to prove – and the results will vary from data centre to data centre.
So even though IBM had hoped Power would displace a chunk of x86 servers, it didn’t happen very often. Customers mostly installed Power boxes to replace incumbent Solaris and HP-UX systems.
The endian of the line and Chip Chalk Talk
However, Power8 is a different animal (no, not a horse, I’m giving up on that metaphor.) The biggest deal is that it’s little endian, meaning endian compatible with x86. When you had an endian mismatch, like with previous Power/RISC processors vs x86, application providers would have to go through what can be a long and complex porting process. Because of this, and a smaller potential market, ISVs weren’t eager to port apps to the big endian Power processors.
Power8 is what IBM calls “bi-endian”, meaning it can access data stored in either big or little endian styles. What this really means is that little endian-based software can run on Power8 boxes with only a recompile to translate the instruction set differences between Power and x86. Software using interpreted languages like Java should run right out of the box on Power8.
The new Power8 boxes also run Linux operating systems (right now Ubuntu – with Debian, openSUSE and maybe RHEL on the way) on bare metal. This opens up a whole new universe of closed and open source Linux applications to potential Power8 users.
IBM has also addressed the cost issue with the Linux versions of their Power8 product line. IBM figures the best place to build their beachhead vs. x86 in a few areas – namely infrastructure and analytics workloads. With that in mind, they’ve priced the boxes (including the O/S and virtualisation) to sell on a par with a comparable x86 Linux/VMware solution. Service and support is also priced at x86-like levels.
Power8 has some technical advantages that can yield significantly better performance vs. x86. For example, each of Power8’s processing cores can support 8 threads vs 2 threads for x86 CPUs. Power8 has larger processor caches – 100MB on chip plus 128 off chip vs the 37.5MB on Intel’s latest chips. The biggest advantage is in memory bandwidth, with Power8 able to pump 230GB/s through the chip, while Intel’s processors are limited to 85GB/s.
However, Intel system can support 1.5TB of main memory on a 2-way box today, while IBM’s Power8 boxes can only support 1TB.
Some might argue that all this hardware talk is irrelevant. Who cares about systems these days? Hardware is a commodity and the profit margins are so low that no one cares anymore. Those people are wrong.
It’s important for IBM to make this Power gambit work.
MBA dimbulb-ery and marginal thinking
The example I’ve been using with clients is to imagine that a newly minted MBA is looking over the products sold by an auto supply chain. In looking at the numbers, he sees that lug nuts (the nuts that hold wheels onto the car) are the highest margin product in their tire-related product line. His advice to the chain is to stop selling lower margin tires, wheels, and those little pine tree air fresheners in order to devote more shelf space towards lug nuts. It’s pretty obvious that this is a stupid strategy, right?
While IBM makes the highest margins on software, from a financial standpoint, that software is usually part of an entire IBM solution. If IBM were to abandon hardware, that will make it more difficult (and costly) for them to provide those complete solutions.
You can argue that customers are rolling their own solution suites, which is true, but a company like IBM can still sell piece parts to play into whatever the customer is trying to put together.
There’s also something to be said for getting more "wallet share" with customers. If you’re successfully selling software and services, adding hardware gives you a larger share of customer spend, and as long as the hardware portion is sold at a positive margin, then it’s always better to have a bigger sale than a smaller one.
In finance, this is the “high Internal Rate of Return vs Net Present Value” problem. Money people present it as an example where Project A has a relatively low cost, but a very high return over several years. On the other hand, Project B has a much higher cost and a much lower rate of return. Which project do you take on? You always pick the project that provides the most cash, after discounting for the time value of money.
The point of all this blathering isn’t to argue that IBM is going to bludgeon the life out of x86 with their Power8 boxes. What I’m saying is that IBM will stay in the hardware business for at least the foreseeable future. The return it gets from hardware and the accompanying ability to sell complete solutions is higher than the cost savings it would realise by abandoning it.
However, IBM needs to put all of its considerable resources behind its new hardware initiatives. It has to execute on its plan to build Power based systems into a worthy x86 replacement, meaning it has to hold the line on prices while continuing to improve the underlying technology.
It’s not going to be an easy road, as I’m sure the comments on this story will point out. ®