GuruFocus Premium Membership

Serving Intelligent Investors since 2004. Only 96 cents a day.

Free Trial

Free 7-day Trial
All Articles and Columns »

Is IBM a Threat to Intel?

June 21, 2014 | About:
FinanceGuru

FinanceGuru

2 followers

When it comes to server processors, Intel (INTC) is by far the most financially successful. While Intel is well known for having the highest-performing microprocessors in the industry, this isn't strictly true. Indeed, if there is one company out there that builds bigger, hotter, and more powerful processors than Intel does for ultra-high end Unix servers, it's IBM (IBM) with its POWER line of microprocessors. With IBM's recent launch of its POWER 8 processors, should Intel be worried?

The power of POWER 8
IBM's POWER 8 is truly a technical tour de force:

Source: IBM.

The POWER 8 CPU core is extremely wide, clocks quite high, and offers what seems to be unmatched single-threaded performance. In fact, if we look at performance results of a system packing four of the six core variants of the POWER 8 clocked at 3.52 GHz, and compare it ith a system packing four of Intel's most recently launched Xeon E7-4890 v2 (codenamed Ivy Bridge-EX), IBM's solution compares quite favorably on a per-core basis.

The four-socket Xeon E7-4890 (which packs 60 Ivy Bridge CPU cores clocked at 2.8 GHz) scores 138,900 SAPS (higher is better) while the four-socket POWER 8 (which packs 24 CPU cores clocked at 3.52 GHz) scores 115,870 SAPS. The Intel system offers greater performance, but on a per-core basis the IBM is much more powerful. Also, from a more geeky technical perspective, the IBM CPU is much wider, the chip has significantly more cache, and it offers much higher memory bandwidth than the Intel chip.

Great, IBM's POWER 8 is a beast. So what?
While the performance of the IBM POWER 8 is not at all in question (this is a magnificentpiece of engineering), this has never actually been a problem for IBM's POWER line of processors. The POWER 7 was a much beefier and powerful processor relative to its Intel Xeon competition, yet Intel has continued to gain share in the overall server processor market. So what are the important questions investors need to ask about POWER 8?

Intel's success has never just been due to performance leadership; a 6-core POWER 8 about matches a 15-core Intel Xeon, and the 12-core variant of the POWER 8 would utterly embarrass it in performance. The problem for IBM is that the equation for data-center dominance isn't simply about performance, but is instead about performance per watt per dollar. While it's too soon to get a read on the performance per watt of the IBM POWER 8 (although it's likely that this kind of performance doesn't come cheaply), the real problem that is likely to plague it is the cost.

IBM doesn't have Intel's scale
The POWER 8 is built on IBM's custom 22-nanometer SOI process (finely tuned for high-performance CPUs) using 15 metal layers and weighing in at a hefty 650-square-millimeter die size. Further, note that the volumes of this processor are likely extremely low and, in case you haven't been following IBM's financial statements, the sales of systems using these processors continue to plummet. On top of that, given that this is quite a large, exotic design, yields may still be difficult (but for what machines based on this chip sell for, IBM can afford it).

IBM's Systems & Technology group continues to bleed out. Source: IBM.

On the other hand, Intel's Xeon E7 (Ivy Bridge-EX) is a 4.31 billion-transistor machine weighing in at 541 square millimeters built on a quite mature 22-nanometer FinFET process in factories that have to date pumped out hundreds of millions of 22-nanometer FinFET processors. While the volumes on this part are low, and while yields on this chip are almost assuredly trickier than a 1.4 billion-transistor Haswell notebook CPU, they are probably better than the IBM POWER 8's at this stage of the game.

In short, even if the performance per watt is better than or equal to Intel's, Intel is likely to be able to offer its product much more cheaply as a result of the massive design leverage, as well as the manufacturing leverage it gets from its PC chip and other segments of its server business.

Conclusion
A big round of applause for IBM's technical teams in executing one of the most interesting and exotic processors on the planet today. While IBM appears to hope for a better future for POWER by opening up the architecture and designs for third-party licensing via its OpenPOWER initiative, the markets that IBM's niche high-end servers that it builds POWER for are still very difficult and not getting any easier.

The key things to watch for:

  • Licensing uptake with OpenPOWER.
  • IBM POWER server sales as POWER 8 systems ramp.

If neither of these points to renewed signs of life for POWER, then the fate of these technically beautiful POWER processors may be sealed and Intel will have nothing to worry about. If, on the other hand, OpenPOWER gets POWER 8 into more markets, then Intel may indeed have something to worry about.


Rating: 0.0/5 (0 votes)

Comments

EnterpriseArch
EnterpriseArch - 2 weeks ago

Excellent article capturing the basic essence of Power performance vs x86. However, I suggest you where you took a wrong turn was with the TDP or the Perf / Watt topic. First, here are is what truly makes Power different. You and your audience would benefit from learning more on each of these features and how to deploy in their environments.

  1. Power servers are virtualized from the start - assuming they are using PowerVM or PowerKVM. PowerVM is the norm so let's stick with that. Use cpu, ram & I/O as dedicated or virtual or both with very few limitations!
  2. Power virtualization is secure, scalable, dynamic and flexible. Dynamically change cpu, ram & I/O up or down to each VM. Move cpu, ram & I/O between workloads dynamically as needed. Make most changes dynamically within AIX and IBM i. PowerVM is secure with no known vulnerabilities unlike x86 hypervisors. Move VM's across server generations such as Power6 to Power8 and back. Great for upgrades, emergencies and planned maintenance.
  3. Secure thru the stack - at the server, hypervisor, OS and even into the application using features like vTPM, PowerSC, Real Time Compliance, AIX Trusted Execution, no-cost Encrpted Filesystems using hardware encryption accelerators to offload off of processors.
  4. Reduce software licensing - unlike x86, with each new Power generation products like Oracle or IBM Software licensing costs should go down as long as the workloads do not go up. Unlike x86 which does not see a per core increase with new processor generations, each Power generation see's a increase with Power8 seeing up to 2X over Power7. Don't argue over the number but focus on the point that it goes up by some %. If it is 50% great! 75% awesome!! 2X Woohoo!!!
  5. The Big x86 Lie - Power is sooo expensive! Power6 and older was because it was competing against Big Iron. Power7 pricing came down significantly due to the competitive landscape changing - it is now x86, Linux and VMware (I'm ok if you want to group Hyper-V in there but I don't see them in too many data centers yet not to say they are evaluating it). Power8 has performance up even more with pricing down even more. Great time for existing Power customers! Better opportunity for those new to Power as they do not have the expensive buy-in. Power can have price parity to be price comparable. If you want Lexus features only available in Enterprise class servers like the Power7 770, 780, 795 those cost more just like a Lexus or BMW cost more than a Chevy, Ford or Toyota. Look at some of the ultra high-end x86 servers and they aren't cheap! Not sure I would run my business on a single high end x86 server, I would and I do know of many, many customers who run their entire businesses on a single enterprise class Power server...for that matter, thousands and thousands of customers run their entire businesses on entry level servers with virtually no issues as well - ask any iSeries customer when the last time they had a issue compared to their x86 environment.
  6. Power utilization can go as high as 90%. You can drive server utilization with a quality of service very high on Power servers unlike x86 - this is just for peaks but sustained workloads. ram, cpu and I/O. The components are designed for it as well as reliability and availability features built-in to the servers are designed to prevent server failures by detecting failures before they happen, re-route into spare and recovery resources if they do to prevent disruption and in the worst case scenario to capture the event the first time so support can diagnose and remediate without requiring the customer to experience it a couple of times to capture traces or other diagnostic data due to a lack of "Fault Isolation" technology.

It is this latter aggregation of multiple workloads where multiple VM's drive a servers utilization levels high is why the performance per Watt does not matter with Power servers. It matters with x86 where there they have the need for horizontal scaling due to unpredictable reliability and lower overall server utilization levels that I see typically not higher than 35% (being generous). So, if you can consolidate 4, 8, 12 or 16 x86 servers into a single Power server and that Power server with 2 x 200 W processors totaling 400 W overall vs 16 servers with 2 x 95 W processors in each server totaling 3.040 W does it really matter that the Power server uses 400 W? x86 sellers, Power haters and the uninformed (often times one in the same) will attempt to treat Power technology like x86 technology. A core is a core. Hypervisor is a hypervisor, server is a server. When they marginalize it like this by taking away the value proposition they reduce Power to being like x86 vs a premium offering at x86 price points.

There are times a business may deploy a single Power server running a single workload but they the server has very sophisticated tmpd or thermal and power capabilities allowing you to know how much power it is using, control those levels (to a min or max), etc. Plus, the servers are very efficient in being able to shut down portions of the server that are not being used. So, just because the processor is capable of using 200 W doesn't mean it is "using" 200 W. Make sense?

That is all I had to say. I will leave it with a few other comments unrelated to the above since I said so much. The high end of the Power8 family is still forthcoming. It should be pretty impressive and from what I am hearing put the 1 & 2 socket servers in their place :) It will bring more performance, scaling, RAS, security, virtualization features in support of Linux, AIX and IBM i. I'm guessing Oracle will release a new SPARCcluster shortly afterward that will outperform it with 1200 SPARC T5 servers and 24,000 cores cost $75M if they used real customer pricing but instead used "fruit fly" licensing terms for the software - they will ignore that and just focus on the fact that they have some world record - ah Larry and his marketing department! Back to the subject though, speaking of Linux, I'm a AIX and IBM i guy but this Linux stuff is cool. The Little Endian stuff is the real deal. Recompile your x86 source code in Ubuntu then run it. Watch it outperform x86. Now you get Power security, virtualization, consolidation, efficiency, RAS, and cost savings on fewer and fewer servers - this ALL translates into lower costs. Lower IT costs means more business capital!

Please leave your comment:


Get WordPress Plugins for easy affiliate links on Stock Tickers and Guru Names | Earn affiliate commissions by embedding GuruFocus Charts
GuruFocus Affiliate Program: Earn up to $400 per referral. ( Learn More)
Free 7-day Trial
FEEDBACK
Email Hide