Very interesting original post and comments section.
* In my opinion, IBM's value is in its ability to control and optimize the hardware AND software elements. The ROI on the POWER architecture should be judged on the systems and software sales this underlying architecture delivers. I think the decision IBM took to not take on x86 in the general server market was the right path. Therefore (and you could argue I am a hardware guy and am thereofre biased!) a shift to software only is a dangerous path. I view those solutions as less "sticky" - opens IBM up to more competition
* I don't believe it is mandatory to own a fab to build compelling high end server products...However, if you don't have your own fab, you certainly need to be a large and influential enough company to get your custom process tweaks implemented that make your particular solution shine
1. For a really competitive CPU, epecially in the server space you need you own fab. Qualcomm, Apple et al are commodity parts and do not really need to fine tune processes. And henec are not valid examples. Oracle is suffering with power issues partly becuase it does not have optimal silicon processes. Intel inspite of a bad server arch. (till recently) partly suceeded because it owns the fab.
2. Having dealt with most fab vendors including IBM since the late 90s. Currently for our processors we use TSMC, UMC and possibly ST for FD-SOI. I can safely say that IBM's problems arose solely due to the way they marketed their fab and Power parts. Even today, every large RDBMS installation in India, uses Power 7+/8. There is a reason for this. Most RDBMs need both threaded and single core performance (as Oracle found out with the T series ) and IBM probably is the best at this. This has a direct bearing on licensing costs which can shoot up 50% with an x86 part.
IBM's mistake was two fold, one of which has been corrected. The first was not to agressively go after Intel in the processor market. They would have easily got 2-40% of the market. Even now if they go commodity with Power 8 they should get a good chunk of the market. Commodity means selling on Newegg. They need to get over their obsession with high margins. Let us hope the Open power strategy is not too late. The market needs an alternative to x86 and Power is technically the preferable alternative. ARM will not count till 2017/8.
The second was not pricing Power systems reasonably but at least now Power boxes with RH have price parity.
3. It amuses me that folks have such religious attitudes towards vertical integration. There is nothing wrong with a vertically integrated IDM. All you need to make sure is that you use up your capacity and your parts are priced right. This is very much a cultural thing. Vertical integration went out of favor in the West after the collapse of the likes of ITT. But Asian business models do not necessarily subsribe to the mantra. In these hyper-competitive times, every piece of differentiation counts. IBM's incompetence in monetising a fab should not be considered as a morality story on the non-viability of the IDM model. Such generalizations belong to case studies at Harvard and have no place in the informed discussion forums of EE Times !
4. On a more technical note, the eDRAM in a Power 8 makes all the difference in the world. Code an RDBMS engine and you will never go to a processor with low levels of L3/L4 cache. If buffer caches could speak, they would agree with me ! Granted there are other ways to get a large L4 cache, say via TSVs but it does not work for L3. In a large core count HPC/server processor, a segmented L3 is key to performance and it has to be on-die. Given the capacities needed for server app caches, eDRAM is the only way to go. I cannot stress this enough, pipeline and TLB flushes are catastrophic in high perf. low latency server SW.
See my point 2. I deal with large banks and deal sizes in the 100-200 M USD. In the core server category, most RFPs specifically blacklist x86 processors. I think this will change with Broadwell and more RAS features but points to the fact that IBM blew it purely on Marketing. In all these deals, the app and webs servers are invariably x86 boxes becuase the hassles of dealing with IBM. If IBM makes owning a Power box easy, a lot of my enterprise deals would have gone the IBM way.
5. ARM is coming along well but they still does not have a high end server arch. in place. They will get there for sure, I keep running into ARM folks in some of my Working Groups so they are going in the right direction. But I think it will be 3-5 years before they are taken seriously.
Qualcomm, Oracle, Apple, Digital Equipment and others have all shown that competitive processors can be built in foundries. The real issue is the cost of IBM retaining its own fabs/process vs. the cost of foundry services. As zSeries and pSeries sales have fallen, and the game business was lost, the cost of IBM maintaining its own fabs has become untenable, and transition to future technology nodes unaffordable.
This article is written as if the design space is very limited. Since IBM is in the system business, there are many design options to achieve overall system cost, reliability and performance. For example, rather than foundry-supplied eDRAM, it might be more cost effective to use RAM on a silicon interposer or RAM attached to the processor chip with TSVs.
Embedded DRAM (I prefer eDRAM as an abbreviation since EDRAM was previously used for Enhanced DRAM) has significant density advantages, but involves significant addition manufacturing cost. This is somewhat like incorporating flash memory into a computation-oriented chip—it adds cost and may have delayed availability or delay the roll out of a new process—, but incorporating flash can more easily reduce system cost for small systems by removing the need for external persistent storage.
If an eDRAM-capable process is available sufficiently later at sufficiently higher cost, it can be attractive to use a smaller ordinary process and use SRAM since the smaller process will shrink and increase performance/energy-efficiency of the logic and make SRAM density more comparable to eDRAM in the older process.
eDRAM is most useful when bandwidth or latency demands integration but the desired capacity would be too great for practical integration if SRAM was used. (eDRAM also has some error resiliance advantage over SRAM.) This makes it great for huge L3 caches integrated into the processor chip (or even just in the same multichip module as the processor chip). It could also be useful as a high bandwidth memory for a GPU.
The lack of great success of Mosys' 1T SRAM (higher performance DRAM abstracted behind an SRAM interface) hints that the market for a memory with its capacity advantage is not especially large. (I seem to recall it was used by HP for a PA-RISC processor's off-chip cache.)
For a foundry eDRAM seems to be something of a niche feature (more so than support for integrated flash memory). Without IBM's server demand for such, the market may not be able to support the extra costs. Process development is already expensive and risky, so Global Foundry's rejecting the extra cost and risk for such a niche feature is understandable.
(I also seem to recall that IBM's eDRAM techology was linked with SOI. SOI also adds cost and seems to be less popular among foundry users.)
A technology can be theoretically very useful ("great") without being very profitable.
By the way, IBM's POWER7 exploited the presence of trench capacitors to provide better power regulation. This additional advantage might expand the niche slightly but only slightly.
If EDRAM is so good, and IBM is talking about selling it to someone, doesn't it give IBM levearge to ask for specific conditions , for example an ability to purchase wafers using EDRAM at a reasonable price ?
And if that's the case, it's most likely a secret from us and company employees ,as common in negotioatios.
And if the tech is so great , why wouldn't global foundries use it ?
IBM's name is seen synomynous with hardware business. Its hard to believe that IBM can be only software development company. There are numerous software companies already being there. IBM stands out from crowd because of their hardware experience.