IBM needs its own unique silicon process technology to drive its server business, but few -- even within IBM's top ranks -- really appreciate that reality.
The following opinion article was contributed anonymously by a current and long-time employee of IBM's semiconductor group in response to our recent series on the outlook for IBM as a chip maker.
Semiconductor process technology is vital to IBM's server business, but it's not clear even some of IBM's top executives understand that reality.
As an employee of IBM's microelectronics division, I am biased but I think there is still a place for our Z-series and P-series computers using processors based on that silicon technology. Perhaps our servers will not sell at the volumes they once enjoyed, but I think they could find a smaller market for banks that can't afford to go over to the cloud.
One of the key process technologies at IBM is silicon-on-insulator and one of its key uses is making trench capacitors for embedded DRAM. This technology offers huge competitive advantages and has been presented at top technical conferences such as ISSCC. EDRAM only really works if you use a trench capacitor not a stack as the ability to wire a stack capacitor would drive extra critical metal levels.
To a certain extent Intel recognizes the value of this technology in its own heavy use of memory with its processors. Interestingly back in the 64 Kbit DRAM days, IBM bought the trench process from Intel as it was exiting DRAM to concentrate only on CPUs. IBM at the time looked at DRAM not just to sell but also to drive lithography for its logic chips.
Having EDRAM for high-end L3 and L4 processor caches is really important. If IBM sold its microelectronics unit to someone such as Globalfoundries as recent stories suggest, it raises questions about whether the buyer would continue the EDRAM process and if so at what price. If the IBM server group has to pay for chips in merchant-market dollars, they may be so expensive that the server group will not be able to market machines competitively anymore.
I don't think that IBM's server processors can make the transition to using a bulk wafer with no EDRAM. Certainly we see when other designs try to move to other fab companies the success rate is not great. Apple has not been able to wean itself off of Samsung to move to TSMC. In fact the rumors are that Apple is looking at Globalfoundries now because its processes are close to those of Samsung so Apple's designs should port over easier than to TSMC.
Speaking of sharing process technology, the technology IBM develops with our partners is called the Common Platform but a better phrase would be a common foundation. Within a node there is both bulk and SOI developed staying as common as possible.
In the past IBM has kept the SOI and the partners took the bulk. However each partner has different tools in their fabs so the bulk process in one company won't be precisely the same as the bulk in another. Until the development of the fully depleted SOI, the partners did not use the SOI in their own offerings.
As for IBM's future as a chip maker, there is a lot of talk about IBM's 300mm fab in Fishkill being long in the tooth. I think this is wrong. The fab does not have a production EUV tool yet but then that tool does not exist. The company also lacks 450 mm wafers but then no one else is using them yet either.
IBM's 193nm wet lithography tools are on a par with any state of the art fab. So, I don't think the stock analysts are correct about the age of the Fishkill fab. What it does not have is the capacity of say the new Globalfoundries fab in Malta.
I think the view of IBM's chief executive Viginia Rometty is that IBM will become a software-only company. This view may be shaped in part by her degree in computer science. When she received that degree, CS was all about programming and not about the actual architecture of the CPU.
So I worry about our willingness to abandon the chip business. We give different messages. On one hand we promote the Open Power Consortium, but on the other hand we show we have no staying power, so why would anyone commit to it. At IBM headquarters in Armonk they are proud of eliminating business that does not perform to stated expectations.
My personal view is that the 22nm node that has just been announced may be the last new process node at IBM.