Intel (INTC), as we all know is the world's largest semiconductor manufacturer by dollar volume. Intel's history has been a story of world changing success with a couple of tooth loosening speed bumps along the way.
Intel's latest and greatest PC CPU chip is called the Haswell. Haswell is a new architecture, distinctly different from the preceding Ivy Bridge and Sandy Bridge architecture; as such Intel has a lot riding on Haswell. On Monday a report surfaced that Intel was having a problem with the power management section of the Haswell. Intel has since denied that Haswell has any problem and is on schedule and everything is wonderful. I guess we will find out the truth in a couple of weeks when Intel reports earnings. I'll be watching to see if the JPM analyst asks THE question.
White Paper - Platform 2015
While trying to understand more about the alleged problem with the Haswell voltage regulator, I ran across this 10-year Intel technology road map (pdf). Notice that, while the document is an Intel white paper, the source of the document is not Intel but the University of Potsdam. Look as I might, I simply can't find this document on the Intel website. I assume Intel has taken the document off the website because of the references to Itanium and Pentium 4, both of which are (or should be) a great source of embarrassment for Intel. Obviously, Intel makes blunders; just as obviously Intel recovers from blunders. Companies that don't make mistakes are not taking enough chances.
Platform 2015 was apparently written around 2005 as a 10-year look down the technology highway. Let's see how it did so far.
Page 3-4, Macro Trends of Pervasive Connectivity, Multi-core processing, Data Growth, and Smart Internet:
The paper is remarkably prescient in recognition of these trends, if anything the growth in these areas is understated relative to what we have experienced during the past eight years.
Page 5, Processing Implications...
Intel recognized that General Purpose Performance was still the driving issue. From Intel's viewpoint, the Macro Trends could not continue without massive performance improvements in data center equipment. As it pointed out in the introduction, very few companies would be in a position to address these future issues and Intel is one of them.
So, Intel had a choice in 2005: Solve the BIG problems (servers) - or- make cellphone chips; cellphone chips that wouldn't work unless the BIG problems were solved. Thus Intel spent the next several years developing very high performance server solutions. I find it ironic that pundits claim Intel "Missed the Mobile Business" while it was busy providing the infrastructure technology for the very business it is supposed to have missed.
Power Management: Power Management is a pervasive topic throughout the paper. Intel was early to recognize that power efficiency, on all levels of computing, would become critical to continued progress. So, the pundits who believe that Intel was not concerned about power consumption of computing devices are simply wrong.
Large Memory Subsystems: In the paper, Intel acknowledges the "Memory Wall," which is the inability to increase processor performance due to limited memory bandwidth. The solution to this is more on-chip memory, "in the gigabyte size range, replacing main memory in many types of computing devices." Decoded, this means that main memory (DRAM) will go on-chip or very near the CPU, perhaps in the package with the CPU chip. I can't see how Intel would allow anything that must go in the CPU package to be supplied by outside sources. I guess the probability of Intel produced DRAM just increased. Today, 4GB of DRAM takes 500 sq. mm of silicon, over three times the silicon used in the CPU chip itself. Those Intel Trigate transistors might be instrumental in shrinking the DRAM silicon down to 125 sq. mm or so.
Page 9 tells us about the HKMG (high k/metal gate) process, which went into production a couple of years later. It also freely mentioned Trigate transistors that went into production six years later. Futures include III-V transistors, carbon nanotubes and silicon nanowires. Intel is not sitting still. Billions-of-transistor chips were forecast. The average Ivy Bridge of today contains 1.4 billion transistors, server and HPC chips and several billion transistors.
I could go on, interpreting this interesting report that is very hard to find, but the point is that Intel has a very long planning horizon and many of the things discussed in the report have come to pass many years later and some others are now just around the corner.
Creative destruction is the way of life in the technology industry. The most recent and glaring example is the near demise of BlackBerry (BBRY) at the hands of a superior product from Apple (AAPL). The list of companies that have gone from first to worst or zero to hero is too long for this article.
Back to the Haswell voltage regulator chip. The Haswell PLATFORM includes the Haswell CPU and the regulator chip. That regulator chip is built on Intel's 22nm process and, even still, is a 105 sq. mm chip. That is two-thirds the size of the current CPU chips; it is a huge chip. We used to call chips that size "curbstones." It has up to 320 "phases." A phase is a channel of voltage regulation (power management). That chip will manage the power consumption of the Haswell to the point that the vast majority of the CPU will be turned off most of the time. This level of control makes a joke of the ARM (ARMH) BIG-little approach.
Now here's the creative destruction part. Currently, a wide variety of companies provide power management chips to the PC makers. There is about $20 worth of silicon used to manage the power consumption of every Ivy Bridge. With Intel sucking that function into the CPU package, these small suppliers will lose business. Intel, on the other hand, by improving the power management with a chip of its own, is entitled to the $20 value of the power management chips that will be replaced. Even in a horrible PC market, Intel will ship 300 million CPU chips, and, presumably, 300 million of those $20 voltage regulator chips as the Haswell replaces the Ivy Bridge. That is $6 billion worth of new revenue for Intel. It also means $6 billion of less revenue for companies like Texas Instruments (TXN), Maxim Integrated Products (MXIM), Linear Technology (LLTC) and others. Six billion dollars is more business than the Apple foundry business would be; $6 billion also represents 20% growth for the client (PC) sector of the Intel business. Six billion dollars is about 40% of the total sales of TSMC (TSM). This is a big deal that is being ignored.
The next candidate for large-scale creative destruction is the PC DRAM discussed above. Integrated DRAM will provide, by far, the largest increase in performance and power saving in the CPU/memory section of a PC. That is precisely why the mobile DRAM in virtually every smartphone is mounted on top of the application processor.
As I mentioned above, 4GB of DRAM consumes about 500 sq. mm of silicon. Today that amount of DRAM cost about $25 or about $.05 per sq. mm. Now, Intel's cost on 22nm wafers is about $.085/sq. mm. The only way this is going to work is if Intel, through the use of 3D structures (think Trigate), can reduce the size of 4GB of DRAM to about 125 sq. mm. Then, at the same $25 for that 4GB of DRAM, Intel would get $.20 per sq. mm of silicon or 57% gross margin - for DRAM! Again, we would see $7.5 billion worth of DRAM business move from the present suppliers (Samsung, Micron (MU), and Hynix) to Intel.
These companies will switch production from DRAM to NAND flash, which will help with the expected explosion of solid state drives.
The point of these two examples of creative destruction is that, once started, both cases ramp to full volume in a relatively short period of time, like a year.
Just perhaps it is this type of business that will fill the immense amount of capacity in place at Intel.