Slashdot Mirror
Apple Switch to Intel Not a Big Loss for IBM
KaushalParekh writes "An interesting article about how Apple's switch to Intel chips may not be that bad for IBM after all. "Apple sees an opportunity with Intel. But IBM continues the same chip development that allowed Apple to claim several firsts and fastests. Now, Big Blue will plow its research efforts into processors for game consoles and other consumer products that might one day knock the PC down a rung." Also, "a lucrative avenue for IBM in China, where the marriage of the Linux OS to PCs armed with [IBM] PowerPC chips presents some intriguing possibilities." And, "Large firms like Sony, Microsoft and Comcast are betting that a home-entertainment device, evolved from a game console or set-top box, will replace many of the PC's functions. IBM plans to be inside these new systems.""
I will play Mr. Smith: It was inevitable.
The cost of producing new leading edge semiconductor processors has always, and will always, grow at a very high rate. The cost of a chip fab is outrageous. It has been doubling for many years now. A new chip fab today costs about $4 Billion US!!! (For a 65nm fab) Most companies have started renovating there fabs to use ever larger wafers instead of building new fabs. Currently many fabs are using 300mm wafers. Many of the older 150mm fabs that have already upgraded to 200mm don't have the room to upgrade further.
Some big new 300mm fabs are:
D1C in Hillsboro, Oregon
F11X in Albuquerque, New Mexico
F24 in Leixlip, Ireland
The "break even" point for development costs has been skyrocketing!
In the 1980's and early 1990's 50,000 units were required to recoup development costs on a chip. At 130nm, fab costs hit more than $1 billion and the break-even point for chips was about 500,000 units. A chip made with a 65nm process needs approximately 5 million units to break even on development costs. With 45nm processes at the edge of the current horizon for chip manufactures, how many markets are there for a chip that has to sell 50 million units to break even!!!
For many years Motorola, AMD and other chip makers were moving steadily away from having there own fabs. Motorola was outsourcing 7% of its chip manufacturing in 1997, and over 30% in 2003. It seems this might have been a bad idea. Apple sited manufacturing capacity as one of its reasons for choosing Intel over AMD; and well, Motorola it seems just lost the Apple contract... (AMD is expanding its Dresden "Fab 36", and considering building a new fab to be opened early 2008. In the mean time they signed Singapore's Chartered Semiconductor to help make AMD64's starting in 2006.)
Currently chip fabs are operating at about 85% capacity; new fabs traditionally start getting built when 90% capacity is reached.
The biggest financial problem that chip manufactures face is the wavelength of light. For several generations of chips we have been using 193nm light. How is this possible you ask when chip feature sizes have been shrinking well below that size for several years now? Well I will let someone else explain:
Quote:
Since the 180nm technology node, the feature size has fallen BELOW the stepper wavelength. How can a 193nm wavelength of light expose gaps and widths that are 180nm wide? The laws of optics tell us that in order to resolve or "see" a gap of X nm in width, we must use a wavelength of light that is itself LESS than X nm in width. Today's feature sizes are down to 65nm and are still being printed with 193nm light! This seeming violation of the laws of physics and optics is being achieved by very clever techniques generally known as RET or Resolution Enhancement Techniques. Since the 180nm technology node, RET has been growing in cost and complexity from simple OPC (optical proximity correction) to PSM (phase shift mask) to the combination of OPC plus PSM, and now on to SRAF (sub-resolution assist features) which is ushering in a new category of RET called X-RET or Extreme-RET. The industry could have reduced the stepper wavelength from 193nm to 154nm, but a detailed analysis showed that simply shortening the stepper wavelength would be cost-prohibitive! Instead, use of 193nm has been extended to the 45nm technology node, but the gap between 193nm and 45nm is quite large and cannot be completely resolved even by the most advanced RET.
Fortunately, something called Immersion Lithography has been introduced. It has been tried before with mixed results, but the need for it has never been as urgent as it is now. By immersing the wafer in water, one can reduce the effective numerical aperture (NA), allowing 193nm light to act as if it were a shorter wavelength. The wafer now has to be immersed in water, however, and this creates new challenges for new types of resist and topcoat materials that can withstand the effects of water contamination. Today, however
Because they ARE inside these new systems. The Revolution will use an IBM CPU (PowerPC variant). The PS3 will use an IBM CPU (Cell). The Xbox 360 uses an IBM CPU (Microsoft even dropped Intel to switch to IBM). The war will go on for years but IBM has secured quite a beachhead in the opening salvo.
That is only if you exclude the Pentium-m from the mix. IIRC the Pentium-m has the best performance per watt for any desktop/laptop chip.
Sorry, teleporters just kill you and then make a copy. A perfect, soul-less copy.
Intel's Montecito is the first Itanium processor to feature duplicate, dual-thread cores and cache hierarchies on a single die. It features a landmark 1.72 billion transistors and server-focused technologies, and it requires only 100 watts of power.
Now 1.72 billion itself is beyond even IBM but managing power of 100W is definitely commendable. So please, do some research on latest work as well. Thanks
As little as 5 years ago, I would not have said this. Having a handful of PCI slots on board was essential for adding capabilities to a computer. Now, everything I might want to add comes in an external box - including TV capture hardware, sound interfaces, and even hard disks. I have a 500GB hard drive here connected via a FireWire 800 interface.
The only other computers I regularly use are a couple of headless boxes, one where I work and the other in a co-located server room. Neither of these need to be expandable because physical access to them is non-trivial.
I am TheRaven on Soylent News
None. It was all pushed forwards by the high-end server market. The POWER line (which are really PowerPC these days, just to confuse everyone[1]) has seriously high performance. They have been shipping dual-core chips since the POWER4+ (the POWER5 is currently top of the line).
The chips they are selling Apple are cut-down versions of the POWER4 (not POWER4+, or POWER5) with a vector unit bolted on. They are not the fastest chips IBM make, they are the consumer versions of an old IBM design.
[1] Originally, there was POWER and PowerPC. These were two slightly different instruction sets with a large common subset. It was possible to compile code that would run on both, or that would only run on one. More recently, IBM dropped the POWER instruction set, and recent POWER-series chips have been server-grade PowerPC chips.
I am TheRaven on Soylent News
Actually, there are 4 Mac clusters in the top 500 (only one in top 100).
It would seem that the advantages of a xServe cluster (altivec, 64-bit and good performance per watt) would be lost when Apple produces an x86 version when compared to a (inevitably) cheaper x86 Linux system.
By the time XServe migrates, I believe Apple will be using x86-64-capable CPUs in that line. Also, IIRC SSE3 has support for double-precision float arithmetic in its vector ops, whereas Altivec doesn't.
Also, Apple is pushing their Accelerate.framework, which should make vectorized code somewhat portable between Altivec and SSE3. Check Apple's guide to SSE for more details.