Lee Re: 1) If this is the case, how much more difficult is it for Intel to produce a chip of this complexity?"
It will be quite a bit more difficult for many reasons. First, it will be essentially a Deschutes AND an HP PA-RISC chip chip blended into one piece of silicon, along with the additional cost of adding these EPIC extensions for parallel execution.
Second, it will be a 64 bit machine - so all execution structures will be 64 bits wide and ALL METAL INTERCONNECT between structures will require 64 parallel data paths (lines/conductors). This will be an amazing amount of aluminum traces running around this chip!
Compound all this with a process migration to 0.18 micron minimum feature size and you have a Herculean task facing Intel.
I have heard that the device will be around 21 million transistors, roughly 3X the Deschutes, which is 131 sq.MM. That would imply a 393 sq. mm. chip size if the 0.25 micron process is used. Since Intel has announced that a new 0.18 micron process will be developed for the Merced and Deschutes follow-on devices, the die will be much smaller than 393 sq. mm.
Theoretically, if ALL layers went from .25 to 0.18 microns, a 49% die size reduction would take place, Realistically, we should expect 36% - the same as when Intel went from 203 sq. mm Pentium II to 131 sq. mm Deschutes.
That would make the Merced about 253 sq. mm in size. and result in 90 to 110 TOTAL POSSIBLE die on an 8 inch wafer (12 inch wafers will come in a few years).
If Intel yields 2 good die ( a very poor number) and sells them for $3600 apiece, they will more than pay for the wafer cost which should be around $3000 - $3500.
At 10 die per wafer and $3600 apiece, they will be back to printing money.
Re: "4) What are your thoughts on how Merced will be able to compete against IBM mainframes? Can it compete against only the lower level, or all IBM has to offer?"
IBM Mainframes will continue to serve the large institutions whose corporate information processing already runs on existing IBM hardware and software (Legacy systems).
But major new enterprise software from SAP, BAAN, Oracle, Computer Associates, Microsoft and SCO will be targeted for the MERCED - Big Silicon.
I'd guess that by 2001/2002 multiple Merced systems (16 to 64 CPUs) will begin to appear and begin to garner corporate MIS commitments for enterprise-wide deployment.
As you noted, IBM/Mainframes won't disappear - they will continue to be purchased to "boost" existing systems, but at a lower and lower rate. New system purchases (as opposed to upgrades) will go to MERCED based manufacturers.
Paul |
