Dojo (HowTo)







  Easter Eggs




  Martial Arts

How a design philosophy effects the real world.

By:David K. Every
©Copyright 1999

To all that don't understand the difference it is as simple as this -

CISC is the traditional design philosophy. Make lots of instructions based on what engineers think is cool, and then compiler writers use a small sub-set of those instructions and the rest sit around taking space or are occasionally used by assembly programmers. But once added these instructions can not be removed (legacy) - so most CISC designs are saddled with 20 year old instruction sets, complex instruction designs, lots of gates (switches and space on chip) to get work done, and are often forced into using micro-code (an emulator for hardware -- which slows down execution) to get all the instructions on chip in the time or space requirements.

RISC is the design concept of simplifying the instruction sets complexity (not necessarily the number of instructions) so that designers can use those freed up gates (switches - which equals space on the chip) to do other things. Usually those gates are put to work making the whole chip faster (like using super-scalar, super-pipelining, caches, branch prediction, caching, etc.). Simplifying the instruction set also means that you can bring a chip to market faster and take advantage of newer processes with less costs... or that you could design multiple specialized flavors of a chip for the same cost.

So you can see it is all a philosophical point of view.

Some notes on these differences

Intel CAN compensate for the inferior design philosophy by just pouring more money into design (to a point). But the yield per dollar is still better for RISC. Intel has been putting more money into design - the question is have they been putting in enough more to compensate for the difference.

Intel can NOT break their CISC based legacy instruction set - so they are saddled with their legacy. Since RISC processors are new lighter designs, and are fast enough to emulate the older processors (at the same or higher speeds) the migration was pretty painless to go to RISC when Apple did. It will be more painful for Wintel. Intel is going to try to solve their problem with Hardware emulation (instead of software like Apple). Which means better initial emulated performance, but higher costs, lots of extra gates being wasted, and slower adoption. (The less dramatic difference between old and new performance the less "forced" they are to adopt the new - and the slower they will do so). Intel is going to have to drag old instructions around even longer AFTER they start to adopt a new design. That wasted space equals cost, heat and performance penalties -- there is no getting around that.

CISC can be fast, or CISC can be power-efficient - but it is hard to do both with CISC (compared to RISC). CISC machines just flat out require more gates to get the same work done - that means more heat/power. RISC is better for laptops and low end consumer machines where power consumption matters. (Why almost all PDA's and home appliance computers are RISC).

  • RISC machines have less complex instructions and a lot larger amounts of their real-estate devoted to Cache. Cache is easier to create (map out) than instructions, and less likely to have bugs. RISC machines therefore are less expensive to design and will have fewer bugs (for same effort) than CISC.
  • Memory and chip capabilities are growing dramatically. As the manufacturing keep growing the chip that is the easiest to design (RISC) CAN be the first to implement that technology (that does not mean they will always do so, just that they could be if they start from the same point). This is part of the reason why most of the big jumps in process or performance are seen on RISC first, and will continue to do so. Money can compensate for some of this - but time is continuous.
  • MMX is a way to make a processor MORE CISC like and MORE proprietary. Intels MMX philosophy is to add MORE complexity to the instruction set (which they will have to carry around forever). Those instructions will only be executed a small fraction of the total time. A better design philosophy is to instead of using $50 of gates on your processor to do this work (and tie up your processor in the process), use a $50 dedicated chip (and evolve it separately) to off load this task, do this task faster, and to leave your processor free to do other work during that same time. In other words - an MMX based Pentium will be slower at both processor and DSP functions than two chips specialized to do each. They will likely cost the same (possibly)- but the Intel approach ties each tasks evolution to redesigns of the other (you can only scale the technologies together). It also makes it harder to split the tasks up, and have multiple processors or DSP's. Also when one of the units is working on a problem it is more likely to have resource conflicts with the other or prevent the other unit from doing what it wants (in the Intel approach than two separate processors).

CISC vs RISC -- recent history

RISC has been winning most of the time. Look at the history.

Macs made the jump in '94 about the same time as the pentiums started coming out. The pentiums were expensive and used lots of power (were hot and temperamental) - the Macs had lots of emulated apps - but the critical apps were quickly native. Pentiums had a bug (History of PentiumBug) that was found out a year later and took a lot of time and money to replace (for both Intel and its users). Then they found another one in the PentiumPro's and PentiumII's. Then there are the multiple bugs with Intel Chip sets that are needed to support the Pentiums (on a computer). The complexity of CISC increases the likelihood for these errors and costs to the consumers.

PPC's have not only been faster at the same MHz (speed) but have been available at faster speeds.

Finally after about 1 1/2 years (after the PPC) and Intel just getting whooped in performance, Intel responded and announced the P6's (PentiumPro's). But they were not widely available (but highly publicized) - they only ran 32 bit Software well (or about 5% of the software out there) - and most of the early motherboard had serious design flaws (the Orion chip set). About 6 months after the P6's were available the PPC's had already surpassed them in MHz (and more in performance) and have been hovering at a fraction (usually 1/2) of the cost -- but people bought the hype of Intel over the fact of PPC.

Intel has been hyping MMX, but the results are in, and MMX just catches up to the last generation of PPC's in DSP/NSP functions -- the current generation often outperforms MMX in MMX specific functions (which are only a fraction of what a computer has to do).

PPC's are still available in much higher MHz --

  • 300MHz for the 603e, vs 233MHz for the Pentium w/MMX
  • 350 MHz in the 604e (Mach5) vs 266MHz for the PentiumII (and the PentiumII is slower than the older Pentium Pro/P6).
  • 300MHz for G3's -- and there is nothing close in the PC camp.

And all that isn't taking into account superior performance at the same speed. The Exponential Chip (a fast PowerPC derivative) at 500MHz couldn't keep up with 300MHz G3's or 350MHz 604e's in the real world (but blew away Everything in the PC camp) -- but in the Mac camp, it was a failure (and died). If the Mac market was like the PC market, those things would be flying off the shelf because they sound faster.

Pentium Chips are almost always behind schedule (P55's and Klamath have both been delayed, P55's are about 4 months late (but finally shipping) and Klamath was about the same). PowerPC's have been exceeding expectations -- both in delivery schedule, MHz, and performance (overall and per/cycle).

So out of the last 3 years (since PPC's were on the scene) it has varied from PPC's being way ahead, to PPC's being a little ahead. PPC's are the leader for 2 1/2 years - with CISC/Intel squeaking ahead for a few (say 6) months for a small sub-set of machines and applications. Then the PPC's once again flew ahead -- and Intel won't have anything to respond with until late '98 or '99 when they go to RISC in their next chip (called Merced). Of course the PPC's are expected to go to their next generation (the G4) as early as spring of '98 -- which is faster, true 64bit, a up to 4 processors on a single chip. Also PPC's are able to use newer manufacturing processes (like .2 micron or less, copper fabrication, etc.) to make it MORE cost effective, and faster.

And all during this time, the PPC's have been getting dramatically better in the low power consumption areas (portable, embedded controllers, etc.) and are drying up the market for Intel down there.

Even Intel has finally admitted that the Merced will be even MORE RISC-like to try to emulate the success of the PPC's. So you can buy the sizzle (and get an Intel inside), or you can buy the steak -- and get a PowerPC!

Created: 02/15/97
Updated: 11/09/02

Top of page

Top of Section