gnu@l5.uucp (John Gilmore) (11/23/85)
In article <1795@peora.UUCP>, jer@peora.UUCP (J. Eric Roskos) writes: > More generally, ...CPUs built with discrete > components (as vs. all on a single IC) tend to be faster. This was > something I was skeptical of myself back during my days of faith in > the microprocessor, and I still tend to believe that is true largely due > to practical considerations than theoretical ones. There are plenty of counterexamples to this, too...e.g. the Vax 785 and Sun-3, the IBM 1401 and the Apple ][, etc. But let's restrict the discussion to things that were designed at the same time, in which case you're mostly right. I think the reasons are complexity and generality. VLSI chips can only handle so much complexity -- once you exceed the number of gates you can fit, you have to go to a board-level design. This was the constraint on Vax-on-a-chip for a number of years. This is especially true in very fast logic families like ECL. The generality comes in when you are trying to make money at VLSI CPU design. You can't afford to design it exactly to one application's specs, so it won't be as well suited for that application as something that *was* designed down to the PALs and gates for that application. On the other hand, due to hardware production volumes and software availability, the "nonoptimal" VLSI CPU design may be a lot cheaper to build, so cost/performance can come out ahead. Another way to look at it is: If you could get the speed you wanted by designing in a VLSI CPU, why would you *bother* to design with discrete ICs? That's why discrete CPUs are faster than VLSI CPUs. :-)
dave@ur-helheim.UUCP (David F. Carlson) (11/26/85)
In article <277@l5.uucp> gnu@l5.uucp (John Gilmore) writes: >In article <1795@peora.UUCP>, jer@peora.UUCP (J. Eric Roskos) writes: >> More generally, ...CPUs built with discrete >> components (as vs. all on a single IC) tend to be faster. > >I think the reasons are complexity and generality. VLSI chips can only >handle so much complexity -- once you exceed the number of gates you can >fit, you have to go to a board-level design. This was the constraint >on Vax-on-a-chip for a number of years. This is especially true in very >fast logic families like ECL. > > blah blah blah blah blah blah blah blah blah blah blah blah blah blah >That's why discrete CPUs are faster than VLSI CPUs. :-) It seems to me that the very "discrete" CPU designs which your boast are faster here are constructed of VLSI components. The bitslice and large gate arrays which are essential to the "modern" "discrete" realization have more complexity on them that those silly 8 bit machines we all confess to have designed/played with in the 70's. Comparisons involving ECL bitslice and NMOS VLSI will get an apples and oranges response from me. ECL is electrically very different that NMOS. Were I to design a 68000 on a chip using ECL I could heat a several homes before the devices melted into slag silicon. Fabrication and packaging complexity limit ECL VLSI more than an inherant "complexity" limit. Now, look at the 68020 vs. (custom) VAX 750's CPUs in the dhrystone benchmarks. The architectures are similar (bulkwise: registers, memory management, intruction types, etc.) When a Clipper system is built (although this is a chip set :-) another clean strike for VLSI will be struck. I welcome this debate. Any takers?? -- "The Faster I Go the Behinder I Get" --Lewis Carroll Dave Carlson {allegra,seismo,decvax}!rochester!ur-valhalla!dave
ron@brl-sem.ARPA (Ron Natalie <ron>) (11/27/85)
In article <1795@peora.UUCP>, jer@peora.UUCP (J. Eric Roskos) writes: > More generally, ...CPUs built with discrete > components (as vs. all on a single IC) tend to be faster. This was > something I was skeptical of myself back during my days of faith in > the microprocessor, and I still tend to believe that is true largely due > to practical considerations than theoretical ones. > I have a big black thing in my computer room which is a supercomputer built out of ECL MSI chips. It's about the best that could ever be done with discrete components. The distance between the chips is just too far for the speed of the electron (when you are dealing with nanoseconds, the farthest you can go is a little under a foot). -Ron
zben@umd5.UUCP (11/27/85)
In article <555@brl-sem.ARPA> ron@brl-sem.ARPA (Ron Natalie <ron>) writes: >In article <1795@peora.UUCP>, jer@peora.UUCP (J. Eric Roskos) writes: >> More generally, ...CPUs built with discrete >> components (as vs. all on a single IC) tend to be faster. This was >> something I was skeptical of myself back during my days of faith in >> the microprocessor, and I still tend to believe that is true largely due >> to practical considerations than theoretical ones. >I have a big black thing in my computer room which is a supercomputer >built out of ECL MSI chips. It's about the best that could ever be done >with discrete components. The distance between the chips is just too far >for the speed of the electron (when you are dealing with nanoseconds, >the farthest you can go is a little under a foot). Originally MOS (mostly NMOS but sometimes PMOS) was used to implement micro- processors because gates could be made smaller, and there was a problem cramming enough gates onto a reasonable sized chip to implement a full non- trivial microprocessor. Also, the power dissipation was lower, making the heat-dissipation criteria a bit less critical. But, the price that was paid for this was speed - MOS couldn't go quite as fast as STTL. Remember that MOS uses "Field effect" transistors, not "Junction" transistors. One major differance is that "Junction" transistors are CURRENT MODE (low impedance) devices, while "Field effect" transistors are VOLTAGE MODE (high impedance) devices. This implies that each Pf (pico-farad) of parasitic capacitance takes longer to charge and discharge with MOS because less current is available to charge or discharge it. The big promise of SOS technology (silicon on sapphire) was that these capacitances would be greatly reduced because the substrate was farther away from the conduction lines. With today's modern production techniques I don't know how much this remains the limiting factor. But, I haven't seen any TTL microprocessors lately, have you? Do you remember the MPY16 parallel 16 by 16 multiplier chip? This animal is TTL and quite fast, but it also has a mushroom-cloud-shaped heat sink growing from it's top, and from all reports it seems to need it! -- Ben Cranston ...{seismo!umcp-cs,ihnp4!rlgvax}!cvl!umd5!zben zben@umd2.umd.edu
kds@intelca.UUCP (Ken Shoemaker) (12/07/85)
> In article <1795@peora.UUCP>, jer@peora.UUCP (J. Eric Roskos) writes: > > More generally, ...CPUs built with discrete > > components (as vs. all on a single IC) tend to be faster. This was > > something I was skeptical of myself back during my days of faith in > > the microprocessor, and I still tend to believe that is true largely due > > to practical considerations than theoretical ones. > > There are plenty of counterexamples to this, too...e.g. the Vax 785 and > Sun-3, the IBM 1401 and the Apple ][, etc. But let's restrict the > discussion to things that were designed at the same time, in which case > you're mostly right. > > I think the reasons are complexity and generality. VLSI chips can only One more things: packaging technology. It is a lot easier to put two 128-bit wide data buses on a mainframe than it is to put the same on a chip with 200 pins. This also has implications on the speed of the output pins in that if you try to switch too much too fast, you get pretty large voltage transients on the chip, which don't sit too well (so you add more power pins...) -- yes, some uncomplicated peoples still believe this myth... Ken Shoemaker, Santa Clara, Ca. {pur-ee,hplabs,amd,scgvaxd,dual,qantel}!intelca!kds ---the above views are personal.