Xref: utzoo comp.misc:1728 comp.sys.m68k:686 comp.sys.mac:11376 comp.sys.ibm.pc:10884 Path: utzoo!utgpu!water!watmath!clyde!cbosgd!ihnp4!homxb!hropus!ki4pv!codas!usfvax2!pdn!alan From: alan@pdn.UUCP (Alan Lovejoy) Newsgroups: comp.misc,comp.sys.m68k,comp.sys.mac,comp.sys.ibm.pc Subject: Re: The New Chips Message-ID: <2101@pdn.UUCP> Date: 25 Jan 88 18:41:41 GMT References: <4746@watdragon.waterloo.edu> <1430@husc2.UUCP> Reply-To: alan@pdn.UUCP (0000-Alan Lovejoy) Organization: Paradyne Corporation, Largo, Florida Lines: 70 In article <1430@husc2.UUCP> sipples@husc2.UUCP (Timothy Sipples) writes: >>IBM is sticking with the Intel line, so the 80386 will probalbly be more >>than the 68020. > >Whoa! Not to diminish IBM's proven wisdom :-), but other manufacturers think >very differently. Ever heard of Apple? Sun? Not to mention the brains >behind Postscript laser printers, including, I believe, IBM's. Also keep in IBM also uses a modified 680x0 (I forget which) and a stock 680x0 together in the AT/370 to emulate the S/370 instruction set. And in the infamous RT, the 68881 FPU is the standard floating point co-processor. IBM's GX graphics accelerator is based on multiple 68020's. The 68020 outsold the 80386 in 1984 (the '386 didn't exist), in 1985 ('386 only in sampling beginning late summer), in 1986 ('386 in full production by late summer) and even in 1987 (thanks to the Mac II and to the '020's wide use as a controller in industrial/commercial devices (such as laser printers)). The '386 is PROJECTED to outsell the '020 in '88, but then, that was the projection for '87. If Apple releases an "SE" type Mac with a '020 in it (say at 12 MHz costing $3000 for the complete system), then the projection may turn out wrong again. Even so, the installed base of '020's is now in the millions and its software base completely dwarfs 32-bit '386 software. 32-bit software for the 386 is not expected to be available in quantity until 1990 AT THE EARLIEST!! There's hardly even any '286 protected-mode (16 Meg address space) software yet! All those benchmarks you've seen showing how fast the '386 is on the Dhrystone (3500-5500 Ddhrystones/second for 16MHz CPUs) suffer from two flaws: 1) Those numbers come from code produced by highly optimizing compilers generating '386 native mode (32-bit) code (most CURRENTLY AVAILABLE application code is for the 8086); 2) the Dhrystone spends most of its time doing string comparison, which the Intel CPU's have special instructions for and the Motorola CPU's don't: and the string comparison operation is assumed to be a primitive whose algorithm is not specified by the benchmark!!!. On a graphics benchmark (requiring bit-field manipulation), the Motorola CPU's leave the '386 in the dust. The Mac II executes the Dhrystone about 4 times faster than the Mac Plus. The '386 machines execute the Dhrystone about twice as fast as the '286 machines. Using the best optimizing compiler available, the Mac II does about 3300 Dhrystones/second. A SUN 3/260 (25MHz 68020) or an Apollo DN/4000 do about 7000 Dhrystones/second using optimizing compilers and static ram caches. The NS32532 reportedly does 16,000 Dhrystones/second (at 30 MHz). The 68030 should produce similar numbers. SUN's 4/260 (SPARC "RISC" CPU) does about 20,000 Dhrystones/second. By the time the Intel/MS-DOS/OS2 world is getting a reasonable supply of '386 code on 30MHz machines (1990; 12,000 Dhrystones/second), the Motorola world will have a larger supply of 68040 software on 40 MHz machines doing 70,000 Dhrystones/second. Really. No kidding. The 68040 will implement most of its instructions using combinatorial logic instead of microcode, and will thus reduce the average number of cycles per instruction from 5 or 6 (on the '030) to almost 1. That's a speed increase of 500% just from that (larger on chip caches, faster clock rates, 3-operand instructions and other changes may mean additional speed ups). Intel likes to talk about compatibility. But it's Motorola who have actually delivered it from the 68000 to the 68040, without needing special compatibility modes. The fact that the 80386 executes three different instruction sets (8086, 80286, 80386) is sugar coating on a fundamentally poison pill. As time goes on, this will become more apparent. --alan@pdn