blanken@uiucdcsb.cs.uiuc.edu (10/21/87)
I was reading a magazine, and at one point the article compared a number
of computers (Hypercubes, parallel and various architectures) by listing
different speeds. I understand that all computers have clocks that run
at different speeds or cycles (Mhz...etc). But the article was using
terms like Mips, Mops, Mflops. Discarding the M as Mega,
ips = instructions per second
ops = operations per second
flops = floating point operations per second
I was wondering how these items relate to clock speed. I understand
that if you have a 16Mhz clock, you don't have 16Mips...etc.
Can anyone help me figure out the difference among these quantitative
descriptions?
Eric Blankenburg ARPA: blanken@a.cs.uiuc.edu
University of Illinois CSNET: blanken@uiuc.csnet
at Urbana-Champaign UUCP: {ihnp4, pur-ee, convex}!uiucdcs!blankenalter@ttidca.TTI.COM (Steve Alter) (10/26/87)
In article <164600007@uiucdcsb> you write: > ips = instructions per second > ops = operations per second > flops = floating point operations per second > >I was wondering how these items relate to clock speed. I understand >that if you have a 16Mhz clock, you don't have 16Mips...etc. I don't know the difference between ips and ops, but I do know why this form of speed metric is prefered over simple clock-speed. The following story is *very* oversimplified but shows the basic concepts. Take two machines by different manufacturers. Let's call one Fred and the other one Sue. Now Fred considers himself a real speed-demon with a clock that zips along at a whopping 40Mhz. Sue, however, takes her time at a leisurely 18Mhz. If we port the same operating system to both machines, with supposedly the same effort expended on each, then you might think that Fred will produce the greater throughput? We don't know yet. More facts: Fred is a machine with a highly complex and convoluted instruction set, and a great many addressing modes and other bells and whistles. Because of this complexity, he is a micro-programmed machine which means that his instruction set is really just a "language" that gets interpreted by a micro-program whose operations are the real hard-wired ones. Micro-programming is a multi-step process, and we can suppose that a typical operation such as an ADD or MOVE, with all of the complex addressing mode handling could take as long as 20 clock cycles (let's call the average 16.) Sue is a RISC (Reduced Instruction Set Computer) processor. She is not micro-programmed which means that her instructions are executed directly by the processor which is feasable because her instruction set is much simpler. She can exeute 90% of her normal operations in exactly 2 clock cycles. Since she doesn't have all of Fred's fancy addressing modes, she requires maybe 2 or 3 instructions to perform what Fred can do in 1 instruction. Now let's put the numbers together: Fred cranks at 40Mhz but takes 16 cycles to do an instruction. Net speed: 2.5Mips. Since Fred can do quite a lot in that one instruction, let's call that amount of work one "operation". Thus on Fred, 2.5Mips is the same as 2.5Mops. Sue moseys along at 18Mhz and requires 6 cycles to perform one "operation". (That's 2 cycles per instruction multiplied by 3 instructions per operation.) Net speed: 3 Mops. The result is that Sue is a slightly faster machine. On a separate point, the ratio between ops and flops can tell you quite a lot about the machine and what it's best suited for. For example, a small machine with no floating point instructions at all have to do all the work using subroutines and have a very low ratio of flops to ops (i.e. close to zero). There exist machines, however, with such fast floating point hardware, that the ratios of flops to ops aren't that far from 1!. -- Steve Alter ...!{csun,rdlvax,trwrb,psivax}!ttidca!alter or alter@tti.com Citicorp/TTI, Santa Monica CA (213) 452-9191 x2541 -- -- Steve Alter ...!{csun,rdlvax,trwrb,psivax}!ttidca!alter or alter@tti.com Citicorp/TTI, Santa Monica CA (213) 452-9191 x2541