I am completely against model numbering systems since they are filled with inaccuracies, purposely misleading terms, and often flat out lying. And by the way, AMD wasn't the first company with model numbers (so Intel isn't copying AMD). Problems with model numbers:
1) Inaccuracies. Suppose you have three programs in your benchmark (I'm keeping this simple, but the same is true with 100 programs). Suppose chip X gets a score of 1000 on all three programs. Thus chip X is given a model number of 1000. Now chip Y comes along and gets a score of 900, 1200, and 1347 on those same three programs. What do you call chip Y? Do you be conservative and give it a 900 rating? Do you be aggressive and call it a 1347 rating? Do you call it a 1149 (the average of the three)? Do you weight one program more than another? Which programs do you use, etc? Basically a model number won't work very well over a wide range of programs since programs and chips are so complex. No matter what you do, you'll leave things out and overemphasis others.
Lets suppose we choose the simple average: 1149. Marketing now comes in. They complain that consumers won't want a 1149 model number. Things must be in round 100 numbers. So do you round this chip down to 1100? Do you round up to 1200? 1200 was the median score and 66% of the programs were 1200 or more. No matter what you do, it'll be inaccurate. Benchmarks don't give round numbers - marketing requires it.
2) Purposely misleading consumers can be rampant. For example AMD claims to have kept the same benchmarks in their PR ratings. Yet several of the programs have been modified and several more added. What does AMD do with these added programs? None of us knows. The claim that the benchmarks are the same is a flat out lie. But things are even more subtle than that. Did you know the AMD ratings are not CPU ratings? Nope they are CPU + motherboard + memory ratings. If a new memory standard comes out and a not yet released AMD chip scores 10% better, then the AMD rating on that chip goes up by 10%. Thought these were CPU ratings? Nope, you are mislead. The 3000+ XP does NOT score double what the 1500+ does in the benchmarks if they are run on the same motherboard and use the same memory - in fact it is significantly less than double.
3) Flat out lying. This I'll leave open to your imagination. It is very easy to fudge a rating score to meet the companies need when consumers (or reviewers like Anand) do not have access to the secret formulas used in finding the rating number. Are the programs weighted evenly? I'd love to know that. Even better: do the weighting remain constant from CPU to CPU? One slight tweak to a program's rating could dramatically effect a number. It is clearly evident that a simple program recompile can add 30%, 50%, or more to a CPUs speed. Will this be done (much like the great potential of the SPEC numbers which have turned into a compiler test more than anything else)? I'm not saying anyone will do it, but the possibility is endless.
I guess I'll add a #4. Model numbers main advantage would be for comparing dramatically different chips. Unfortunately using a different rating for different chips makes this useless. We all know the AMD64 3200+ is far faster than the Barton 3200+ XP, but you'll never know it from the model number. The consumer NEEDS to know Celerons are crap, the model number could help, but it will be done in a way that the consumer still won't know the real weaknesses of the Celerons.
Facebook
Twitter