Originally posted by Tom West
How do we know it the benchmarks are fake?
Perhaps the most obvious problem with the benchmarks was the fact that Bryce, which does not use multiple processors, sped up substantially using a dual. There are numerous other somewhat questionable aspects to the benchmarks.
I'm not saying I believe the benchmarks but I haven't seen a really compelling reason why they are blatantly false. Most of the reasons are just as weak as the benchmarks. They offered the explanation that a more recent (beta) version of Bryce was used -- given that the rumour stated that Apple was doing these benchmarks this is entirely plausible since Apple commonly has pre-release versions of software for testing on new hardware.
The 80% figure is an educated guess based on the speculative SPEC benchmarks that IBM posted.
There are hundreds of different metrics that can be used to compare processors. IBM has almost certainly advertised its strongest rating, so I'm assuming that it's probably somewhat weaker in the non-advertised aspects and in a real world situation (i.e. Apple is not going to supply unlimited cache, fastest possible memory, etc.). The p4 with 800 MHz front side bus is a pretty hefty creature and is significantly faster than the previous P4.
IBM is usually fairly conservative with things like estimated
SPECmarks for a processor they don't yet have silicon for. They've gotten in trouble with things like this before so I wouldn't necessarily discount their estimates, nor would I assume that Apple will come to market with a significantly weaker system. Looking at the processor's specifications, clock rate, and the performance of the POWER4 I'd say that the estimate were entirely reasonable. There were also Moki's comments about some sandbagging going on, and the various rumours of IBM being pleased with how the 970 has turned out.
On the other hand you're right that the new P4 is a beast. The 970 might only be 80% of its performance, but its possible that Apple will beat that number. We just don't have enough information to make a reasonable guess, but even 80% of the new Intel hardware is going to be a lot better than we've got right now!
I'll show where I got the 1% figure (although it is a guesstimate based on relative expenditures, since IBM doesn't break out microprocessor research by chip type) in another post when I have the time. Suffice it to say, the 970 was done very cheaply.
This is a good thing, by the way. If IBM had spent a billion on research, just amortizing the R&D would make the chip cost $500 before even getting to production expenses. If Apple is going to use a few million a year of these, they'd better not have spent more than $100 million on chip R+D. Compare this to the 10 billion that Intel has spent over the last few years (roughly $4 billion/year).
I think you're pretty much right on the money here. IBM is using some advanced circuit design tools to build the POWER series, and the 970 directly leverages the POWER4 work. This has already served IBM very well and it will allow them to continue to compete with Intel into the future on a far more cost effective basis. It may cost them some potential performance, but it might also allow them to focus on high level improvements rather than getting bogged down in detail improvements. If the tools they are using can stay on the bleeding edge it'll allow IBM to take advantage of new techniques & processes faster, but if the tools don't evolve quickly this could turn into a liability. The single biggest advantage in IBM's approach is that they should be able to build massive processors much more quickly than Intel can -- when chips reach into the hundreds of millions of transistors, IBM will be able to do more interesting things with that sooner. Unless, of course, Intel adopts the same strategy.