At least ONE .09 micron A64 cpu with REAL Core Enhancements

[formulav8]

Here is the link to the review that did a very nice job at checking for any performance or temp differences between each A64 core process. The 3200+ .09 he has does really have some improved core enhancments. Some times there is a pretty nice increase in performance increase clock for clock. So it kinda looks like there are both tweaked cores and just shunk .09 process A64 cores around. Kinda wondering. 😕



Jason

 

[ViRGE]

With no offense to the guys that did this article, their numbers are all over the place; the S754 chip is winning half the time it seems, and the rest of time they're near the experimental error point for comparing the two S939 chips. There's nothing conclusive in this article at least showing that 90nm chips are any faster clock for clock.

 

[formulav8]

Not sure what you mean by the 754 winning that much. It barely wins any.

Did you even read the configuration?? The 754 had 1MB cache! and running 22210 memory. The other setups are using 2.5338 memory. So having twice the L2 cache and stronger memory timings speaks very well as to why the 754 cpu won a couple here and there. And it still doesn't win near half the benches as you say.

He did a fine article. Especially with the 3500+ and 3200+ at the same exact speeds and ht speeds. There is absolutely no software changes that take effect. So it would not at all change the results of the benches like they show if there was no core enhancements.


Jason

 

[ViRGE]

I didn't see that it was a 1MB chip, but reguardless, there's no conclusive proof. The PCMark scores differed by 50 points, 3DMark by 600(out of 22,000); his game benches show some interesting data, but at the same time, why is the 3400+ beating the FX53 in UT2004? There isn't enough data at this point to claim that the 90nm chip is faster, especially when some of the other reviews are showing no change.

 

[ludditeTx]

It is easy to get a couple of % variation, for example by loading the app into a slightly different context and hence having slight misalignment on code in memory.

I would want a better understanding of the test methodology, and I couldn't find any description of that. For example, are the benchmarks run N times, with the highest and lowest values discarded, a mean of the remaining values derived ...?

The numbers may be noise, or based on a reproducable methodology, I just can't tell.
AMD do admit that they rework part of the design every 6 months with a goal of improve performance (upto 10%), and it may simply be that the other 939 is from a previous design. All current 939's may be identical designs and performance.