Replies to post #3283 on Advanced Micro Devices Inc (AMD)

[yourbankruptcy]

I recall when AMD made it's first Athlon (and kept it faster than the fastest Intel's cpu for the whole year) on 0.18 but used some small parts at 0.13, Intel said that "AMD eats up into its future". Should they say that about themselves now?

[fyodor]

Klaus: As you obviously have profound knowledge on cache matters, maybe you have an opinion on the following as well

My knowledge of CPU architecture and the like is limited to what I read here and there. I have a decent background for understanding most of it, but "expert" knowledge is not something I can claim to posess in this field. There are several on this board who are far more knowledgeble on the subject of cache (and CPU architectures) than I.

I don't have a solid die size for Banias, so I would have to go with the 80mm² or so speculated on this thread. Considering that Opteron's 1MB L2 seems to take up around 81mm² (source: die photo measurement), that would seem very impressive. Recall, however, that Tbred'C' has a die size of 87mm², compared to 101mm² for Barton. Assuming the difference between the two is just the 256kB cache, that results in a cache density of 56mm² per MB. The PIII core was always significantly smaller than the Athlon core, so it's not unreasonable to assume that the cache density of Banias is "only" something like 10-30% higher than Barton's.

-fyo

[Dan3]

Re: Cache density of Banias is really impressive compared to Thoroughbred

They're pretty different cache designs. Intel uses a very simple, but fast cache that helps it do well in "copy every fourth byte" applications like windows media encoder. AMD's cache is complex but slower, which helps Athlon/Opteron do well on SpecJBB and complex SQL queries.

In a nutshell, AMD's cache is less likely to overwrite bytes it shouldn't, but it takes longer to make that better decision.

If the application is a very simple, repetitive one, there isn't much contention for cache blocks, and Intel's fast design is the better approach. On complex applications, with more of the memory space in active use and more active allocated blocks, AMD's solution is the better one.