Replies to post #4826 on AMD strictly moderated (AMD)

[mas]

On multithreaded benches Conroe's apparent lead shrinks accordingly. It's not just me claiming it's the memory controller it's AMD and whole world vs you, the amatuer sophistrist. I will go to the trouble of digging out an even more specific quote as you are too intransigent to connect the obvious dots.

K8's improvements were in the SSE engine (which do not tend to be used in cache bound synthetic benchmarks), the front end (which would require more diverse traffic and instruction mix than one designed to be completely cache bound), and some aspects of the execution pipeline (which might not be affected by the type of computation used in the synthetic test in question).

Firstly the benches weren't synthetic and your characterisation that the K8's uarch changes can't be used on cache bound work is just yet another example of your biased ignorance being paraded as knowledge.

[mmoy]

> That would not entirely surprise me, depending on the
> benchmark. K8's improvements were in the SSE engine (which do
> not tend to be used in cache bound synthetic benchmarks),

The vast majority of Windows applications do not use SIMD.
Some benchmarks do but in typical user applications, users
would not see benefits due to SIMD. I think that most will
admit that older AMD chips had really, really lousy SSE(1)
support and that you were probably better off not using it.

I'm not sure if you're including the addition of SSE2 support
as an improvement and it certainly would be for applications
that take advantage of parallel instructions. It's not clear
to me that using SIMD scalar instructions beats f87.

> the front end (which would require more diverse traffic
> and instruction mix than one designed to be completely
> cache bound), and some aspects of the execution pipeline
> (which might not be affected by the type of computation
> used in the synthetic test in question). Why not name the
> benchmarks and speak to its strengths like an intelligent
> person, rather than closing your eyes and pointing randomly?

I pointed out the front end in another post and this seems
to be related to the pipeline as it was given as one of the
reasons for increasing the pipeline in an article that I read.

How much time does this save on average though? It wouldn't
seem like much to me.