The AM2 mem controller

[DrMrLordX]

With Brisbane's rev G1 looking like more and more of a dud, I've been wondering what, if anything could be done to tweak AM2 performance to be more competitive with Core 2 (if anything). Something tells me nothing, especially considering the fact that Core 2 is faster per clock and can hit 4.0+ ghz on air, whereas you're lucky to get a K8 to 3.6 ghz or higher on exotic cooling (3 ghz on air, typically).

One thing that caught my eye, however, was this old review by Anandtech:

http://www.anandtech.com/showdoc.aspx?i=2800&p=5

Interestingly enough, it seems Core 2 approaches a latency floor of about 45-50 ns or 30 ns, depending on whether you believe Sciencemark or Everest. The AM2 A64 mem controller does not seem to have a similar limitation.

If you could theoretically push the memory latency of an AM2 chip to 10 ns or below, how would that affect overall performance? Would it enable an X2 to compete with Core 2 on a clock-per-clock basis in anything?

And just how would you go about reducing latency to that point anyway? It seems to be that latency drops with higher CPU clock (check out this http://www.anandtech.com/showdoc.aspx?i=2741&p=7), higher memory clock, and tighter memory timings. I have no idea how fast everything would have to be running to drop latency below the 10ns mark, though, nor do I know if it'd be worth the trouble.

 

[BitByBit]

It sounds interesting, but memory latency just doesn't seem to have much of an impact on performance these days. There are a few applications in which it can make a notable difference, but sophisticated caching algorithms and large cache sizes mean that 99% of the time, the data/instructions a processor needs are already cached, reducing the impact of memory latency.
The impact of performing such a tweak would probably be similar to using lower latency memory - anything from negligable to a few percent.

 

[zephyrprime]

The ram plays a big part in latency so I don't think you could attain anywhere close to 10ns latency even if the processor itself took zero when performing a memory access.

 

[DrMrLordX]

Originally posted by: zephyrprime
The ram plays a big part in latency so I don't think you could attain anywhere close to 10ns latency even if the processor itself took zero when performing a memory access.

This is true, but you can effectively reduce latency by running the memory at higher speeds (or with tighter timings). What speeds would be necessary for sub-10ns latency, I don't know . . .

Originally posted by: BitByBit
It sounds interesting, but memory latency just doesn't seem to have much of an impact on performance these days. There are a few applications in which it can make a notable difference, but sophisticated caching algorithms and large cache sizes mean that 99% of the time, the data/instructions a processor needs are already cached, reducing the impact of memory latency.
The impact of performing such a tweak would probably be similar to using lower latency memory - anything from negligable to a few percent.

Right, I'm just sort of curious as to which applications would benefit and how much. Anandtech's old review showed how lower latencies did not help gaming performance much(http://www.anandtech.com/showdoc.aspx?i=2741&p=8), though it should be noted that in that old benchmark, running the CPU at a higher speed also did very little in two of the four tests, so it may be that the games tested were largely GPU-limited. I would estimate that any application that benefits from a large l2 cache would also like ultra-low memory latencies coupled with memory bandwidth saturation, but these days, it seems cheaper/easier for AMD and Intel to push large caches on their cores rather than try to get system integrators to support high-performance memory platforms.

 

[Roguestar]

"AM2 memory controller" is a misnomer, Socket AM2 CPUs all have their own memory controller and they're different for each type of CPU . Just being a pedant .

 

[DrMrLordX]

Not so, the memory controller on AM2 CPUs is quite different from the memory controller on s939 CPUs, which is why I mentioned it as such.

That being said, there should be very little difference in the memory controllers of different AM2 CPUs, ranging from lowly Semprons to FX processors.