Any disadvantage to an Integrated Memory Controller?

[thilanliyan]

So are there any disadvantages to an integrated memory controller like what's on the Athlon's??

Just curious by the way.

 

[MrPickins]

Increased die size, and cpu (not total system) power draw?
Just my guess. But, neither of those affects performance.

 

[Aluvus]

If you want to support a new kind of RAM, such as DDR2, it can be a bit inconvenient. Hence Socket AM2.

 

[lopri]

You're basically tied to the memory technology of which the IMC is based on. So up until AM2's introduction the memory choice for A64 was just DDR. On the Intel side, a socket 775 P4 can be run on 945 chipset (DDR) or 975 chipset (DDR2). Well, there even is a Conroe board with DDR support.

 

[HopJokey]

Intel had a Integrated Memory controller planned for a processor earlier in the decade. However, the memory controller was tied to RAMBUS memory, so the part was canceled.

So basically, the disadvantage is that with an IMC you are tied to a specific type of memory technology (ofcourse you can ofload memory stuff to the chipset, but that negates the performance advantage of an IMC).

Other than that and the extra transistors that MrPickins mentioned, probably not too much disadvantage. The performance advantage outweighs the disadvantages. Memory technology has been pretty stable over the last few years as well.

 

[uOpt]

As mentioned, the main disadvantage is that you cannot pick your kind of memory after you make the die. The die already hardcodes whether you use DDR or DDR2, whether you use unregistered/registered or fully buffered DIMMs. Otherwise, the northbridge would.

That is a problem on multiple levels. First of all you can't get away with just picking one. As AMD has involuntarily demonstrated, you have to have at least registered and unregistered as unregistered can't hold the amounts of RAM servers demand and registered is too expensive for consumers, not to mention the overclockers come down your throat. But when you have an integrated memory controller that means you have to make different dies, which means more engineering/design, then more production lines, more stockpiles and less flexibility in designating the final CPUs after binning.

There is also the risk of killing the memory controller by screwing up something in the board or RAM, in which case you kill the CPU if it is integrated.

In the end, however, the integrated memory controller for AMD has proven not only to be a huge advantage at the time of it's design, it continues to be one of the major advantages left to AMD after it has to puts its AMD64 line against an entirely new Intel technology. So I'd say it was worth the trouble for AMD several times over.

 

[thilanliyan]

But Intel proved (at least for now) that it's not needed with the Core 2 launch right??

I seem to recall that some Intel engineer(s) had said an IMC was a very good idea. And it will eventually be used in their CPUs...but what is it that's coming down the line that would benefit greatly from an IMC??

Would an Intel quad-core benefit greatly from the IMC??

 

[coldpower27]

Originally posted by: thilan29
But Intel proved (at least for now) that it's not needed with the Core 2 launch right??

I seem to recall that some Intel engineer(s) had said an IMC was a very good idea. And it will eventually be used in their CPUs...but what is it that's coming down the line that would benefit greatly from an IMC??

Would an Intel quad-core benefit greatly from the IMC??

They proved there are multiple avenues to attaining performance. IMC is one of them but not the only way in the universe to do so.

 

[MrPickins]

You may be locked in once the die is set, but isn't K8L supposed to support ddr2 as well as ddr3? Once again, I'm sure that adds even more transistors, but at least there is the possibility of building in a little flexibility in an IMC, and could keep it from becoming obsolete before the architecture of the rest of the die.

But, building for the next generation can be dangerous, i.e. Rambus.

 

[imported_Crusader]

Originally posted by: coldpower27

Originally posted by: thilan29
But Intel proved (at least for now) that it's not needed with the Core 2 launch right??

I seem to recall that some Intel engineer(s) had said an IMC was a very good idea. And it will eventually be used in their CPUs...but what is it that's coming down the line that would benefit greatly from an IMC??

Would an Intel quad-core benefit greatly from the IMC??

They proved there are multiple avenues to attaining performance. IMC is one of them but not the only way in the universe to do so.

They've stated in interviews that they (intel) would also be moving to an on-die mem controller as well.. as they finally admitted the performance advantages of one.