IMC

Nehalem's IMC supports up to three DDR3 channels, with two DIMMs per channel up to a total of 24GB of RAM. 800MHz, 1,066MHz and 1,333MHz are supported by default, with faster RAM support possible in the future.

Although Nehalem's memory controller doesn't support RAM as fast as Intel's own X48 or nVidia's 790i chipsets did (up as high as 1,600MHz to 2GHz respectively), because the controller is on-die rather than in the chipset it doesn't need to.

In theory a 64-bit (8-byte) dual-channel DDR3 memory controller with 2GHz DDR3 could offer about 31.2GB/s (8 bytes times two channels times 2GHz divided by 1024 to get from MB/s to GB/s). However, even the fastest Penryn system only has a 1,600MHz FSB and that only offers at most 12.8GB/s of bandwidth not all of which can be used for accessing RAM, so in the real world the limitation isn't memory speed.

Nehalem's 64-bit memory, triple-channel, controller, by contrast, can actually offer that same 31.2GB/s figure, but with slower 1,333MHz RAM. That theoretical bandwidth doesn't actually translate into real-world figures, but the basic point stands: Nehalem has much more memory bandwidth available than Penryn did.

That's not to mention Nehalem's added benefit insomuch as slower-clocked RAM is also generally lower latency, making data access faster yet again. All told, Nehalem has all the bandwidth a system could realistically need and then some, while also reducing latency over previous architectures.

Interestingly, it seems that memory voltage is tied directly to the CPU, as there is a 1.65V limit imposed over which, it is warned, damage to the processor could occur. That could pose a problem as some RAM kits available now are rated as high as 1.9V or even 2.1V. You have been warned!

The final consideration of Intel's IMC is that in multi-socket systems, each CPU will have its own dedicated RAM pool and if a CPU needs data in another's memory allocation, it will have to pull it over its QPI link with that CPU. QPI's fast speed and large bandwidth should mean that the latency introduced as a result of this is minimal, though - and it's not going to be of concern to normal consumers anyway.

comments powered by Disqus