We have always advocated larger memory capacities over higher memory bandwidth to improve a PC's real-world performance. The effect of high memory bandwidth have always been overhyped by memory vendors, because it makes good business sense. This is the situation we now see with the introduction of the new Intel Core i7 processor. The Core i7 processor supports up to three DDR3 memory channels. Using DDR3-1333 memory, triple-channels would give you a memory bandwidth of 64 GB/s. That is a tremendous amount of memory bandwidth. But do we really, really need that kind of bandwidth? That is what we will be looking at today. Link : The Intel Core i7 Memory Bandwidth
What about the opposite? Very good article. I read somewhere else (cant remember where) that although you may have sufficient bandwidth, you can under-voltage the RAM and other tweaks to get very low latency 5-5-5 or lower? Could you get a stable system doing that? Would there be a read advantage in doing so? Thanks Nikos
Reduce voltage to reduce latency? I don't think that's possible... Maybe if you reduce the clock speed.
benchmark applicability The author indicates the following This would be indicative of an application which is CPU bound. As a consequence, memory bandwidth once it has satisfied the CPU requirement becomes irrelevant to execution time.
Indeed. That's why memory bandwidth is kinda "irrelevant" for the Core i7. There's so much memory bandwidth, you do not have to bother overclocking the RAM or even use high-speed DDR3 memory.
Maybe this will change when Gulftown is launched. It's 6 cores will have higher memory bandwidth requirements to keep them fed. QPI seems more like a preparation for the multi-core era ahead. Still, I feel 3 channels on i7 weren't necessary. Lynnfield's 2 channels should deliver similar memory performance. But, again, you don't get a hexa-core on LGA1156, probably due to bandwidth limitation. Btw, this is my first post here.