It seems that everyone getting an Intel 1366 platform (i7 9xx CPUs) has to, and absolutely must, get triple channel DDR3 RAM.  People suggesting builds always get it, and people who decide to only go with dual channel seem to get bonked.

Admittedly, RAM isn’t too expensive (despite prices going up recently) especially compared with the base X58 platofmr, and, as it does support triple channel, one might say “why not”.  I guess this may, however, be a point against choosing the typical i7 930/X58 platform over a generally cheaper i7 860/P55 platform; the main advantage of the X58 being triple channel RAM (and perhaps greater PCIe throughput, if you plan to go dual GPUs, but I don’t see that many people doing that).

So why is triple channel RAM so much better than dual channel?  From what I see, I think people have some idea that triple channel is like “3x speed” whereas dual channel is “2x speed”, so triple channel makes your RAM 50% faster than dual channel.  This is true in a sense, however, this only affects the bandwidth figure.

For transfering data, there are often two primary speed metrics: bandwidth and latency.  It’s the latter that is often ignored, however, more than not, it’s the latter that really matters.  Bandwidth can easily be increased by multiplexing multiple lines of channels (for example, the case with triple channel memory), however, latency is something which is fixed, and generally cannot be improved so easily.

High latencies of RAM have been a primary performance killer for a while (it takes in order of around 600 clock cycles for a CPU to retrieve a value from those DDR2/3 sticks) – that is, ignoring all the insane CPU optimisations (L1/L2/L3 cache, branch prediction etc) that have gone in to try to get around this issue of RAM being incredibly slow.  However, bandwidth has never been particularly much of an issue for most applications either – I’d still say that latency is still of primary importance for performance; various high performance applications, such as x264, explicitly optimise for this by trying to reduce the amount of cache misses.

My 2x2GB DDR2-800 sticks in dual channel can read at about 4.5GB/s if memory serves me correctly.  Very few applications even need anywhere need that speed – or in other words, constantly using a large portion of memory (most of the time, applications use a small portion of memory frequently at a particular point in time).  And even with that read speed, it’s unlikely that your CPU could process so much information, let alone your HDD writing it.

So going back to the issue at hand of whether triple channel is worth it or not.  There are two primary benefits of it over dual channel, one being increase bandwidth (discussed above), the other being greater SMP efficiency.  On this front, it means that the memory can serve multiple data requests simultaneously, rather than queue up the requests.  This may be useful if one is running a multi-threaded application, or several applications, which need to request memory that hasn’t been cached.  However, as alluded to above, most applications use a small portion of memory at a particular point in time, which means that the working memory set is most likely cached when it is used, so the benefit of more channels is perhaps limited, and gives minimal benefit only for certain applications.  However, as the number of CPU cores increase, the advantages of more channels of memory will become apparent, but for the typical quad core i7 930, I don’t think there’s much of a benefit.

Benchmarks also seem to agree with my assessment here as well.
(note, synthetic benchmarks may show a difference as they will often stress the RAM; of course, very few practical applications do this)

Conclusion: triple channel really does not provide as much benefit as many people seem to be thinking.

Disclaimer: I don’t claim to be an expert on this field, rather, I’m basing my reasoning from knowledge I’ve obtained.

Last year, our Uni upgraded all PCs in the IT labs to ones with a Core 2 Duo E8400 CPU and 4GB RAM.  A friend studying at the ANU said that his Uni upgraded all their PCs to E8400/4GB configurations, and noted the serious overkill with the CPU.  This year, my Uni upgraded some PCs in the business faculty’s computer labs to similar configurations.

Some (only a few actually) computers at my work got upgraded to E8400/2GB configurations – again, same overkill CPU but half the RAM.  Recently, it seems that BCC libraries upgraded most of their public computers to an E8400/1GB configuration (they appear to have 80GB HDDs).  Now seriously, this is an overkill CPU compared to RAM given.  Note that these PCs are only being used for web browsing, and maybe some document editing, and they have DeepFreeze installed to prevent viruses and the like slowing the computer down.  They also replaced their catalogue computers with these setups, even though the only task the catalogue computer needs to do display one website.

With the exception of my Uni’s IT labs computers (where you have computer geeks and people doing 3D rendering), the E8400 is definitely overkill for the basic tasks that get done.  But it’s not a cheap CPU either.  In fact, it’s being sold for AU$219 at Umart (which is more than the E8500 going for $210).  At this price, it costs more than an i3 540 ($165), slightly less than an i5 650 ($224) and C2Q Q9400 ($234) and costs more than a Phenom II X4 965 BE ($205; though these companies tend to never go with AMD).  All the while, it’s much more expensive than the cheap $55 Celeron E3300, which should be more than enough to run everything needed for the various tasks.  Okay, these machines probably have been assembled before the i3/i5 range, but even back in those days, it was fairly expensive compared to the rest.

So why so much love for the E8400, even pairing it with 1GB of RAM?  My guess would be Intel’s vPro platform, of which, these machines have a lovely stick advertising the fact.  This platform requires a Core 2 branded processor with Intel-VT, despite few applications (okay, screw Win7’s XPMode; businesses here still uses XP) really using it.  The Core 2 branding requirement is obviously trying to push businesses to buy the more expensive CPUs, as some of Intel’s cheaper “Core 2 based” CPUs (eg Pentium Dual Core E6xxx range and some Celerons) have Intel-VT.

Out of the Core 2 Duos, only the C2D E6xxx and E8xxx range support Intel-VT.  E6xxx is phased out, which leaves the latter, and the cheapest E8xxx is the E8400 (can’t seem to get E8200 and E8300 over here).  So perhaps that’s why everyone’s going with the E8400.

Now considering that these machines came with 1GB of RAM (around $50 maybe?) and 80GB HDD (around $30 probably), a cheap case+PSU, the Intel motherboard and CPU probably consume a huge proportion of the cost.  So is vPro really worth such a huge premium?