How good is triple-channel memory?
We started by setting up a test system with an Intel DX58SO motherboard along with a Core i7 920 processor, 3x1GB of Qimonda 1066MHz DDR 3 Ram and an Intel X25-M SSD running 32-bit Windows Vista Ultimate. The nature of the Intel motherboard with its four memory slots meant that we could install 3GB of memory in triple-channel mode, 2GB or 3GB in dual-channel, and 2GB in single-channel mode.
SiSoft Sandra Results
Memory Bandwidth in Gigabytes per Second (GB/s)
Longer bars are better
Memory Latency in Nanoseconds (ns)
Shorter bars are better
SiSoft Sandra shows that the memory bandwidth climbs in big steps as each memory channel is added with 7.4GB/s for single channel, 12.7GB/s or 12.8GB/s for dual channel, and 17.1GB/s in triple channel mode. The memory bandwidth didn't scale as steeply as we might have hoped, and increased by some 70 per cent as each extra channel was added. The latency remained steady at 86ns or 87ns.
PCMark Vantage Results
Longer bars are better
We used PCMark Vantage to see what effect these huge changes in memory bandwidth would have across a range of tasks and were only slightly surprised to see that the answer is ‘very little’. The system delivers stacks of performance with 2GB of memory in single-channel mode which barely changed when we switched to 2GB in dual-channel mode. 3GB of dual-channel memory is a surprisingly effective configuration, while triple-channel appears to be a small step backwards.
Don’t worry too much about triple-channel memory as two channels work just as well.
Next page: Memory speed and overclocking Core i7
@Dustin... be sure you comprehend before accusing someone of being ignorant
Ian didn't say Windows "couldn't do PAE - period". He said it couldn't do PAE because too many drivers couldn't handle it, having not been written with PAE in mind.
Yes, I find it truly surprising that an 8-DIMM dual-opteron setup was not tested in this article for Core i7 memory configs!!
This is very worthwhile reporting!
It is little known that the only fully performant memory configuration for dual processor AMD Opterons has been exactly 8 DIMMS of identical density, 4 on each socket, at least according to my tests. Other configurations give poorer measured performance, which may or may not be reported by the BIOS.
I have only tested with an in house tool, Opteron versions up to Barcelona. Anybody concerned about memory performance should repeat the tests on more modern hardware.
I find it amazing that such basic information is not clearly documented and is also rarely tested and reported-...
@jolyon - Yes, it's mainly a driver issue
Here is what Microsoft say on the issue -
And the wikipedia page on Physical Address Extension says -
"However, desktop versions of Windows (Windows XP, Windows Vista) limit physical address space to 4 GB for driver compatibility reasons."
Microsoft themselves confirm that >4GB is a no-go with 32-bit XP and Vista.
So, very limited PAE with Windows on the desktop and deeply scary compatibility issues with PAE on both servers and desktop. We've tried PAE on desktop and server Windows and it quickly became clear that the pain of moving to 64-bit was less than the pain of trying to get PAE stable and effective.
With Linux, we installed a PAE kernels, rebooted, and the servers all worked exactly as before but with much more memory. We are now moving to 64-bit (with virtualisation where required) but it has bought us a few years.
It matches my rule of thumb ...
... which says that more bog-standard memory trumps less but faster memory every time.
Slightly surprised that 3-channel offers no noticeable advantage. Perhaps it's time will come with future iterations and speed steps of Intel's new architecture. Anyway, there's a financial advantage: 12Gb without needing to buy expensive 4Gb DIMMS.
PAE and multicore CPUs means that 8Gb or even 12Gb may be sensible with 32-bit Linux: 2Gb or 3Gb per process, each running flat out in its own core. But if you aren't constrained by some sort of historical relic, 64-bit Linux should be today's default. I doubt I'll be doing many new 32-bit installs in the future.