This is a topic that has intrigued me for a long time. Most synthetic memory benchmarks (Maxxmem, Aida...) show the 2600K/2500K and the SB-E lineup have identical memory bandwidth. Actually SB's dual channel has a lot more memory bandwidth than Bloomfield's triple channel due to it's improved memory controller, or the newer architecture in general, and I've never seen a single test, nor synthetic or real life application where the memory channels made a difference at all.

 

Many of those apps/benchmarks used to compare CPU performance are heavily threaded also, and still the 2600K performs just the same

Even if benchmark/whatever is CPU intensive, it doesn't mean that it is memory intensive. In some benchmark cases, it is totally possible that benchmark is running almost solely in CPU cache, and you don't get that much memory reads/writes to show the difference. For CPU test, they even try to minimize the memory's effect on measurements.

 

Every benchmark is also different. For this reason AIDA memory benchmark gives pretty much same result for SB and SB-E processors and SANDRA benchmarks show pretty huge difference (somewhere around 21GB/s vs 37GB/s). Why so? I think, and this is just my opinion, AIDA doesn't measure properly aggragated memory bandwidth where your program(s) are executing several simultaneous reads, writes and copys from/to/in memory. SANDRAs algorithms do this and it shows the performance difference. 39XX "slowness" in AIDA memory test is not wrong though. It tells the performance in those instructions it makes and shows how the larger latency of SBE memory controller affects to the measured bandwidth.

 

So, no synthetic performance test will tell you everything like no single "real" program will give you the right answer also. In that sense benchmarks and applications are the same and one's performance can't tell the truth how some other bench/app will perform on a certain system. It all depends what program is doing and what is exactly measured. Fact is, memory bandwidth of SBE is there, it is purely a mathematical question. Wholly another thing is, if your system will ever utilize it to truly benefit from it.

Even if benchmark/whatever is CPU intensive, it doesn't mean that it is memory intensive. In some benchmark cases, it is totally possible that benchmark is running almost solely in CPU cache, and you don't get that much memory reads/writes to show the difference. For CPU test, they even try to minimize the memory's effect on measurements.

 

Every benchmark is also different. For this reason AIDA memory benchmark gives pretty much same result for SB and SB-E processors and SANDRA benchmarks show pretty huge difference (somewhere around 21GB/s vs 37GB/s). Why so? I think, and this is just my opinion, AIDA doesn't measure properly aggragated memory bandwidth where your program(s) are executing several simultaneous reads, writes and copys from/to/in memory. SANDRAs algorithms do this and it shows the performance difference. 39XX "slowness" in AIDA memory test is not wrong though. It tells the performance in those instructions it makes and shows how the larger latency of SBE memory controller affects to the measured bandwidth.

 

So, no synthetic performance test will tell you everything like no single "real" program will give you the right answer also. In that sense benchmarks and applications are the same and one's performance can't tell the truth how some other bench/app will perform on a certain system. It all depends what program is doing and what is exactly measured. Fact is, memory bandwidth of SBE is there, it is purely a mathematical question. Wholly another thing is, if your system will ever utilize it to truly benefit from it.

 

True, Sandra is the only memory benchmark where memory channel affect the score. The bandwidth seems to be more of a theoretical value because it's not so much the type of load but the type of memory load, since other memory intensive apps and benchmarks (most of them out there) don't benefit from it at all. I'm a software engineer, slower server RAM and the channel count have never been a limitation in any of the apps I've worked with, which of course doesn't prove anything, I'm just curious what RW applications are those where memory channels help, because I've never seen one

True, Sandra is the only memory benchmark where memory channel affect the score. The bandwidth seems to be more of a theoretical value because it's not so much the type of load but the type of memory load, since other memory intensive apps and benchmarks (most of them out there) don't benefit from it at all. I'm a software engineer, slower server RAM and the channel count have never been a limitation in any of the apps I've worked with, which of course doesn't prove anything, I'm just curious what RW applications are those where memory channels help, because I've never seen one

Interesting conversation going on here, though little offtopic. I'm not a software engineer (I'm more like fiddling with comms and networks as a profession), but the way I see things is like this: Practically every threded application that uses a lot of memory benefits from the bandwidth as that usually means lots of simultaneous reads, writes and copys. On the otherhand single applications which only uses 1 or 2 threads don't usually use more than one or two channels of memory thus using always less than 50% of the maximum theoretical bandwidth of four channels in X79 case, usually a lot less. That's why we see so little games which can benefit from huge bandwidth. So, multiple programs or heavily threaded software using a lot of memory is the application what we are looking for. It's certainly not FSX or any other 32-bit game with 1-4 simultaneous threads using up to 4GB of RAM: no bottleneck there. You'll manage fine with even 1 channel as your CPU doesn't acces memory mostly but with one thread at a time.

 

And then of course server applications, even if you haven't found BW so useful. SBEs have their roots in Sandy Bridge EP -processors (=Sandy Xeons) which came out late last year, and many server applications have great need for high memory bandwidth. There may be several multithreaded CPUs constantly accessing same memory addres space size of 384GB or so, something that seems totally ridicilous from the desktop user point. Also in servers, where you have large amounts of RAM in gigabytes and in physical sticks, you have to seek the bandwidth by other means than just rising memory chip speed. This is because first, your memory controller can't handle large amounts of high speed RAM and two, high speed usually means higher latencies. We also have to remember, that servers usually use either ECC or registered memory or both, and those technologies tend to lower memory performance.