v5.1.0 reports quad-channel bandwidth below 20 GB/s?

[jwagner]

What kind of reported main memory bandwidths do you get on your quad-channel systems... ?


Here I'm seeing memtest86 (v5.1.0) bandwidth of only ~17 GB/s on a 3.9 GHz i7-4820K, Intel X79 chipset, UEFI, speedstep is off, memory is four 4GB DDR3 DIMMs (16GB total) with identical 9-9-9-24-2N timings running at 1600 MHz. The motherboard has 8 DIMM slots. With single DIMMs in each memory channel (quad-channel mode) the memtest86 bandwidth is ~17 GB/s. Isn't that very slow? I also tested putting two DIMMs in just two channels (dual-channel mode) and the reported bandwidth remained about ~17 GB/s. I was expecting it to be half...


Is it possible that memtest86 v5.1.0 UEFI can generally not reach over ~20 GB/s on Core i7?

[David (PassMark)]

MemTest86 V5 reports multiple benchmark numbers (up to 4).
If your figure of 17GB/s is for the main RAM, then this is OK.

There is no standard way to measure bandwidth. You can get a lot of different results depending on the nature of the test. For example,
- Does the test read and write, or just one direction?
- Is the access linear, or somewhat random?
- How much RAM is used for the test and what % is cached?
- How many CPU cores are used?
- 32bit or 64bit access (of SIMD)?
- How the test algorithm is coded and how much CPU load is introduced? In high level languages (Java, C++, Scripting, etc..) often the RAM is fast enough to keep up with the CPU. So the CPU is the blocking some of the time. You need to use hand coded assembler to really push the RAM, and keep the CPU usage low, but often this results in a rather artificial situation.

Switching from dual channel to quad, never doubles bandwidth. I don't have a quad system to test with, but I would be surprised if it was more than a couple of % better on average.

There are also articles like this saying Quad is sometimes slower,
http://www.tweaktown.com/articles/44...sis/index.html
"...it's just this personal issue I'm having at the moment with the numbers we're getting out of Quad Channel memory in our normal benchmarks. It doesn't make sense that from a bandwidth perspective, it offers no more performance than Dual Channel, if not slightly less at times."

[jwagner]

Thanks for the info! That makes sense.


So if memtest86 throughput is CPU-limited on this sytem, then I'm happy with the 17 GB/s main RAM result (and the error-free run in now 24h).


Interesting article you pointed out. Additional googling turned up a newer article from 2013 with AIDA64 v3.00.25 (the 2011 article had v2.00.17): http://www.guru3d.com/articles_pages...review,13.html While it doesn't compare quad- vs. dual-channel results of one and the same system, the page does list i7-4820K quad-channel bandwidth results that are about twice better than those of e.g. a i7-3820 dual-channel setup. I suppose those tests are simple sequential read, write, and copy, likely without any arithmetics or data comparisons -- makes sense that memtest86 actually does useful extra work rather than just shuffling around data, and thus needs more CPU.