Memtest86+'s memory performance figures, are they considered vaguely accurate?

[mikeymikec]

Both machines have DDR3-1600 RAM, though mine has tighter timings.

Memtest86+ v5.01 reports my Ph2 960T's memory stats as follows:

L1: 31032MB/sec
L2: 18467MB/sec
L3: 9285MB/sec
Memory: 4747MB/sec

Same version of memtest86+, on a Haswell Pentium G3220:

L1: 199891MB/sec
L2: 49153MB/sec
L3: 37017MB/sec
Memory: 10483MB/sec

I realise a couple of things (-edit - or maybe three):

a) Comparing 2010 hardware to 2014 hardware
b) Raw throughput isn't the whole picture in terms of performance
c) That Pentium can probably give my 960T a good spanking performance-wise in a lot of respects

I just wonder whether people have read reviews that suggest the memory performance of current Intel processors would so drastically outstrip a Ph2, or if the figures aren't really representative of anything, then why should memtest86+ even bother to mention them? I suppose if there was a serious memory throughput performance issue, it would show up because a pass took say two hours longer than expected.

 

[gbeirn]

I can say that by using Memtest86+ for years at work, yes the figures are pretty accurate. Now I don't know if the actual MB/s numbers are reported correctly but they are consistent. Meaning you can rely on them to judge the speeds of two different computers.

This comes from testing the same CPUs and RAM side by side along with tons of variations: 3rd gen to 2nd gen Core, Ph to Ph2, etc.

I don't really pay that much attention to the MB/s numbers since I'm looking for bad memory but when I do every once in a while them seem to be where I'd except them to be.

 

[DrMrLordX]

I have noticed that the numbers do not agree with Sandra, if that matters.

 

[Homeles]

Vaguely accurate? Maybe. Accurate? Doubt it.

The general picture that it's painting in that comparison is correct, from my understanding. To elaborate, Haswell should have 4 times the L1-read bandwidth of K10, twice the L1-write bandwidth, and twice the L2 bandwidth. Here, it's somewhere between 6-7 times higher for (presumably) L1-read. I can't comment on L3 without doing more reading than I'd like to.

System memory bandwidth is something Intel has excelled at for the past few years. However, those numbers are way below their theoretical maximum for DDR3-1600, for both SKUs.

I'm not sure why the Haswell Pentium is reporting 199GB/s for the L1, because its max theoretical L1 bandwidth should be 192GB/s (3.0GHz * 64 bytes/cycle). Pentium models don't have turbo (usually, or ever), and that model certainly shouldn't, so I'm not sure why it'd report a higher number.

Regardless of the gap between the two, keep in mind that Intel only has such monstrous cache bandwidths for 256-bit AVX -- most workloads won't even make use of that kind of bandwidth.

 

[mindbomb]

well, the haswell number is pretty much correct if this is single channel ram. For the amd system, try manually specifying the memory frequency.

 

[24601]

Vaguely accurate? Maybe. Accurate? Doubt it.

The general picture that it's painting in that comparison is correct, from my understanding. To elaborate, Haswell should have 4 times the L1-read bandwidth of K10, twice the L1-write bandwidth, and twice the L2 bandwidth. Here, it's somewhere between 6-7 times higher for (presumably) L1-read. I can't comment on L3 without doing more reading than I'd like to.

System memory bandwidth is something Intel has excelled at for the past few years. However, those numbers are way below their theoretical maximum for DDR3-1600, for both SKUs.

I'm not sure why the Haswell Pentium is reporting 199GB/s for the L1, because its max theoretical L1 bandwidth should be 192GB/s (3.0GHz * 64 bytes/cycle). Pentium models don't have turbo (usually, or ever), and that model certainly shouldn't, so I'm not sure why it'd report a higher number.

Regardless of the gap between the two, keep in mind that Intel only has such monstrous cache bandwidths for 256-bit AVX -- most workloads won't even make use of that kind of bandwidth.

Perhaps the Mobo is overclocking the base clock a tiny bit to cheat benchmarks like most enthusiast mobos are known to do.

Also maybe some of the instructions are getting saved in the re-order buffer giving higher effective bandwidth in that specific benchmark.

 

[mikeymikec]

well, the haswell number is pretty much correct if this is single channel ram. For the amd system, try manually specifying the memory frequency.

Correct, it was single channel. The AMD system has dual. I should have mentioned this already.

Perhaps the Mobo is overclocking the base clock a tiny bit to cheat benchmarks like most enthusiast mobos are known to do.

Also maybe some of the instructions are getting saved in the re-order buffer giving higher effective bandwidth in that specific benchmark.

What would you suggest if I wanted to check to see if the board is doing this? It's a Z87 Gryphon btw.

I have noticed that the numbers do not agree with Sandra, if that matters.

Do they agree with Sandra, relatively speaking, comparing to the 960T / DDR3-1600?

 

[24601]

Correct, it was single channel. The AMD system has dual. I should have mentioned this already.



What would you suggest if I wanted to check to see if the board is doing this? It's a Z87 Gryphon btw.

It doesn't really matter. They usually only do it by setting the base clock to 101-103MHz or something.

It's only enough to get them slightly ahead of the exact same setup on the same setup of a competitor's board.

Do they agree with Sandra, relatively speaking, comparing to the 960T / DDR3-1600?

Yes Intel has always been about twice as fast as AMD per channel comparing the latest to the latest ever since Intel integrated their memory controller.

As I'm going to assume K10 CPUs have a much lower performance compared to Kaveri, the roughly 4x better effective memory bandwidth is within the realm of reason.

 

[Homeles]

Perhaps the Mobo is overclocking the base clock a tiny bit to cheat benchmarks like most enthusiast mobos are known to do.

Also maybe some of the instructions are getting saved in the re-order buffer giving higher effective bandwidth in that specific benchmark.

I've not really heard of memtest being widely used as a benchmark. It's generally a diagnostic tool. Seems silly to cheat on, even if it's a simple "white list."

 

[mikeymikec]

As I'm going to assume K10 CPUs have a much lower performance compared to Kaveri, the roughly 4x better effective memory bandwidth is within the realm of reason.

Right now I'm laughing at myself because this post initially slightly ruffled my feathers, and I started looking for a Kaveri CPU on Anandtech Bench in order to compare to mine, then I thought, "so what if it is?"

 

[DrMrLordX]

Do they agree with Sandra, relatively speaking, comparing to the 960T / DDR3-1600?

Not sure in your case, but in my experience running Memtest86+ on an old Sempron 2800+, an x2-3600+, a Sempron 140, and an Athlon II x4 635, that Sandra consistently produces system memory bandwidth numbers at least twice as high as does Memtest86+. It's like Memtest86+ can't/won't detect dual-channel memory configs on AMD platforms. I wasn't paying attention to cache bandwidth, and I'm not sure that the versions of Sandra that I used back then reported on cache bandwidth accurately anyway. Maybe it did, I just don't recall.

 

[BrightCandle]

My personal experience of tweaking memory settings on the basis of memtest86+ bandwidth figures did translate into similar improvements in Windows and applications. So whether the absolute figure it produces is accurate or not the comparison of the numbers within the program is remarkably consistent and useful for tweaking. I would say the comparison you have is quite representative.