MacPro Memory Question

[Loa]

And I don't honestly think you can say that article underscores any of the important points in this particular discussion.
.

I quoted 2 articles in there, one of them focusing on triple or dual channel when using DDR3 chips.

Also, I'm waiting on the name (or details) of a test you will accept as good. I'm ready to do RetouchArtists test: is it still current for testing in 2009?

I have all afternoon to do these tests, but I don't want to start one you'll dismiss out of hand.

Loa

 

[Tesselator]

True, but the net effect of interleaving memory across multiple channels is an increase in memory bandwidth... which can also be achieved by increasing the speed, or reducing latency.

Using your own analogy, it's the same as improving storage performance... you can improve HD performance by RAIDing more disks, decreasing access times, or improving the transfer rate of individual disks. The overall effect is improved STR in any case (and likely a proportionally small impact on overall system or application performance).

For memory, the fact is (due to the two points I mentioned above) increasing memory bandwidth beyond about 10-12GB/s does not materially impact application performance significantly... small amounts for very specialized apps... yes, but it's still more important to have enough memory, than faster memory... else you end up relying on orders of magnitude slower storage.

Yeah, I can agree with that. But once you get past the amount needed to typically avoid system wide VM hits and etc. the triple-channel should make a noticeable difference. At least I have measured the differences between single and dual channel - and I occasionally notice it as i use my system too - enough to recommend trying to attain both enough RAM and also configuring it in triple channel orientation. Or dual channel if that's all you have.

Of course we can take the logic here to the extreme in order to show the point. Imagine your system with only 1MB of RAM. We'd be at the mercy of our storage devices. Of course we need "enough" RAM. That part should be obvious. After that the tuning (speed, CL, channel orientation) is also important. To say it's not and that it "has zero impact" is a mistake and not correct.

So that's my caveat anyway.

 

[Loa]

To say it's not and that it "has zero impact" is a mistake and not correct.

So help me prove you're right then. Because right now all you have are theoretical arguments and tests you made on an older machine with different architecture.

For the third time, give me a test I can run right now and I'll do it and post the results.

Loa

 

[Tesselator]

I have all afternoon to do these tests, but I don't want to start one you'll dismiss out of hand.

Loa

Retouch wouldn't be it. Maybe a section of it.

But just because I know that kicking a monkey in the butt and counting rocks till he hits the ground is not a good model for keeping time and doesn't apply to something you would hang on your wall or wear on your wrist, doesn't mean I'm a master watchmaker.

If you're set on using PS to bench this I would suggest creating some scripts yourself. As one idea I would maybe try creating a few very large images (16-bit, blank white, blank black, gradient fill) and then duplicating and deleting that layer repeatedly a few hundred times. I dunno, creating a proper test takes a lot of work, research and testing. There's almost nothing intelligent on the Mac in OSX that I know of. All bench marking apps (what few there are) are pretty lame if not just plain broken. They seem all to be too high level and we end up testing the efficiency of the author's code more than the performance of the hardware. My benchmarking suit is attached below if you're interested in what I have and haven't tried. I guess you can know by icon images only.

Install both dual and triple channel setups and you WILL notice some differences. When you do try to isolate what they are and why they are happening and then create a test loop for it.


.

 

[Loa]

I do want to test this in PS as it is just about the only "demanding" app I regularely use.

As one idea I would maybe try is creating a few very large images (16-bit, blank white, blank black, gradient fill) and then duplicating and deleting that layer repeatedly a few hundred times.

I'll do that.

I dunno, creating a proper test takes a lot of work, research and testing.

I agree, if you're trying to make a universal test. What I liked about RetouchArtists test is that it's a series of commands that I use often. (You know, real-world...)

Install both dual and triple channel setups and you WILL notice some differences.

We'll see. Only argument I'm expecting is that I will not be able to have the same amount of RAM in my machine for both tests. Triple channel will be 6GB, and Dual channel will be 8GB.

But while you may object to that "flaw" in my test, it completely fits in our specific debate: it's for our own 2009MP machines. Not Windows like in the graphs I provided, nor a 2006-2008 MP. It also fits the "real-world" aspect of the test I want to make.

So I'll test the "artificial" blanks and gradiants images you suggested, as well as use one of my own images with RetouchArtists's test. I'll then report the findings (maybe tonight, if not; tomorrow).

Loa

 

[Tesselator]

We'll see. Only argument I'm expecting is that I will not be able to have the same amount of RAM in my machine for both tests. Triple channel will be 6GB, and Dual channel will be 8GB.

According to that article you linked to it won't matter as long as there is free RAM during all tests. LOL

It does actually matter but I dunno how much tho. Probably not measurable.

Just run any tests you do many many times and average the results while tracking best and worst.

EDIT:
This is actually close to what I would expect: https://forums.macrumors.com/posts/7984940/ = 8.6% increase. Also VirtualRain mentioned 7% increased render speed. That sounds about right too. I would expect 7 to 12 percent increase where the operations are dependent on RAM I/O. I see a little better than that when looking at the differences between single and dual channel modes on my machine - and everyone told me the same things as I'm hearing now - "it won't matter, you can't tell", etc. 7% increase? Hells ya! I'll take it!


.

 

[VirtualRain]

Yeah, who doesn't want a 5-10% increase in performance... that's worth a lot.

The problem is, some folks figure that going from 2 to 3 channels (a 50% increase in memory bandwidth) or from DDR2-800 to DDR3-1066 will yield a 50% increase in performance... it just wont' happen.

So Tess... how much is a 10% increase worth to you... enough to pawn your 2008 MP and get a Nehalem?!

 

[Tesselator]

Yeah, who doesn't want a 5-10% increase in performance... that's worth a lot.

The problem is, some folks figure that going from 2 to 3 channels (a 50% increase in memory bandwidth) or from DDR2-800 to DDR3-1066 will yield a 50% increase in performance... it just wont' happen.

So Tess... how much is a 10% increase worth to you... enough to pawn your 2008 MP and get a Nehalem?!

No.

I couldn't bring myself to buy a 2009 system. If someone gave me one or traded me for this one it would be worth it to buy 24GB or ram instead of 16 or 12GB instead of 8 tho!!!


And yeah.. for sure I guess some people get their expectations up too high. Even 10%... that's NOT a 10% system wide increase. It's only bandwidth to/from the RAM and the CPU cache (mostly).

 

[Loa]

Hello,

Tests results are in. I started a new thread here to make room for the complaints! hahaha

https://forums.macrumors.com/showthread.php?p=7986352

Loa