Should we consider RAM content?

[Psi*]

Trying to stick to "RAM content", there were memory tests that intensely tested the memory system with tests like walk through 1s with all of memory set to 0 and then looks for bit errors ... and many variations of this ad nauseam ... and can take a very long time. The programmers had some knowledge of the hardware and PCB layout as the testing could find crosstalk issues in the physical layout of the memory system. Crosstalk could be the coupling of a signal from 1 trace on a PCB to another causing some kind of memory error. BUT crosstalk could occur because of just a bad board layout near some sockets.

Certainly the tests could find flaky or marginal chips as well. And, as I vaguely remember could identify particular chips so you knew what to replace.

So there are 3 topics relative to memory testing; the memory chips, the sticks or modules, and the mother board.

Now that we have gotten past my own self interests about my build, I remember that this kind of memory testing is what 1st came to mind. Clearly an exhaustive RAM test of this nature would take a long time given GBs of RAM, but very fast processors ... so no telling. It looks like there might be some programs available for today's systems.

Testing might even need to be focused on the mother boards and keep the same set of chips ... the "gold" set(s) if you will. Or the other way, have "gold" performance mother board(s), and change the chips. Either way, probably some obvious economy in there some where driven by $$ & time. And, I would be happy with some kind of gray scale ... good luck figuring that out.

And, who knows, you might get half way into this & truly prove that memory bandwidth makes for nice testimony and that is about all. Or, maybe there are some mobos that might have ample OC features in BIOS, but really do suck (not to get too scientific or anything) at actual capability because of hardware issues.

Last, about *my* stuff, all of my programs have scaled directly with CPU speed ... so my assumption is that CPU/ALU is the limiting "channel". Although not all of the programs peg the CPUs. But, not all number crunchers are created equal or are based on the same solving methods. Some programs might use 80% of the CPUs (or cores) which does cause one to wonder what is the neck of that funnel? Memory channel, inefficient programming, or .... ... So for me, I guess fast memory may be sort of like religion for some people; they "go" just in case.

Until I find out otherwise (from you???) I am going to buy the cheapest CAS 6 or CAS 7 1600MHz RAM, thanks Kougar, mediocre memory that I can find. I will keep 1866 CAS7 for the next build hopefully later in the year if the economy allows.

[Rob Williams]

Whoops. I did kind of get off the memory topic there.

I have to be honest here... the tests you describe are severe overkill for our goals. While the results could prove interesting, testing for things like crosstalk sounds more like something that should be tested in a proper lab, with equipment we certainly don't have. I really don't think crosstalk is a huge issue, and if it is, I'd like to be corrected. Memory is more stable than ever, and even when overclocking I'm not sure issues like this would be introduced.

I've tested many stock-speed kits before and have never received errors even after long tests (8 hours). The tests you describe would no doubt take days, and in all seriousness, it would make no sense for us to dedicate so much time to something very, very people would be interested in (we do have limited PCs we can work with).

The most important factor we'd want to take a look at is simply where the bandwidth can be put to good use, and what real benefit can be caused from tighter latencies. Some companies sell DDR3-1333 7-7-7, while others sell DDR3-2000 7-7-7... it'd be our goal to find out exactly why that faster memory is worth looking into.

Regarding your personal scenario, are you using common applications, and is there any way to possibly "benchmark" it with the same workloads over and over?

Originally Posted by Psi*

Last, about *my* stuff, all of my programs have scaled directly with CPU speed ... so my assumption is that CPU/ALU is the limiting "channel".

Sadly, I think that goes for most any application. That's the problem :-/

[Psi*]

I wasn't going to add anymore to this thread, bu-u-ut I cannot resist.

I have finally just run Everest on my Supermico H8DC8 e/w dual Opteron 290s. It is fascinating that memory latency is between 59 ns & 60ns. This is with the cheapest possible RAM I could get at the time .... Patriot DDR, ECC, registered, unbuffered, 166 MHz, 2.5, 6, 6, 7.

Needless to point out that with a server board there are very limited OC capabilities. I did just try Nvidia's ntune program ... what can I say? I had time on my hands & my mind was wandering plus I just waited a few days (128 hours) for a model to run.

To answer your question about common applications ... "well, not really". They are commercial apps. but quite expensive @ > $60K. And, the DVDs don't even play music video.

I am looking into some freeware number crunchers that may be similar. Oft times there is a university behind such software & they could be multi-threaded. So a benchmark could be setup running standard models with these programs to compare various h/w setups.

Good benchmarks with these programs would be contingent on problem size & type. It would be great if I could devise just 2 test models; one memory pig but fast simple solution (meaning more scalable with CPU speed) and, the other might have a lot of memory access overhead (meaning a more complex model). More about that on another time.

Back to everest ... FPU Julia is 4760 on all 4 cores on H8DC8. So you can see how I am very interested in more CPU speed.

And, back to Nvidia's ntune ... I blindly jumped in fully committed & set the program up for auto-tweaking, or what ever it is called. Came back a while later & the machine was BSOD-ed. I wasn't really surprised. In attempting to make slight adjustments, manually, I became quite adept in shorting the BIOS reset pads. I could muster an ~9% over clock so my 128 hour run became 123 hour. ehhhh! Not worth the gamble for stability. There is no option of voltage adjust and any tweaking of the RAM was immediate BSOD.

BTW memory bandwidth is 8.40 GB/s per Sandra.

What does this all mean? I don't really know maybe you guys can tell me. I think that it means I need faster CPU like a i7 clocked as high as possible.

[Rob Williams]

Originally Posted by Psi*

To answer your question about common applications ... "well, not really". They are commercial apps. but quite expensive @ > $60K.

How on earth does a piece of software cost $60,000? I could see a site license being up there, but a single application?! And here I thought 3ds Max was expensive! Given it's a commercial application, can you share what it is? I'd love to better-understand what could require so much memory and benefit from both huge bandwidth and low latencies, because I've yet to see a great example.

Originally Posted by Psi*

Not worth the gamble for stability.

No sir... not at all. That's what I was hinting at earlier. Imagine running a render for four days and then it BSOD hours before the render was to finish. I can't even picture how aggravating that'd be.

Obviously a faster CPU is going to benefit you, especially if the scenario is heavily multi-threaded. In our Core i7 tests, the HyperThreading feature benefited a LOT of our workstation apps, especially 3D Rendering and to a degree, also our video rendering. I just posted a thread the other day where I'm looking for suggestions and the like for upgrading our CPU testing machine... maybe you could have a look? It's not memory-related per se, but I'd love to hear some recommendations from someone who actually works with such hardcore scenarios day in and day out.

Not to go too far off the intended path here, I wish it were a simple task for me to include both consumer and true workstation benchmarks, but it's a matter of acquiring said applications and also scenarios, because what we have now is a little bit lacking. We're fortunate to have a copy of 3ds Max, but even the model we uses there is lackluster... it's included on the samples disc! I wish there were a resource of free but fully robust projects out there for us to benchmark with... I think our results would greater help the professional if we were able to get some.

</tangent>

Back to your problem though, like I said, it's REALLY hard to say whether faster memory is going to improve much, but at this point it seems like the CPU is the biggest roadblock. If you have the money, though, I'd recommend just buying what you can... a dual-socket Xeon (Nehalem) machine with 12GB of DDR3-1600 RAM would simply be stellar for what your doing (again, as long as your scenarios are truly multi-threaded, and for $60K, they freaking better be).

One a dual-socket motherboard, you're likely to be able to run even more than 12GB, but I'd start off with 12GB (6GB to each CPU essentially) and see if more is needed later on.

[Psi*]

haha ... well there are a few packages that I know the price with certainty.
www.ansoft.com www.cst.com www.remcom.com to name 3 3d electromagnetic field solvers. Their primary programs can predict with good accuracy everything from how antennas for communication with satellites & antennas embedded in notebooks & antennas in cell phones next to the human will work to creating a video of an electric charge moving through a PCB & connector.

And, this is just the electrical side of things. www.ansys.com www.flomerics.com have software for calculating mechanical dynamics to "computational fluid dynamics" ... this is how Apple, Dell, HP & the rest figure out how many vents & what kind of fans go into their enclosures. They also predict how molten plastic flows in a mold for computer housing to connector housings.

Soooo, all of that geeky stuff requires a lot of geeky people around the planet writing multi-threaded software (a $10K option from some & almost old hat now a days), distributed network processing (another $10K option from some), to ... make sure you read about MPI clusters that most of these companies have implemented ... no idea how much that costs!

When business was really good, I had 3 of the field solver with several options. "Just tools in the tool box" as they all have their better features. One has a 3D repositionable human body ... necessary for figuring out how/where to put antennas with best range amongst other things.

The SPEC software touches on one of these solver techniques, but those 1st 3 companies actually represent 4 or 5 solution methods AND each use the CPU & memory quite differently. I am happy with just the multi-threaded stuff.

Yes, I have been drooling over the a dual w5580 system that I might build for ~$5K. But not all and even not many projects require that intensity ... I have learned. I believe that the I7 975 is equivalent to "a" w5580. The I7 975 is about $700 less (@ Newegg) and 4 cores are a great thing and not all problems will use all 4 cores. Some, because of smaller size, only setup maybe 2 threads & half the I7 would be used. That would be a big waste with dual w5580s.

Yes, I build my machines rather than pay the extra to Dell. One must be diligent and quick about getting it all together and built ASAP in case there are issues. In that case ... gotta love newegg! The $$ that doens't go into this build (not building dual w5580 system) will go into the next system or software.

I don't know what I could do to help, but I would at least offer my $.02 worth. I get into all of the current PC technologies when I need to upgrade. I am there now, unfortunately the economy has slowed that down but it does give me more time to study it all for the best bang. I will try to keep on the freeware number crunchers to come up with suitable problems.

I think since I am a system builder, it is why I OC. I have an old dual Athlon water cooled system that has been OC-ed for .... well for ever. For the past 3 years, since the dual Opteron box, it has been the best darn email & report writing machine ever! Still running Win2K, I'm too cheap to upgrade it and I am afraid. But, I not afraid of nor unfamiliar with OC-ing.

That's why I have been polluting this thread with questions & comments quite off the original topic about the I7 975 mix with what kind of memory that might have the most likely combination of maximum CPU clock speed. Well, I thought it to be a little related & if the forum were more heavily traveled like some of the engineering or auto forums I would have been more careful, because I sure would have been flamed long before now.

[Rob Williams]

Thanks for the URLs and information regarding those applications... some of what those can do is simply amazing. This is one application to catch my eye... seriously cool stuff. I think I'll stick to SPEC, hehe... since the results we'll get from that are still going to be quite good. I do have to wonder just how important PC speed is to people using those applications mentioned above... I have my doubts that anyone benchmarks them.

I agree on the i7-975 vs. 2x W5580. It's too bad it's not easy for you to simply find out if that much processing power could be utilized. If you're rendering a single project, it's doubtful that it will use all 8 cores / 16 threads, but again, it's really hard to say. You'd have to ask people who would know, I guess... including the company itself.

Originally Posted by Psi*

I think since I am a system builder, it is why I OC. I have an old dual Athlon water cooled system that has been OC-ed for .... well for ever. For the past 3 years, since the dual Opteron box, it has been the best darn email & report writing machine ever!

AMD chips from that era were amazing for overclocking. I had a single-core AMD 3200+ at 2.0GHz and had it overclocked to 2.7GHz for three years straight without an issue. I only got rid of it because I upgraded... I was long overdue for a multi-core chip, hehe.

As for the memory issue, I still don't think it's anything you really need to fuss over. A fast kit of memory isn't that expensive anymore, and I truly believe that you're never going to hit the bandwidth that DDR3-1600 7-7-7 would offer. Especially on the Core i7 platform, where the memory bandwidth is between 3x and 4x the previous Intel architecture, and 2x AMD.

http://techgage.com/reviews/intel/i7_975/13.png

[Kougar]

Psi, since you were interested in building a Core i7 rig, were there any bench data that you'd specifically wish to see? I know how Opteron's used to be king, but 60ns latency should not be that surprising.

Using Everest and the settings in my sig, I get a consistent 43.8ns latency, FPU Julia of 17863, FPU Mandel of 9552, and FPU SinJulia of 7898. That should give you some ballpark figures, a stock clocked 975 isn't going to be much slower. Interestingly two E5462's (Quadcore 2.8GHz 12MB L2 cache, 45nm Penryn cores) scored 19,545 on the Julia test.

After reading your comments about the type of programs used, it strongly reminded me of the Stars Euler3D program. Here's the site: Euler3d Benchmark

TechReport loves this program and uses it in even their server testing. Because they do a good job describing the benchmark in detail I'll give them a brief mention here where you can see the program in action on the Core i7 975. If you want I'll run it on this system to show what the clockspeed boost would net.

I fully agree with Rob in that the memory isn't going to make much difference at this point. It sounds like these apps are as CPU-dependant as one could get... so sticking to midrange memory for the best price/performance and spending the money saved on a higher clocked CPU would be best. 6GB of CAS 7 1600MHz DDR3 is floating around $70 still, so that's pretty much the best option just now in my opinion.

[Rob Williams]

Kougar, the Euler3D application looks good at first glance. Would you say it's something that would prove useful in our content? I can't look too deep into it right now, but I will next time I get a spare chance.

[Kougar]

The Euler3D program is pretty good... it is not the full program, but a slice of it the author used to run a basic level generic test case scenario. Still it is entirely CPU bound and scales well, especially on Core i7. It'll take a little work to set up as it's all command line console based, but if you are looking for a program to test CPU's with, then yes I'd recommend it.

[Psi*]

I have over looked this thread ...

Maybe a run of Euler3D with changing only the memory speed of an I7 system is possible? Keep the CPU clock, motherboard, memory timings as consistent as possible in other words.

[Psi*]

Originally Posted by Rob Williams

http://techgage.com/reviews/intel/i7_975/13.png

I meant to comment that this link compares memory bandwidth of systems including the affect of the CPU clock rate ... it would seem. I compared the % change of CPU clock and float MB/s between the i7 systems and it does not directly track the CPU clock. I'm looking for the test methodology page ...

Sooooo, I think my question all along has been "have you compared changes in the speed of the memory clock?" And, you probably have a table someplace that I have been seeing & but not reading.

[Psi*]

http://techgage.com/article/intel_co...t_memory_kit/2

[Psi*]

... I plotted some of the data from that article, "Intel Core i7 - Choosing the Best Memory Kit". The attached screen capture from Excel suggests (from the last data point) multi-core performance may benefit from higher clock rates. Only 1 data point to go on here and not to go irrational, but this is begging for higher clocked memory to be tested.

And, I think that this is suggesting that the slower memory bandwidth limits multi-core performance? Depending on what faster memory speeds would show of course.

[Rob Williams]

Just to be clear, when I tested memory configurations in that article, the CPU clock was not changed. That's part of the fun of Core i7... the ability to adjust memory speeds without touching the CPU. When that article was posted, we were testing with pretty common kits, but the landscape has changed drastically since then.

That said, we really, really should follow-up to this article, but I'm really strapped for time right now. I'm hoping that things will change in the near-future, because I'm actually rather interested in testing out even more configurations, especially with faster memory. I'd especially like to introduce even more real-world scenarios, to better get an idea of the benefits for the regular consumer.

As for slower memory limiting the multi-core performance, that's a possibility, but I think the real-world implications would be negligible. After all, there were few tests in this article that showed incredible gains when even doubling the frequency.

[Psi*]

It is even more difficult to understand the difference between the multi-core values of 40.46 vs 44.13 suggests for what ever real world applications happen to be.

And, now there is a D0 stepping of the i7-975 which somewhere I just read adds a performance boost vs the C0. Technology changes like this obsolete reviews pretty fast.