FAQ-DDR3 Memory Selection, and Scaling ------

[Mr Fox]

I'm a pragmatist when it comes right down to it....


There have been some very interesting articles that have come down the pike...

Both at Tom's Hardware, and Anandtech's Gary Key...

First I will layout the Tom's Hardware Articles,

There are two articles that are clearly looking at this subject, and they are all excellent !!

The first article looks at all three platforms :

http://www.tomshardware.com/pi...mory-scaling-ddr3.html

This Article is Corei7 Specific

http://www.tomshardware.com/re...y-scaling-i7,2325.html

The below was excerpted from here :

http://www.tomshardware.com/re...caling-i7,2325-11.html

Tom's Hardware

Conclusion

2:00 AM - 06/09/2009 by Patrick Schmid and Achim Roos

The memory industry, which typically focuses on offering faster (and more expensive) memory products, must find these results rather annoying?there is simply very little benefit to fast DDR3 memory on a fast processor such as the Intel Core i7-975. More than ever, premium RAM can be compared to a high-end sports car: it may provide better performance, but the benefit in everyday life is often very limited.

DDR3 memory kits capable of running beyond DDR3-2000 speeds at increased voltage are the best you can get. However, we limited the testing to DDR3-1600 speed, and stayed at a voltage level of 1.65 V for this article, which represents a reasonable maximum for most users (hitting 2133 in our i7-975 review required a screaming 1.75 V QPI). Our testing included all other selectable RAM speeds, all the way down to DDR3-800, trying both tight and relaxed latencies for each of the RAM clock speed settings.

High End Memory Is Not Worth The Money

The results are obvious: going from one memory speed to the next, e.g. from DDR3-1066 to 1333, does not provide major benefits. Even the replacement of slow DDR3-800 RAM by DDR3-1600 memory will mostly yield disappointing results. While the performance advantage is measurable, it is never noticeable.

Exceptions, however, do exist. Compressing files with WinRAR is significantly quicker on fast, low latency DDR3-1600 RAM. Some applications, such as games, can at least take minor advantage of the upgraded memory horsepower.

What If?

In this light, we decided to add a few more benchmark results at overclocked processor speeds. We decided to accelerate the CPU by one clock speed increment, which reflects exactly what happens if you decide to purchase a faster processor instead of high performance memory.

Click here for all faster CPU benchmarks results.

As you will see, a higher CPU clock speed typically provides better performance in most of the benchmarks?but not in all of them.

Memory Recommendation

Knowing all of these results, it is obvious that highest speed DDR3 memory only makes sense for serious enthusiasts, or for those with unlimited budgets. Everyone else should focus on mainstream clock speeds of DDR3-1066 or DDR3-1333, and go for a trustworthy brand and the quickest timings their budget allows.

I have nothing but the highest regard for Gary Key, and feel that he has led the charge at being honest about FUD when it is found..

The Article is here :

http://www.anandtech.com/memory/showdoc.aspx?i=3589&p=1

Below is excerpted from here:

http://www.anandtech.com/memor...owdoc.aspx?i=3589&p=13

Anandtech - Gary Key

We raised the question at the beginning of this article: is there any value in using memory faster than DDR3-1066 on the Core i7 platform? In certain situations that answer is a definite yes and in others we really doubt the actual value of using anything more than a good triple channel DDR3-1066 6GB kit running at tight latencies, at least CAS 7 and preferably CAS 5.

What we discovered is that faster memory certainly makes a splash in our synthetic Everest benchmarks with memory read, write, and copy speeds showing improvement in the 40% range when moving from DDR3-1066 C7 to DDR3-1866 C7. Latency improvements improved by 30% in the same tests. However, we expected this, and for that reason we did not run the standard Super Pi or early 3DMark tests that heavily depend on memory and cache speeds for best results.

Once again, as we moved to real-world applications, those impressive synthetic benchmark improvements did not translate into results that would justify spending three times as much for a memory kit for most people. We had mixed with certain applications like WinRAR producing a 20% improvement from DDR3-1066 C7 to DDR3-1866 C7 while several applications showed minor performance improvements under 2%. If your primary job is to compress and archive files for a living, then the expenditure for fast low latency memory is justifiable. However, the decision to spend additional funds on higher performance memory is quickly up in the air after this point.

In our 3D rendering tests like Cinema 4D R11 and LightWave 3D 9.6, which we know to be sensitive to memory speed, additional memory bandwidth can provide tangible performance improvements of up to 7% or greater. Certainly, our multitasking benchmark showed the benefit of both greater memory and reduced latencies with performance improving 8% as we heavily loaded the system with multiple tasks and a large memory footprint. For users in this category, we have to agree that improving memory bandwidth will be beneficial.

However, common desktop applications such as Excel, Photoshop CS4, iTunes, and others just do not benefit that much from improved memory bandwidth or latencies. That brings us to games. Average frame rate improvements improved up to 7% by increasing bandwidth and reducing latencies but we never noticed the difference when actually playing the game. However, we noticed minimum frame rates improving up to 14% as we increased bandwidth and reduced latencies. This is an important fact as minimum frame rates are a better indicator of performance than averages in most cases. In our particular benches, the improved minimum frame rates took Dawn of War II from being a stutter fest in heavy action sequences to relatively smooth when moving from 1066 C7 to 1333/1600 C6.

Once we overclocked our system, the playing field equalized for the part as latency improvements had just as much impact on performance, if not more so at times, than bandwidth in most of our applications. In fact, in our multitasking test that showed an improvement of 8% at stock speeds, the difference between 1200 C5 to 2000 C8 was only 2%. Even our top responding application, WinRAR, managed just a 4% performance increase when moving from 1200 C5 to 2000 C8 compared to the 20% increase when moving from 1066 to 1866 in the stock test. The increase in CPU speed outweighed any potential gains in memory bandwidth or latency improvements in our benchmarks.


Based on today?s overall results, we have to question the validity or purchasing high-end memory for most users. Whether we like to admit it or not, most of us home users tend to be single task users when it comes to running an important application or game. Sure, we might have a few IM programs open, several browser windows, email, an occasional video or audio application, maybe Word, and then we have a major application like Photoshop or a game open. Although most gamers I know will close just about everything down to get the best video performance, several still run many applications in the background while gaming.

We might consider this multitasking but in reality, we have many programs sitting in the background while concentrating on a single task like Photoshop or Premier Pro as an example. We typically are not encoding last week?s party video in MainConcept Reference, have Cinema 4D R11 and LightWave 3D 9.6 rendering our latest artistic creations, and Photoshop is making us look a lot better in a swimsuit than we ever could while watching the latest BD title with Grandma singing Boom Boom Pow on Skype Video.

Of course, those that are doing all of these activities and more will benefit from purchasing fast low-latency memory and we even suggest getting 12GB while you are at it. For the rest of us, the primary applications we run and whether or not we overclock will have a large say in what is best for us. There is also the budget to consider, as most of us actually have to adhere to one. Looking at it from a budget viewpoint, is it worth paying 225% more for a DDR3-2000 C8 6GB kit over a DDR3-1066 C7 kit for an average performance improvement of 5% across a wide range of today?s most popular applications?

For a significant portion of users who run at stock or near stock speeds with system stability being paramount, we think not. For these users we suggest a DDR3-1066 C7 kit like the one we used from Patriot that has the capability of performing at 1066 C5 with a small bump in voltage and that can reach DDR3-1600 C9 (an excellent comprise setting due to pricing changes this past week) at warranty and system friendly voltages. This allows some growth potential in the system or the ability to increase bandwidth in the future at relativity low cost. Another excellent option in this price range is the GEIL DDR3-1333 C7 6GB kit we used that had no problems running at 1333 C6 or 1600 C8 with a small yet warranty friendly bump in VDimm and VTT. In fact, this particular kit or ones like it hit the performance sweet spot for users wanting very good performance with an eye to future overclocking efforts.

We know there is another section of users, enthusiasts to be exact, who balance their needs between stability but also extracting as much performance out of their systems on air or water cooling who need something more than DDR3-1066, even at CAS5. For these users we highly suggest going with a DDR3-1333 C6 6GB product or one that will do C6 with minimal voltage increases as a base choice. Based on recent price adjustments, a good budget to performance compromise would be DDR3-1600 C9. However, in individual module testing our 1333 C7/C6 kits had more clocking headroom along with the ability to run lower timings than our 1600 C9 sample. For those who tend to overclock 24/7 while running a multitude of applications, a DDR3-1600 C6 6GB capable kit is our best suggestion without going into debt on the DDR3-2000 kits.

We tend to favor running low-latencies at any given memory speed to ensure the best possible performance and as such we tend to stray from any of the CAS8 or CAS9 kits below DDR3-1866, unless they are capable of running much lower timings on like voltages. As such, we think DDR3-1600 C6 offers the best overall application performance in the market today for the enthusiast if you are willing to pay for it. That could be for a C6 certified kit similar to the ones we used from Mushkin or several of our DDR3-1600 C7 or DDR3-1866/2000 C8 kits had no problems running 1600 C6 with proper VDimm and VTT settings. In the near future, we will review all of the kits utilized today to determine the best value in each category, but for now, consider our choices to be safe ones.

Of course, for the extreme users dedicated to benchmarking, you can toss out all of our recommendations and just go for the DDR3-2000 C7 or better kits. In the end, there is no denying that the Core i7 processor will always perform better when paired high bandwidth low-latency memory. Just how much better depends on the application or situation, but Intel was certainly in the ballpark when they designed this platform around low-latency triple channel DDR3-1066.

You really do not give up that much performance with DDR3-1066 when compared to the more expensive alternatives and that is a good thing to know when putting a system together on a budget. As such, there is still great value in using it. If you have budget freedom, first off I am sure you will be a memory company?s best friend. Secondarily, temper your desires to go straight for the DDR3-2000 kits. It should be a comforting thought that purchasing a DD3-1333 C6 or DDR3-1600 C6 capable kit will offer the balanced performance you are seeking at price that should still allow a night out on the town this month.

Hopefully this will serve the community, and end the endless questions....

Thank You !!

 

[Idontcare]

No one should really be surprised, there is a reason the on-die cache hierarchy has increased to L1, L2, and now L3 caches totally megabytes in size.

These caches, if they function as intended, combined with hardware prefetchers will make the system all the less dependent on the underlying ram's speed provided it is merely above some silly low critical minimum bandwidth so the prefetchers can function to their design.

http://forums.anandtech.com/me...112855&highlight_key=y

Now if we had 4GHz quad-core rigs that were attempting to function with 16KB L1$ alone, and everything else relied on the ram speed, then you'd probably see a linear correlation between ram speed and computer performance.

Still though, its good to see the topic getting attention so more people stop and think about it. Personally I don't need super duper fast memory, I just need super duper quantities of it. (48GB would suffice for my current application needs) And I'd rather not spend $6k on it either, so until those 8GB density IC's roll out of samsung in 6-8yrs I'll still be doing the distributed computed thing across multiple cheap nodes with their paltry 8GBs.

 

[endervalentine]

Prob. a stupid question, but what's stopping cpu manufactures from increasing the cache size or creating another hierarchy, ie. L4? Is it silicon real estate?
If that's the case, why create another intermediate place between cache and DRAM?

And with all the interest in SSDs, why not have a bank reserve for SSDs that's tied to the motherboard, this should still give a boost of performance without the spike in cost of springing for a SSD drive?

 

[Idontcare]

Originally posted by: endervalentine
Prob. a stupid question, but what's stopping cpu manufactures from increasing the cache size or creating another hierarchy, ie. L4? Is it silicon real estate?
If that's the case, why create another intermediate place between cache and DRAM?

Yep, basically cost. Until the cpu performance is hampered by dram speeds there won't be a compelling reason to add yet another level of cache, but the day will come. Maybe 5-7 yrs.

Originally posted by: Idontcare
FWIW while I was researching something else I came across this article by Goto-san regarding die-stacking dram and cpu.

Original link in Japanese

Google translated to English

Notice this slide from Intel IDF last year, kinda funny they use the term "feed the beast".

There are even IDF slides detailing the various trade-offs to dram die stacked on top or on bottom of the the cpu. Looks like we are about a year late in our conversation here :laugh:

http://pc.watch.impress.co.jp/...008/1226/kaigai_8l.gif

from: http://forums.anandtech.com/me...=2334782&enterthread=y

Originally posted by: endervalentine
And with all the interest in SSDs, why not have a bank reserve for SSDs that's tied to the motherboard, this should still give a boost of performance without the spike in cost of springing for a SSD drive?

What you describe is functionally close to what Intel previously called Turbo Memory and Microsoft calls Readyboost. Intel was going to rebrand it as Braidwood in future desktops but it looks like those plans are on hold at the moment.

 

[alcoholbob]

The minimum framerate topic will certainly resound with gamers

 

[endervalentine]

Originally posted by: Idontcare

Originally posted by: endervalentine
Prob. a stupid question, but what's stopping cpu manufactures from increasing the cache size or creating another hierarchy, ie. L4? Is it silicon real estate?
If that's the case, why create another intermediate place between cache and DRAM?

Yep, basically cost. Until the cpu performance is hampered by dram speeds there won't be a compelling reason to add yet another level of cache, but the day will come. Maybe 5-7 yrs.

Originally posted by: Idontcare
FWIW while I was researching something else I came across this article by Goto-san regarding die-stacking dram and cpu.

Original link in Japanese

Google translated to English

Notice this slide from Intel IDF last year, kinda funny they use the term "feed the beast".

There are even IDF slides detailing the various trade-offs to dram die stacked on top or on bottom of the the cpu. Looks like we are about a year late in our conversation here :laugh:

http://pc.watch.impress.co.jp/...008/1226/kaigai_8l.gif

from: http://forums.anandtech.com/me...=2334782&enterthread=y

Originally posted by: endervalentine
And with all the interest in SSDs, why not have a bank reserve for SSDs that's tied to the motherboard, this should still give a boost of performance without the spike in cost of springing for a SSD drive?

What you describe is functionally close to what Intel previously called Turbo Memory and Microsoft calls Readyboost. Intel was going to rebrand it as Braidwood in future desktops but it looks like those plans are on hold at the moment.

Thanks for the rely. So I'm curious now, do you have the link to the IDF that lists the different trade offs between the dram die stacking?

Also, any ideas on why Braidwood was held off? Seems like a natural progression to me.

 

[Idontcare]

Originally posted by: endervalentine
Thanks for the rely. So I'm curious now, do you have the link to the IDF that lists the different trade offs between the dram die stacking?

Not to the original IDF presentation but Goto-san has the three slides of relevance linked here:

MCM'ed DRAM

Stacked DRAM on top

Stacked DRAM on bottom

The tradeoffs are relatively straightforward, signal length (MCM) versus thermal load on dram and cpu (dram on top) versus limited IO and power connectivity (dram on bottom).

Now the IDF pictorials show the dram die being as large as the CPU die, but my expectations are that they would intentionally/intelligently design a cache block in the CPU to have a floorprint that is amendable to producing discrete dram IC's as well, and then they'd die-stack the dram's over the L3$ (or some other "intrinsically cooler" part of the cpu IC where thermals aren't so debilitating) using TSV thru the cpu backside (so the dram stack is between the cpu die and IHS) and then use a "keyed" IHS that accommodates the change in step-height of the underlying IC package (the backside topology) such that the logic areas of the cpu die are not dissipating their heat directly thru the dram IC's as the heat is conducted to the IHS but at the same time the dram stack itself is also cooled by contact with the IHS.

Something like this crude hacked together picture I whipped up: http://i272.photobucket.com/al...tackeddramunderIHS.jpg

This minimizes signal distance, reduces thermal load on drams as well as improves thermal conductivity for logic areas of the die, and retains full IO and power connectivity options.

Originally posted by: endervalentine
Also, any ideas on why Braidwood was held off? Seems like a natural progression to me.

Cost versus performance. SSD's coming down in price and capacities increasing to the point that having an 80GB SSD in your system is feasible and entirely negates all the benefits of braidwood.

Braidwood was like a poor-man's effort to get 75% of the benefits of replacing your spindle-drive with an SSD for OS and apps...but with poor-man's SSD's coming down in price so much it makes the even cheaper alternative (braidwood) a non-marketable feature and in the end these guys are only interested in building in features that yield product differentiation they can sell.

Presumably someone ran the latest marketshare and $/GB projections for discrete SSD's and concluded Braidwood would have limited marketability over the timeline needed to justify its development/implementation cost overhead.

 

[pcslookout]

Originally posted by: Astrallite
The minimum framerate topic will certainly resound with gamers

Yeah minimum framerates are pretty important in games that are demanding so you don't notice any kind of slowdown even at the most demanding parts. I didn't expect there would be such a large gain!

 

[jimhsu]

Hm.. on minimum framerates - I can see the following bottlenecks:

1. CPU limited. AI/physics calculations. Um... upgrade to a faster processor?
2. Memory to GPU swapping limited (what is discussed here).
3. GPU limited - drawing too many polygons, etc. Upgrade to a faster GPU.
4. Hard drive limited - the classic "loading time" and stuttering that most people consider in minimum FPS. Helped by not solved by SSDs.

The problem is, in light of all the other bottlenecks, is 2 ever a realistic bottleneck? Anandtech shows that it can be, but is it the latency or the raw bandwidth that is the main factor here? Crucially, the compromise question - is it better to sacrifice latency for bandwidth, or vice versa? This is the question that has many overclockers up in arms.

Though, minimum framerates are certainly important. The experience of a constant 30 FPS is arguably vastly superior to a 60 FPS experience with a minimum FPS of 10.

 

[RockNRule]

I just shopped for ddr3 for a new i5 build. They are not even selling the slow stuff anymore as far as I can tell. I picked up a 4gig Corsair kit for $80 after a rebate. CL9 but I doubt it matters for real apps. I benched the build and it's just fine. So far a perfect match for the MSI p55-53