gavrik

RAM for Intel 975X - Do enthusiast grade modules improve performance ?

Recommended Posts

The interesting discussion about the need for ECC modules inspired me to post this question.

Leaving the ECC or not to ECC question alone for now -- what about the RAM clock frequency? I am also thinking of building a system with the 975X chipset and am wondering about the efficacy of using RAM modules with "enthusiast" timings.

I am assuming that the enthusiasts are either overclocking the FSB and want to run the RAM at a frequency that 'matches' the FSB or that they are running the RAM "faster" then the FSB. By 'matches' I mean that the RAM data transfer rate matches the FSB transfer rate. The FSB is 64bits wide and a dual-channel memory system is 2x64 = 128bits wide. If the FSB is clocked at 1066M data transfers per second then a memory system clocked at DDR2-533 would match the FSB because their data transfer rates are the same.

I know that calculations like this -- based on simple nameplate specifications -- are not the whole story, but I want to understand how this works. I understand that reducing latency improves performance. But if the latency in nanoseconds was held constant and the memory clock frequency is increased above the frequency that matches the FSB frequency (above DDR2-533 for a 1066 FSB) what happens to memory system performance? Once you clock the memory system above DDR2-533 the memory system becomes asynchronous. What is the consequence of this? Is the memory system buffer is deep enough that this is not a problem? Would a system with a 1066 FSB and modules running at DDR2-667 actually perform better then with modules running at DDR2-533 and the same latency in seconds?

Share this post


Link to post
Share on other sites

First let me say that the enthusiasts or more specifically, the overclockers, are hobbyists much in the same fashion that automobile performance enthusiasts are. In both cases you have classes of people who think they know better than the engineers who designed the product, or simply prioritize their own needs above the tradeoffs made by the engineers who designed the product. The tradeoffs specifically are lifetime and reliability.

Internal combustion engineers design an engine that makes 250HP and lasts 300,000 miles. Then an enthusiast gets one, changes timing, compression, breathing, whatever, squeezes another 50HP out of it, but it only lasts 50,000 miles before major engine parts fail, or it over-stresses and breaks the transmission. Or both.

Same with the computers, except the overclockers think they're smarter than the racer boys.

Guys who don't know an electron from an electrician will tap the sage advice of others who know just as little, make some seat of the pants assumptions about how much they can get away with, and start twiddling the knobs and adding massive cooling systems, trying to squeeze more performance out of the rig. And the results are predictable. People burning out motherboards and video cards, damaging memory and CPUs, and for what. Not productivity. Not when you count the time spent tweaking. It's just like the racer boys - it's mostly about bragging rights and the thrill.

OK, diatribe over.

Enthusiast RAM offers slightly lower latency potential and yes, performance increases can be had. In some cases you can gain up to 3% or so in overall memory bandwidth just by really tightening up the timings, i.e. from 5-5-5-15 2T to 3-3-3-9 1T. But memory bandwidth, especially with the Core 2 Duo systems, generally does not translate directly or linearly to application performance. When you look at the cost of enthusiast RAM you'll see that a percent or two in performance ends up being very costly. Especially now with the inflated cost of DDR2 RAM.

Enthusiast RAM also offers higher frequency operation. Here's where the real benefits can be achieved. But it's not always cut and dry. For example if you read the C2D overclocking results in some of the enthusiast forums, you'll find that running 667MHz out of sync actually decreases performance. The C2D platform generally responds well to running 1:1, so if you want to use the RAM at 667MHz that would mean running the system clock at 333MHz and either leaving the CPU multiplier as-is and putting extra cooling in place, or dropping the multiplier by one or two to bring CPU frequency back to something closer to design limits.

Or, use 800MHz RAM and choose a different set of tradeoffs.

There's a LOT to this. No need for me to re-invent the wheel for you, just search out the info on your own. But everyone agrees it's good to run 1:1.

-Brad

Share this post


Link to post
Share on other sites
Enthusiast RAM offers slightly lower latency potential and yes, performance increases can be had. In some cases you can gain up to 3% or so in overall memory bandwidth just by really tightening up the timings, i.e. from 5-5-5-15 2T to 3-3-3-9 1T. But memory bandwidth, especially with the Core 2 Duo systems, generally does not translate directly or linearly to application performance.

Let me second Brad.B on this one. The reason a faster DRAM doesn't translate linearly into performance is simple - not all memory-accessing instructions actually go all the way to DRAM. That may sound nonsensical, yes, but the reason is simple: caches. Most loads and stores actually get resolved into the L1 and L2 cache (a Core 2 Duo has about 4MB of L2 cache ... that can accomodate most applications' working set quite easily).

BTW, a load that hits into L1 has a latency of about 4-5 cycles. If it misses into the L1, but hits in the L2, the latency is around 10->20 cycles. Finally, if it misses all the way to memory (i.e., DRAM), the latency is around 200->400 cycles (!!!).

Share this post


Link to post
Share on other sites
Do enthusiast grade modules improve performance ?

If the enthusiast RAM modules have lower timing and/or higher operating frequencies than RAM that you might otherwise purchase, then yes. Depending on the application, these benefits can range from 0-5%. A better question would be:

Is it worth the extra money to purchase enthusiast grade modules to improve performance?

The answer to this is more formulaic, and requires work on your part.

1. Find reviews that compare the speed of different DIMMs in the types of applications you will use, preferably with the same chipset you will use.

2. Find the average speed increases.

3. Find the cost of stock DIMMs and enthusiast grade DIMMs.

4. Total the cost of the computer system you are building with the stock and enthusiast DIMMs.

5. Calculate the % difference in speed between the computer with stock and enthusiast DIMMs to the % speed increase offered by the faster DIMMs.

If the speed increase is less than the cost increase, you are probably better off spending the extra money upgrading some other part. Here is a synthetic example:

2. Normal speed in my favorite game is about 2%.

3. Stock DIMM is $200, and the premium DIMM is $230.

4. System base is $1800, so with RAM is $2000 and $2030 respectively.

5. Difference on the system's price is 1.5%, and the difference on the performance is 2%.

So buying the premium DIMM would increase the performace/$ of my system for my application.

This same formula can be repeated for all common upgrade items to see where your money is most effectively spent. If reviews aren't available for your scenario, you could always make your own benchmarks. (If you have a Fry's in the area, you could buy many different system parts, benchmark them all, and then keep what gives you the best performance for the money you are going to spend. ;) )

Share this post


Link to post
Share on other sites

One other note. Core 2 processors do better with 'matched' memory speeds (i.e. the DDR claimed MHz half the processor bus claimed MHz, so for a 1066 MHz bus processor, 533 MHz memory would be best.) But if you happen to be running an old Pentium 4, Pentium D, or Pentium Extreme Edition, the faster the memory, the better. It's all based on the memory prefetching of the two architectures.

Either way, having lower CAS, RAS, etc, numbers at a given MHz memory speed will improve performance a slight amount. Atamido has a good plan, find real world benchmarks that show what increase YOUR workload will see, and see if it's worth the money.

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now