Experiments in RAM: Memory Myths Put to the Test

RAM, like water, is a commodity. And just as there’s a clear difference between putrid L.A. County tap water and water choppered in from the peaks of Mt. Everest, the quality of RAM can vary wildly. But quality is not the sole factor to consider when you’re trying to achieve optimum memory performance from your system.

These days, a user is faced with a plethora of options spanning different technologies, speeds, and capacities. We’re here to help you make heads and tails of all that so you’re prepared when you configure your next rig. Armed with a slew of RAM-based benchmarks, we set out to answer three of the hottest questions in memory today: Is DDR3 for AMD’s new AM3 Phenom II CPUs worth the expense? Should you pay for high-speed RAM or stick with the standard stuff? Finally, just how much memory is enough? We test three common amounts of RAM for Intel’s Core i7 to identify the sweet spot.

 

Intrigued? Then read on.

DDR2 vs. DDR3

AMD’s newest AM3 Phenom II CPUs finally bring DDR3 to the table, but is it worth the price?

If there’s one thing AMD knows, it’s how to be fashionably late to a RAM party. Sure, the company showed up early to the DDR dance and was the belle of the ball way back when, but these days, AMD shows up just in time to stack the chairs on the table.

Just as the company lagged far behind Intel in supporting DDR2, AMD has just now—more than a year and a half after its competitor—launched new AM3 chips that support DDR3, although its marketing of the new RAM standard is anything but enthusiastic. In fact, the company maintains that most consumers will prefer DDR2 because of its lower prices. And in a further display of halfheartedness, the company said that if we somehow got a hold of an AM3 board, we could test it, but the company recommended we use DDR2 and an AM2+ board for official Phenom II testing.

 Look very closely: A Socket AM3 CPU (on left) has two fewer pins than a Socket AM2+ chip.

Not exactly confidence inspiring, eh? To find out what’s behind AMD’s blasé attitude about DDR3, we decided to test a 2.6GHz Phenom II X4 810 with both DDR2 and DDR3.

Before we go on, let us explain just how AMD enabled both DDR2 and DDR3 in the same CPU. Generally, CPUs with integrated memory controllers are locked into the type of RAM they support. With Intel, for example, the company decided to support only DDR3 with its new Core i7 CPU. Intel could afford to make the break because there were no legacy Core i7 DDR2 users. With millions of AM2+ boards out there, AMD couldn’t burn its bridges, so it built the new Phenom II with both a DDR2 and DDR3 memory controller. Plug an AM3 CPU into an AM2+ board, and the DDR2 controller is activated. Plug it into an AM3 board, and it runs in DDR3. Got it?

The theoretical advantage that DDR3 memory has over DDR2 is the former is capable of deeper data prefetching; DDR3 can prefetch eight bits of data per clock cycle vs. four bits for DDR2. So, let’s see how this difference plays out in performance.

AMD builders are faced with the thorny question of whether to upgrade from DDR2 to DDR3.

The Test

For our testing, we plugged the 2.6GHz Phenom II X4 810 into the AM2+ MSI DKA790GX board. We slotted in 4GB of Patriot DDR2 RAM running at DDR2/1066 speeds. We then removed the Phenom II X4 810 and put it into the new Asus M4A79T Deluxe AM3 board. To this, we added 4GB of Centon DDR3/1333 in dual-channel mode. Both were tested with a GeForce 8800 GTX, an Intel X25-M SSD, and Windows Vista Home Premium in 64-bit.

To see just what you get for the trouble of building an AM3 box, we selected a set of synthetic memory benchmark tests in addition to gaming and application tests. We settled on 4GB of RAM for both platforms, since it seems like the sweet spot for dual-channel configurations. We ran the DDR2 at 1066MHz—the highest speed available—and the DDR3 at 1333MHz. If DDR3/1333 doesn’t seem like enough of a spread, we also ran a few spot checks with the DDR3 set at 1,600MHz and found essentially no difference.

The Results

No difference. That’s the story of the day with Phenom II in either DDR2 or DDR3 trim. Well, to be fair, not enough of a difference to write home about. From media encoding to application tests to games, DDR3 seemed to have no impact on performance over DDR2. Even in the synthetic memory benchmarks that push the RAM to its theoretical limits, we didn’t see DDR3 make much of a case for itself. That surprised us somewhat, because you can usually count on increased bandwidth to show an improvement in the bandwidth tests. Even with the RAM ticked up to DDR3/1600 rates, the performance increase was minimal in the synthetics. It’s no surprise that we didn’t see an improvement in the majority of our application tests. PCMark Vantage, which stresses various application workloads, actually saw a slight performance decrease from moving to DDR3, but in the synthetic gaming test, 3DMark Vantage, DDR3 pulled ahead by a slight amount.

With these numbers in front of us, we’re not surprised that AMD wasn’t hot to push us on DDR3 for Phenom III. For AMD fans who have a DDR2-only AM2+ motherboard and are considering chucking it for AM3: Don’t bother. At this point, it’s just not worth it. You’re better off just adding more RAM.

For those looking at a new build of a Phenom II system, the choice is far tougher, but we’d probably lean toward DDR3 and AM3. Sure, DDR2 is cheaper, but DDR3, especially at the lower 1,066MHz and 1,333MHz speeds, isn’t going to break the bank. With AMD’s current AM3 chips, the DDR3 performance isn’t spectacular today, but what if a new spin enhances the performance? We know, it’s wishful thinking, but it’s probably worth the bet. 

 

 

Next: DDR3/1066 vs. DDR3/1600

---

DDR3/1066 vs. DDR3/1600

Before you pay for the extra bandwidth, read this

Despite AMD’s statements to the contrary, DDR3 is pretty darn cheap. Yes, DDR2 diehards will scoff, seeing as how you can practically get 4GB of DDR2/800 with a Starbucks coffee, but for those of us who remember paying $500 for 2GB of DDR3 just two years ago, today’s DDR3 pricing is a hell of a bargain.

But that doesn’t mean you should spend indiscriminately. Besides the typical concerns over quality—name-brand memory is always preferable—users face another question: How much bandwidth  is worth the cost?

Paying for higher-bandwidth RAM will get you more performance, but not in every application.

Before we get too far into this topic though, let’s review the basics. Like DDR and DDR2, DDR3 uses double data rate technology, which sends and receives data on both the rise and fall of the RAM’s clock—DDR3 just grabs twice as much data as its predecessor. There are two common designations for DDR3: One is the module name, which describes the theoretical bandwidth, as in a PC8500 module or a PC10600 module. These numbers represent the peak transfer rate in MB/s and can become quite confusing. Quickly, tell us the effective data rate of a PC8500 DDR3 module. We’re still waiting.

Sometimes the module will also helpfully list the effective clock speed—the other common RAM designation—e.g, 1066MHz. But when it doesn’t, you can figure out the effective clock speed of a module by dividing the bandwidth by eight. For example, a PC8500 module divided by eight is 1,066MHz, and PC10600 becomes 1,333MHz. Mind you, this is not the actual clock speed, which is 1/8 the effective clock speed. The actual clock speed of a 1,066MHz module is 133MHz. Why all these inflated bandwidth and clock numbers? It’s a legacy of the old war with Rambus, when the DDR guys didn’t want the original DDR/200 (which ran at a core clock of 100MHz and had a bandwidth of 1600MB/s ) to sound inferior to Rambus’s PC800, which ran at 400MHz and had a bandwidth of 3200MB/s in dual-channel mode. Thus, the single-channel-only DDR/200 module became a PC1600 module. Voilà, it’s a bigger number than PC800! We don’t want to pick old scabs, so the take-away is that we prefer to use the nomenclature of DDR3/1066 or DDR3/1600, which is far less confusing and far more useful.

The Test

For our test, we settled on Intel’s big, bad Core i7. Why? Our previous tests proved it’s pretty much a wash between DDR2 and DDR3 on the Phenom II platform. Furthermore, Intel’s Core i7 features a triple-channel DDR3 controller, and its ballistic speeds seem best suited for our bandwidth test. We used an MSI X58 Eclipse SLI board with a 3.2GHz Core i7, an EVGA
GeForce GTX 280, an Intel X25-M SSD, and Windows Vista Home Premium in 64-bit trim. For RAM, we used 6GB of Corsair Dominator DDR3/1600 modules and 6GB of Qimonda DDR3/1066 RAM. Both were in tri-channel mode. For benchmarks, we selected a variety of synthetic memory, gaming, and application tests.

The Results

Coming into this from the uneventful Phenom II DDR2/DDR3 testing, our expectations were low. That’s not because there’s anything wrong with having extra bandwidth, but we’ve rarely seen it have much of an impact. Our experience has shown us that few applications actually require the added bandwidth that’s available today.

We believed that to be doubly so with Intel’s Core i7 series. The chip has so much excess memory bandwidth that anything more than DDR3/1066 seemed like overkill. With three DDR3/1066 modules in tri-channel mode, the chip has more than 25GB/s in memory bandwidth available. With three DDR3/1600 modules, it has in excess of 38GB/s available. The typical dual-channel DDR2 system today has about 13GB/s of memory bandwidth. While we could imagine the step up to DDR3/1066 having an impact with a Core i7, we were skeptical of the chip needing more than that.

We were wrong. At least partially. We were pleasantly surprised to see the Core i7 kick up some benchmarks by a noticeable amount with the DDR/1600. The extra bandwidth didn’t help everywhere, of course. It made no difference in Adobe Photoshop CS2 or Premiere Pro CS3. Likewise, 3DMark Vantage saw no difference going from 1066 to 1600.
But in PC Mark Vantage in 64-bit mode, we saw a rather healthy 9 percent boost going to DDR3/1600. Quake 4 (which we’ve long used as a bandwidth gauge) also showed a 7 percent improvement. Minor bumps also came in our ProShow Producer and MainConcept Reference encoding tests. The improvements weren’t huge, but they were unexpected, since the tests have historically been purely compute bound. Photodex, the maker of ProShow, has long told us that its hand-written encoder loves faster CPU-to-CPU interconnects, and it apparently gets a small increase in performance from RAM bandwidth too.

More predictably, the synthetic memory benchmarks saw sizable increases from the DDR3/1600. SiSoft Sandra showed a whopping 44 percent increase, while Everest Ultimate showed similar gains of 35 percent to 45 percent.

The upshot of this: Unlike the Phenom II, giving the Core i7 more bandwidth will be worth it for folks who want the absolute tops in performance. Expectations must be tempered, however, because a great majority of applications will likely see very little difference. So if you’re on a budget, saving cash by going with DDR3/1066 or DDR3/1333 isn’t going to kill you. That’s assuming you’re running tri-channel though. Our initial tests with Core i7 showed that there are severe performance penalties for running RAM in single-channel mode. Dual-channel is better, but tri-channel will give you the best performance
possible on a Core i7 system.

 

 

Next: 2GB vs 3GB vs 6GB

---

2GB vs. 3GB vs. 6GB

Less may not be more, but is it just enough?

You can never have enough money, ammo, or RAM. But is that really true? To find out, we looked at three very common memory configurations a person would face when building a new Core i7 machine. The first config is the upgrader. This person adopted DDR3 in the Core 2 days, and after paying $500 for 2GB of DDR3/1333 RAM, he sure as hell is going to migrate it to his Core i7 box, even if it means running the RAM in dual-channel.

The second configuration is the person who wants to scrimp a bit on RAM to perhaps buy a larger hard drive, so he buys 3GB of DDR3/1333. The third configuration is the geek who, sick of never getting more than 3.25GB of RAM in Windows XP, wants to break that 32-bit barrier. So he outfits his new Windows Vista Home Premium 64-bit machine with 6GB of DDR3/1333 RAM—and he damn well better see it say 6GB of RAM in the device manager!

Who has the optimum setup in terms of memory? Let’s find out.

The Test

For our test, we again reached for our MSI X58 Eclipse SLI board and first outfitted it with two 1GB sticks of OCZ DDR3 RAM to simulate our Core 2 Quad upgrader. The second configuration used three 1GB sticks of CSX Diablo DDR3 RAM. The third configuration used three 2GB sticks of Corsair Dominator DDR3 RAM. The Diablo RAM is rated for DDR3/2000 (although we could never boot our board at that speed with it), and the Corsair RAM is rated for 1600. We ran all three scenarios at DDR3/1333 speeds.

Again, we outfitted our X58 Eclipse SLI board with a 3.2GHz Core i7 Extreme Edition, an EVGA GeForce GTX 280 GPU, an Intel X25-M SSD, and Windows Vista Home Premium in 64-bit mode. We reached for the same combination of synthetic memory benchmarks, application, and gaming tests that we used in our other two challenges. Because we suspected that our standard benchmarks would not be enough to stress the amount of main memory, we also ran additional tests that would simulate a heavily loaded multitasking machine. The first was to launch Premiere Pro CS3 and open an HDV editing project. We then launched Photoshop CS3 and opened a 600MB PSD file. We also ran our Valve Particle Gaming test, which simulates multicore particle workloads. And we created an uber-workload by launching Premiere Pro CS3 and opening 30 images in Photoshop CS3, then running either Crysis or PCMark Vantage in 64-bit mode.

The Results

We saw fairly minimal differences between the dual-channel 2GB configuration and our 3GB and 6GB configuration in 3DMark Vantage and PCMark Vantage. Likewise, Valve’s Particle Test, Quake 4, and ProShow Producer performed the same regardless of memory amount. Everest Ultimate also showed very little difference among the three configurations, which was a bit surprising, since we’ve seen a difference between dual-channel and tri-channel in this benchmark in the past. SiSoft Sandra’s results, however, saw the tri-channel configs showing additional bandwidth. Again, we didn’t expect much of a difference with our standard tests, which ran fine with 2GB of RAM in dual-channel mode. We did, however, see
pretty good gains in our Photoshop CS3 test and Premiere Pro CS3 tests. Both handily favored the 6GB configuration over the 2GB and 3GB builds. We also saw minor improvements in ProShow Producer and MainConcept Reference.

Having 6GB of RAM will yield the best overall performance, but we were surprised by how well 2GB and 3GB did.

OK, so what about our tests designed to specifically stress big memory holes? That’s where we were first thrown for a loop. With Premiere Pro CS3 and Photoshop eating up a chunk of RAM, running our Valve Particle Test showed absolutely no difference. Why? We believe that all three apps were still below the 2GB threshold. That’s where we expected Crysis to make a big difference. With Premiere running and 30 large images open in Photoshop, surely we’d see a frame rate drop in this monster of a game, right? Nope.

Crysis is said to need 4GB of RAM to run smoothly, but the benchmark spit out the same frame rate with 2GB, 3GB, and 6GB installed. What gives? We suspect this is the result of one of the things Microsoft did right with Vista. The OS is far more efficient at memory management, and it likely just pushes aside applications to make room as needed. Performance was likely ameliorated by our use of Intel’s superfast X25-M SSD. This drive’s 200MB/s-plus read speed and zero access time is quite capable of masking disk caching that would be more apparent with a slower drive. We have to note that although the benchmark frame rates are fine, actual gameplay would probably suffer with these apps open—level loads would be slower, given that the OS simply has less RAM to play with.

PC Mark Vantage combined with Premiere and Photoshop actually gave us the results that we expected and in our opinion is a more reliable predictor of heavy memory loads. Unlike a gaming benchmark that’s fairly small and runs for only a few seconds, this application simulates application performance and tries to measure how well a system runs. For example, in one section, the app plays a sample of HD footage and measures the frame rate. We could visibly see PC Mark Vantage struggle at times to complete runs with the other application open, and even the video playback would stutter with the 2GB and 3GB configurations. It was only at 6GB that we saw acceptable performance with PC Mark Vantage.

Our take-away: You can actually get by with less amounts of RAM if you are a single-tasking kind of person. Even the 2GB, dual-channel combos yielded acceptable performance. We don’t recommend it though. The minimum a Vista 64-bit PC should run is 3GB and the recommended is 6GB. We say this because today’s CPUs are designed to have multitasking workloads thrown at them. With so many available threads, you’re expected to run multiple applications without significant performance slowdowns. By not having enough RAM, you hurt the machine’s multitasking capability and you won’t work as efficiently. 

 

 

 Next: RAM tips for buildling your new PC

---

Location Matters: Fill Your Slots Wisely

It’s a typical rookie PC move: In the excitement of building your machine, you jam the RAM into any slot available. Unfortunately, that’s a practice that can lead to instability or even lower performance.

Here are a few basic tips to help you avoid problems when building any new PC:

• Make sure you’re running on all channels. Even seasoned system builders can get confused by the RAM coloring schemes on motherboards. Asus, for example, used to require that you put the RAM into slots that were the same color to get dual-channel support on its motherboards. MSI, however, used to do the opposite, so you had to have the RAM in slots that weren’t colored the same. This would cause even old salts to pull a brain fart and set a new motherboard in single-channel mode by mistake. The solution is to read the manual before you build the board to find out what the recommended slot config is.

• Put the RAM in the slots closest to the CPU—except when you need to put them in the slots furthest from the CPU. Yup, you got that right. With Athlon 64, Phenom, and Phenom II, it’s generally recommended that you use the slots for each channel that’s closest to the CPU. For Core i7, it’s the opposite. In fact, many motherboards won’t even boot if you put the RAM in slots closest to the CPU first. Again, RTFM.

• void the fourth slot. Intel’s Core i7-based DX58SO motherboard features four slots for RAM. You should fill only three of these for tri-channel mode and avoid filling the fourth, since the company says it will impact overall bandwidth. Intel won’t be the last to design a Core i7 board with just four slots, so consider yourself warned.

 So You Want to Run the Most Memory Possible?

One of the most common questions we get is, why the hell doesn’t Windows XP show 4GB of RAM? The short answer is, 32-bit Windows XP can address only 4GB of RAM. Since address space is also set aside for the graphics card memory and other devices in the system, the majority of 4GB systems end up with 3GB to 3.5GB of RAM available to the OS.

Although you might think that 32-bit Windows Vista overcomes this, it doesn’t. Microsoft just changed the OS to report the total amount of RAM installed—part of that RAM is still being set aside for addressing memory used by videocards and other devices. The only way to fully utilize your 4GB, 6GB, or 12GB of RAM is to run with a 64-bit OS such as Windows Vista 64-bit, which inherently offers additional space for addressing memory other than RAM.

But that’s not all. We’ve found that the still-immature X58 boards and BIOSes aren’t the best at booting 12GB of RAM and seem to get confused on how to configure the memory. Before you fill all DIMM slots, install just three, do your build, and update the system to the latest BIOS. For example, we had to swap 6GB of Patriot RAM for 6GB of Kingston to go with our 6GB of Corsair RAM before we could POST our MSI board. Another trick you might try: Add a tiny bit of voltage. How do you do this if the system won’t POST? Fire it up with just three sticks of RAM and change your BIOS settings. Then power down, unplug, discharge, and install your three other DIMMs and reboot.

Finally, we recommend that if you truly intend to fill every dang DIMM slot on that new Core i7 board, that you go specifically with RAM from the QVL, or Qualified Vendor List, that each motherboard maker publishes on its website. This will increase your chances of having a flawless build. 

Core i7 Memory: Locked or Not?

If you read our Core i7 overclocking guide in the April issue, you know we reported that Intel locked the memory multipliers on its non-Extreme CPUs. We said Core i7-920 and Core i7-940 were simply incapable of running DDR3/1600 at its rated speed without overclocking the bclock. Well, we were wrong.

Most of our Core i7 testing was based on engineering samples of the CPUs. These chips are supplied to hardware vendors and the media to perform tests before the chips go into full production. Normally, the difference is nil between ES parts and production parts, but Intel made a change that unlocked the memory multiplier (and QPI speeds too). The company said it made the change at the last minute in response to “customer requests.” The change was so last minute that PC OEMs and even some Intel field engineers were unaware of it.

This should make cheapskates everywhere smile: Retail Core i7-920s feature unlocked memory multipliers.

This is, of course, good news for folks interested in running higher-speed memory without having to resort to an overclock. Production Core i7-920 and Core i7-940 chips are fully capable of running DDR3/1333, DDR3/1600, or higher-rated RAM, and not locked down at DDR3/1066. This also makes the budget chips just a heartbeat away from the top-end Core i7-965 in capabilities. With the ability to increase the QPI from 4.8GT/s to 6.4GT/s and run the RAM at any speed you want, the only real difference between Extreme and non-Extreme today is the overall multiplier lock of the CPU, fine-grain Turbo Mode settings, and some thermal override settings. This makes the budget chips far more attractive than they seemed just three months ago.