AMD is upon one of its greatest platform launches in a very long time and just the fact that Socket AM2 uses a, to AMD, completely new memory technology makes it extra interesting. We have taken a closer look at what to expect from Socket AM2 and its DDR2 support.

When it comes to AMD’s transfer from DDR1 to DDR2, we all knew that this logical succession would come sooner or later. Almost two years ago Intel announced its transition to DDR2 and that PC3200 would be the last officially supported DDR1 frequency. AMD, on the contrary, had different plans. They had built and optimized their CPUs for DDR1 and saw no reason to transfer to DDR2 if it, though higher frequencies, would mean slower access times. Not only did AMD keep DDR1 alive, it was encouraging the evolution and density of DDR1. Now, however, the limitations of DDR1 were starting to show, at the same time DDR2 could now be had with fast access times and high frequencies. It was now time for AMD to enter the DDR2 market.

Today, AMD is launching their new Socket AM2 with DDR2 support to match. The reason for the new socket is that the communication from the CPU to the memory is not identical between DDR1 and DDR2. The CPU’s built-in memory controller is thus also updated to cope with the new specifications. What these changes mean, technically and physically, and what kind of performance the change to DDR2 will bring is the main point of this review.

We will begin with listing the differences between Socket AM2 and Socket 939.

Physically there is not much that differs between the two sockets, Socket AM2 and Socket 939 respectively. The new socket accommodates one more leg, making the total 940.Though the number of legs is shared with some Opterons, they are not pin compatible. As seen in the picture above the blanks have been altered to prevent the insertion of a non-compatible CPU. Otherwise there is no change in either Socket or CPU.

Around the socket we see a prominent difference, the retention bracket and its mounting holes. AMD claims this new solution gives ”increased reliability and stability”, though we have never experienced any such problems. We find it to be rather a drawback as those who have invested in an expensive heatsink will have to get some kind of adapter, a compatible mounting mechanism or a completely new heatsink. No further restrictions to the surface surrounding the socket have been made, this is bodes well for extreme cooling devices.

Finally we have the new memory slots with 240 pins, for DDR2, as opposed to 184 pins for DDR1. The length of the slots is still the same so the number of pins per inch has increased. As seen in the second picture.

On the next page we’ll take a look at what is architecturally new with the new AM2.


It’s not very correct to claim that DDR2 support is a feature as it is a vital part of the CPU. AMD has, for the last couple of years, used an integrated memory controller that handles the communication between the CPU and the memory. This has evident performance advantages compared to Intel’s design that relies on an external memory controller integrated in the Northbridge. The news with the memory controller is that it now officially supports DDR at 533, 667 and 800 frequencies. The memory latencies have been taken up a notch and the three most significant adjustments starts at 3-3-3 and goes up to 6-6-6. However we have noticed a definitive drawback with the way AMD implements the memory ratios.


Memory frequencies (at 200MHz HTT)
Memory multiplier Memory frequency Change in performance
14x (2.8GHz) 7x 800 0%
13x (2.6GHz) 7x 742 -7%
12x (2.4GHz) 6x 800 0%
11x (2.2GHz) 6x 733 -8%
10x (2.0GHz) 5x 800 0%
9x (1.8GHz) 5x 720 -10%
8x (1.6GHz) 4x 800 0%

The memory frequencies are dependent on the CPU multiplier in correlation to a memory multiplier, e.g. we get DDR800 with a CPU with a 12x multiplier via a 6x memory multiplier, i.e. 200HTT * 12x / 6x = 400MHz = DDR800. A problem arises when using a CPU with an uneven multiplier for example 11x. As there are no ”half multipliers” you get the closest one above, in this case 6x. Using the above calculation we get 200HTT * 11x / 6 = 366MHz = DDR732. A quick check in the table reveals these CPUs lose between 7% and 10% memory performance because of this.

AMD Virtualization

AMD has added support for virtualization, which is equivalent to Intel’s Virtualization Technology. Virtualization is all about allocating different parts of the system to certain software. There are many different applications of this and to mention one is to separate critical systems from each other, for increased security, and letting more resources share the same hardware, to increase efficiency.

Energy Efficient (EE) – CPUs

With the AM2 AMD has introduced new alternatives to existing types, where extra care has been given to power consumption in the CPUs. These alternative types will be available in all speeds except the fastest ones. The x2 CPUs has had the specified power consumption lowered from 89W to 65W and the single cores will go from 67W/89W (dependent on type) to 35W. These are noteworthy improvements of the power development.

We’ll have a look at the CPUs about to be introduced for Socket AM2.

AMD Athlon64 FX – Socket AM2
Model Frequency Cache Power consumption
FX-62 2.8GHz 2x1Mb 125W

AMD’s FX series hardly requires any further introduction. The FX processors are intended for the most extreme enthusiasts and have specifications to show for it. As expected AMD has increased the frequency from FX-60 to 2.8GHz and with it a fitting price tag.


AMD Athlon64 X2 – Socket AM2
Model Frequency Cache Power consumption
5000+ 2.6GHz 2x512kb 89W
4800+ 2.4GHz 2x1Mb 89W (65W EE)
4600+ 2.4GHz 2x512kb 89W (65W EE)
4400+ 2.2GHz 2x1Mb 89W (65W EE)
4200+ 2.2GHz 2x512kb 89W (65W EE)
4000+ 2.0GHz 2x1Mb 89W (65W EE)
3800+ 2.0GHz 2x512kb 89W (65W/35W EE)

AMD has expanded its X2 series with a few extra models but also added EE versions, which you can see in the parentheses. We have reason to believe that there will be arrive even more processors in this series and then first of all a 5200+-model with equivalent FX-60 specifications. We also see that the 3800+-processor will be available as two different EE models where one has a TDP at a mere 35W.


AMD Athlon64 – Socket AM2
Model Frequency Cache Power consumption
3800+ 2.4GHz 512kb 62W
3500+ 2.2GHz 512kb 62W (35W EE)
3200+ 2.0GHz 512kb 62W

The Athlon64 series has been stripped somewhat and now only contains 3 processors all with 512kb L2 cache each. We can see that it will only introduce the 3500+ model as an EE version.


AMD Sempron 64 – Socket AM2
Model Frequency Cache Power consumption
3600+ 2.0GHz 2562kb 62W
3500+ 2.0GHz 128kb 62W (35W EE)
3400+ 1.8GHz 2562kb 62W (35W EE)
3200+ 1.8GHz 128kb 62W (35W EE)
3000+ 1.6GHz 2562kb 62W (35W EE)
2800+ 1.6GHz 128kb 62W

The budget series Sempron will be expanded on the other hand and also include EE versions. The tendencies to reduce the cache to save silicon and thus manufacturing costs are still there.

Before we start with the benchmarks we will take a look at the test system.


Test system

Reference system
Motherboard Asus M2N32-SLI Deluxe Asus A8N32-SLI Deluxe
Processor AMD Athlon64 5000+ AM2 AMD Athlon64 FX-60 939
Memory Corsair XMS2 8500 (2x512MB) Corsair XMS 3200 (2x512MB)
Graphics card ATI Radeon X1900XTX
Power supply OCZ PowerStream 520W
Operating system Windows XP (SP2)
Chipset drivers nVidia nForce 6.82 (x16)
Graphics drivers ATI Catalyst 6.2
Monitoring program Asus AI Booster
Benchmarking programs SiSoft Sandra 2005 SR3

3DMark2003 3.6.0

3DMark2005 1.2.0

3DMark2006 1.0.2

AquaMark 3

WinRAR 3.60


Everest Ultimate 2.80

Idle One hour in Windows without load
Load One hour with Prime95 running
Stable No errors reported by Prime during load
Processor temperature The temperature reported by AIBooster

In our tests of the AM2 we have chosen to use the Asus M2N32-SLI Deluxe motherboard because it uses nVidia’s nForce 590 chipset. We will not go into more detail because we will review this motherboard later on. Those of you who are observant have realized that we will test two CPUs with the same CPU frequency, but with different amounts of cache. This will of course not result in a completely fair comparison, like we would get with a direct comparison. That’s why we will focus on the memory performance, compare how different memory settings affect the performance and in the end show some general benchmarks and game tests. As we discussed earlier, we don’t have the possibility to activate true DDR800 on our test CPU and the maximum memory frequency will be at DDR740.

We will begin by testing how different memory timings affect the performance of the memory.

Because we wanted to make a good comparison between different latencies, we ran the memoryies at a certain frequency and only changed the latencies. For some reason, the motherboard refused to start with latencies at 3-3-3-8 when DDR800 was set, despite the fact that we, on another platform, verified that they worked with these latencies. We had to run them at the DDR667 setting which gave us a frequency of 326 MHz/DDR652. We’ve chosen to focus on the three most common latency settings: 3-3-3-8, 4-4-4-12 and 5-5-5-15.


We clearly see that in the pure memory tests latencies influence the read and copy speeds the most, while the writing speed is less affected. The latencies, which are what the memory settings control, reflects how the differences in the access times are affected by the different settings.


WinRAR benefits, in some degree, of the memory speed and we see a linear reduction of performance between the fastest and slowest settings.

Next we will examine the differences in performance when using different memory dividers.

Like we did on the previous page, we will lock the one of the parameters; the memory latency at 4-4-4-12, and test how the different memory dividers works and how they affect the performance.


We see the same trend with an exception of more separated results. Comparing our results with the previous page concludes that it’s more important to reach high memory frequencies for optimal performance.


Here we see that WinRAR is very dependent of the memory frequency to perform well. While we take a huge step upwards when using the DDR800 divider and we take an equally huge step downwards with the DDR533, with regard to the DDR667.

We combine the best timings and dividers and compare these with the Socket 939 on the next page.

Time to compare a setup of configurations which will most likely be the most commonly used with Socket AM2. We’ve chosen to run DDR667 3-3-3-8 latencies and DDR800 with 4-4-4-12 latencies and compare the performance with a Socket939 system running DDR400 and 2-2-2-5 latencies. Here we will once again acknowledge what we discussed under Test system, that we in fact use two processors with different amounts of L2 cache.


We can see that AM2 performs well compared to DDR1 when it comes to pure memory speed as the new platform takes a big lead. The difference in delay is something that has been discussed since the launch of DDR2 and here DDR1 is really hard to beat. An interesting comparison is between DDR667 at 3-3-3-8 and DDR800 at 4-4-4-12 where we see that DDR800 with the ”slower” settings is actually faster. The frequency has more significant impact on the delay between the two and we will most likely see more and even faster memories than 4-4-4-12, and then DDR2 will be up to par with DDR1. AIDA is not as consistent with this test and we assume that this write test use the L2 cache more than Everest’s tests does.


When it comes to WinRAR we suspect the difference between the two processor’s L2 cache has a big impact on the performance, so please consider these figures with a pinch of salt.

We continue the comparison with some general benchmarks.

The Socket939 system is victorious all the way through, but we have to emphasize the difference in L2 cache once again.

We see the trend where the DDR800 configuration offers better performance still goes, even though is close to minimal here.

Let’s move on and do some game benchmarking.

The trend continues and the DDR400 system remains on top through all game tests. Between DDR667 and DDR800 it is still even, even though the DDR800 configuration has a small, but constant advantage. Once again, we should point out that the main portion of the difference between the two systems comes from the different sized L2 cache.

We move on and conclude our experiences of Socket AM2 on the next page.

Physical changes
On the physical side of things, not many changes have been made from Socket 939. The thing that has been changed will most certainly annoy those who already own a high-performing heatsink made for two holes. Though, it’s even worse for those who use more exclusive cooling devices like water cooling, Prometeias and VapoChills. These usually have long periods of adaptation before any new products or adapters arrive for the new socket.

We haven’t had much possibility to test a large amount of different DDR2 modules and see how these behave with the new memory controller, but the pairs we’ve tested have worked without any problems. In the present-day situation, most memories are of the conservative type with slow SPD settings, which reduce the risk for troubles. The one compatibility issue (though a minor one) we ran into was the inability to use 3-3-3-8 with the DDR800 divider. We made verifications to ensure that these memories worked in this speed with good margins on an Intel platform, and they shouldn’t have encountered any problems on the AM2 platform. It also remains to be seen how AMD will solve the DDR800 setting on processors with uneven multipliers, which currently risk taking a performance hit of up to 10%.

Performance and stability
When it comes to raw memory performance, there’s no doubt about the fact that AM2 is a step in the right direction. DDR2 memories will continue to arrive with more and more aggressive speed specifications and we are convinced that AMD will add additional memory dividers to offer support for DDR1066 and further speeds. At the same time as faster memories surface, there will also be those with tighter timings, which offer a certain performance increase. As we can see in our tests, we earn more when setting the memories to higher frequencies, and therefore we don’t really mourn the absence of the 3-2-2-4 setting which is offered on Intel’s platform. We are certain that higher frequencies with tighter timings will arrive, and when they do, we won’t see any performance gain with DDR1 anymore.


AMD has with Socket AM2 a platform that will stretch from the budget segment Sempron to the real high performance models with the Athlon 64 X2 and FX series. It’s not very hard to imagine what AMD was thinking it wanted just one platform, but transferring to the new platform comes at a cost, and what a cost. Not just the processors themselves but also the new motherboard and memory you will need for the new AM2 platform. The processor prices that has appeared so far are identical, more or less, with the current Socket 939 models, which is of course a positive thing and it shows that AMD really wants new consumers to transfer to its DDR2 platform. Our sample, Athlon 64 X2 5000+, costs about 6800 SEK in Swedish stores but we haven’t been able to locate it in any North American stores, but we would assume the price would be $680-700, which is quite a lot. But since AMD has moved more or less its entire assortment to the Socket AM2 there are processors for all budgets, big and small. Also motherboards with NVIDIA’s new nForce 5xx chipsets seems to have equivalent prices as the nForce 4 models for Socket 939. Thanks to the fact that AMD has waited so long with moving on to Socket AM2 the previously expensive DDR2-SDRAM memories have become quite affordable and this includes the DDR2-800 modules, which we strongly recommend for the AM2 platform, alas they are still more expensive than the DDR-400 counterpart, although not by much. AMD seems to have done the right thing with its pricing of the Socket AM2 platform, simply not letting them become an obstacle.

In conclusion, we feel that even though AMD has worked for a longer period of time with optimization of the memory controller, there are still some things left to be done. It’s possible that these things will be solved via BIOS, while some things will remain until the next update of the memory controller. The memory performance is present, but it doesn’t show its full potential in comparison to a processor with a larger amount of cache. We hope to return with an updated test shortly with two equal processors to really analyze the performance difference.

Socket AM2

+ Memory performance
+ EE version available of many processors

– No true DDR800 with uneven processor multipliers
– 3-3-3-8 latencies did not work with the DDR800 divider

We want to thank AMD for sending us the processor for evaluation.


Leave a Reply

Please Login to comment
Notifiera vid