Memory divider, Socket, fixations and chipset

Divider & memory frequency

If the integrated memory controller is beneficial for latency and performances, it also has its downsides. The release of DDR2 is also the occasion to address an often forgotten problem. Memory frequency is obtained from processor frequency, to which a whole number divider is applied.

With a 200 MHz memory frequency and processors whose frequencies were multiples of this value, like with the Socket 754 and 939 and DDR400, this restriction wasn’t particularly problematic unless you wanted to use DDR266 or DDR333. With the release of DDR2 and its varied frequencies, there are many cases in which the memory simply won’t run at full speed. For example, at 2.6 GHz, if you divide by 6 you get 433 MHz, which is too much for DDR2-800. The coefficient of 7 is applied, or in other words, a speed of 371 MHz.

Here is the frequency table:


As you may have noticed, each frequency has disadvantages. DDR2-800 will only be fully exploited at 2.8, 2.4, 2.0 and 1.6 GHz and the DDR2-667 only at 2.0 GHz. Of course 371 MHz instead of 400 MHz isn’t too dramatic, but if AMD could have introduced half whole number dividers (6.5 here) we would have been closer to the “real” frequency.

Modified socket and CPU cooler fixations

The Socket AM2 is physically different from the Socket 939, first of all, because of the number of pins, which increases from 939 to 940. The position of the pins is modified, so it’s impossible to fit a Socket 939 processor on AM2 and vice-versa (unless you push really hard).


It isn’t the only platform modification, because the fixation system is also different. The previous one used a 3 point fixation per side while now it has been reduced to one point. This isn’t a real problem in itself, however, because many older coolers work with the new system and vice-versa.


There are systematic problems for CPU coolers that don’t use this system and which require a direct fixation on the motherboard. Indeed AMD’s retention system is no longer fixed on the motherboard by two screws but rather four, which makes that many high end fans incompatible without an AM2 kit. We wonder if this was really that useful as we have never heard of a ripped off retention system.

... no chipset modifications !

If changing your memory means changing processor and motherboard, this isn’t case for the chipset. It’s still connected to the processor via a 1 GHz HyperTransport bus. NVIDIA, ATI, VIA, ULi and SiS current product lines are compatible. NVIDIA decided to release a new chipset, the nForce 5, with the AM2, but the former will also work with Socket 939 or 754. The nForce 4 won’t be a problem on AM2 too.

AMD's product line on AM2

AMD’s product line on AM2

Up until now, we found the following Socket 939 (Athlon) and 754 (Sempron) processors:


The product line released by AMD on Socket AM2 is:


There are no major modifications for the Athlon. Of course, there’s the release of the FX-62 at the same price as the FX-60 Socket 939 despite a higher frequency, and the Athlon 64 mono core with 1MB of L2 cache isn’t on the Socket AM2. We noted the release of two new X2, the 5000+ and 4000+.

There is a general increase of the Sempron P-Rating by 200 with the modification from Socket 754 to AM2 and of single channel DDR to dual channel DDR2. We found a couple of surprising modifications. For example, the modification from 1.6 GHz to 1.8 GHz increases the P-Rating by 400, while the modification from 1.8 GHz to 2.0 GHz increases it by 300 with 128 KB of cache, and by 200 with 256 KB of cache. We are still looking for a logical explanation.

In addition to the standard version, AMD will release an « Energy Efficient » model of these processors. With Sempron, voltage will be decreased from 1.35-1.4V to 1.2-1.25V and the TDP from 62W to 35W on the « EE » 3400+, 3200+ and 3000+. This has an increased cost as they are at $145, $119 and $101, respectively, as compared to $97, $87 and $77 for standard versions.

With the Athlon 64, the voltage and TDP reduction is the same for the 3500+ at $231, as compared to $185 for the standard version. Finally, for dual core there is a single « EE » model at 32 watts and 1.025-1.075V. It’s the 3800+ priced at $364 as compared to $303 for the standard version. All other X2s, except for the 5000+ are available with a TDP of 65W in 1.2-1.25V, as compared to 89W in 1.3-1.35V: the 4800+, 4600+, 4400+, 4200+, 4000+ and 3800+ are $26, $43, $44, $52, $25 and $20 more expensive than the standard versions.

To release this type of processor, AMD uses a die that requires less voltage to accomplish their tasks. According to the quality of the die produced by AMD, it’s possible that some of the “standard” processors could be under power levels supported by “Energy Efficient” models. Of course, only this label goes with the warranty and AMD makes you pay for it.

Processors and test platforms

Test processors

For this test, we selected three Socket AM2 processors:

- Athlon 64 FX-62
- Athlon 64 X2 3800+
- Sempron 3600+


In tests we reduced the Athlon 64 FX and Sempron multiplying coefficient to measure the Athlon 64 X2 4800+/4400+/4000+ and the Sempron 3400+ and 3000+ performances.

We could have expected this because of the TDP announced at 125 watts. Even if AMD always overestimates TDP, the Athlon 64 FX-62 begins to reach power consumption levels usually associated for Intel and not AMD. Of course, it isn’t at the same level as the Pentium D, but overall power consumption of the AM2 configuration in load under Prime95 perfectly shows this (2x512 MB of DDR2 phenomenon, with Radeon X700, and Raptor in addition to the motherboard and the processor):

- Sempron 3600+: 137 Watts
- Athlon 64 X2 3800+: 163 Watts
- Athlon 64 FX-62: 221 Watts

The difference in power is 84 watts with the Sempron and 58 with the Athlon 64 X2. Of course, these losses are recorded from the PC power supply and the power stage of the motherboard, but these figures are rather high. In full load, AMD’s box should reach 4,000 Rpm to cool down the computer and at this speed it’s rather noisy. We clearly prefer something bigger with a 120 mm fan on top.

Test platform

For this test, we selected ASUSTeK’s M2N32-SLI Deluxe. Based on NVIDIA’s latest nForce 590 SLI, it embeds all functionalities, and amongst other things, 6 Serial ATA and two Gigabits networks ports. There is HD audio with a soundMAX AD1988B codec and ASUSTeK has also additionally included a Texas Instruments PCI FireWire 400 chip, a Silicon Image SiL3132 for 2 additional SATA including one external, and a WiFi 802.11a/b/g Realtek RTL8187L controller, which uses a USB hub. Four DIMM for the DDR2, two PCI Express x16, one PCI Express x4, one PCI Express x1 and two PCI are also included.


This impressive density of functions is a little problematic since many SLI motherboards that use dual slot cards will block the PCI-E x4 port and one PCI port. We noted that two out of the six SATA natively supported by the nForce 590 also have difficult access in this type of configuration. To make a little more room, ASUSTeK placed the WiFi on some sort of mini daughterboard. The M2N32-SLI Deluxe, cooled down by a passive cooling system based on heatpipes is available at 225€.

For memory, we have 2 x 512 MB XMS-2 8500 Corsairs, DDR2-1066 in 5-5-5-15. Because the AM2 only initially supports DDR2-800 (which isn’t bad), this type of memory module has no real interest except in overclocking. It did make it possible for us to test a wide range of frequencies and timings.

DDR2 in practice

DDR2 in practice

What does DDR bring in practice compared to DDR2 for AMD? This is what we have wanted to verify from the start by comparing DDR-400 and DDR2-800 with a CPU frequency of 2.8 GHz (Athlon FX-62 on AM2 and FX-60 at 62 on 939) for the DDR2-800 to be really at 400 MHz (see page 2).

To do this, we measured performances in four domains. We first focused on a reading bandwidth test and random reading latency test with RightMark Memory Analyzer. These results are respectively expressed in MB /s and in number of cycles. Finally, two applicative tests complete the above, with WinRAR and Far Cry, which are particularly dependent on the speed of the memory sub system.


For reading there is a relatively interesting increase of 20% and this is approximately the same percentage as the performance loss for latency in this test. In the end, practical results aren’t really surprising. We expected something better. WinRAR remains in the lead with DDR-400 and with Far Cry, the DDR2 improvement is only 2.5%.

Here are other figures, this time obtained at 2.6 GHz. Note that DDR2-800 is only at 93% of its potential, while DDR2-533 and 667 are at 98% compared to 96 and 93% at 2.8 GHz.


Results are rather dramatic. If we take an entry level DDR-400 memory module at 3-3-3-8, we see that with WinRAR you have to select the highest end DDR2 to be faster. With Far Cry, you have to use DDR2-667 in 3-3-3-8. The most worrying aspect, however, is the performance gap between entry level and high end products, where the performance loss is 9% with WinRAR and 7% with Far Cry in DDR and 17 and 13% in DDR2. We’ll let you guess, which memory will equip computers sold in superstores…

The next question was, which memory should we use for the rest of the test? Up until now we were using DDR400 2-2-2 on a DDR platform and DDR2-667 4-4-4 on a DDR2 platform. We decided to continue this and also test the FX in DDR2-800 at 4-4-4. You might criticize this choice but we felt that it wouldn’t be appropriate to test all AM2 processors with an expensive DDR2-800.

Here is our test protocol. Because of the enormous amount of results (50 configurations), we will separate each test into two parts. We made comments on the first and it focuses on the AM2 compared to the Socket 754/939. The second is just a graph with all the results.

3d studio max 7

3d studio max 7

For 3D-studio max we used a rendering via the 3Ds internal engine (scanline). Developed by Studio PC this scene mainly uses radiosity. The result is more realistic in terms of lighting and is also slower. 80% of this scene is based on this type of effect.


Here, Socket AM2, 939 and 754 are relatively close, because even if we note some gaps, this type of load isn’t really hungry in terms of memory speed. Therefore, we mainly see the inadequacy of AMD’s P-Rating in some cases since the AM2 Sempron 3600+ isn’t faster than the Socket 754 Sempron 3400+.

Maya 6

Maya 6

We used a scene developed by Yann Dupont of 3DVF (whom we thank for its use) rendered via Mental Ray.


Like with 3ds max, the influence on memory bandwidth or latency is minimal and doesn’t make it possible to create significant differences in any way.

Mathematica 5.2

Mathematica 5.2

The following tests are scientific calculation programs, starting with Mathematica 5 from Wolfram Research and the test suite developed by Stefan Steinhaus.


This time there are performance gaps in favour of DDR2, a rather good surprise compared to preliminary results. Of course, the improvement is only 3% with DDR2-800, but this is still better than nothing. We noted that despite the modification to dual channels, the progression of the Sempron AM2 compared their Socket 754 homologue isn’t that obvious.

WinRAR 3.51

WinRAR 3.51

A total of 588 MB of 493 Word and Excel files (69 MB), 22 Eudora e-mail box (251 MB) and one audio wav format (268 MB) file were compressed using the most advanced option via WinRAR 3.5.


Preliminary results already showed the catastrophic performances with WinRAR. Performances are passable in DDR2-800 and terrible in DDR2-667. This is the reason why the X2 3800+ S939 is a little faster here than an AM2 X2 4400+! Also, the performance gain due to the Sempron dual channel is more than diluted in the losses due to the DDR2. An AM2 Sempron 3600+ is required to beat a Sempron 2800+ S754!

TMPGEnc 3.3 Xpress

TMPGEnc 3.3 Xpress

With TMPGEnc 3.3.3.7, we encoded a 10 minute 16 second DV file to MPEG-2 format in 720x576 with an average bitrate of 4500 Kbits and in two paths. The video preview display is activated during this test and the DV file is decoded via a Mainconcept codec, which is faster than the decoder in TMPGEnc.


The AM2 is presented here in a more favourable light. If the DDR2-667 vs DDR-400 duel goes in favour of the latter, DDR2-800 brings a slight performance gain (2%). Also, the improvement of dual channels is more significant than the loss due to DDR2. This is the reason why the Sempron AM2 is faster than their S754 homologue at equivalent cache and frequency. Unfortunately, the P-Rating developed by AMD doesn’t go in their favour.

DiVX 6.1 & VirtualDub 1.6.11

DiVX 6.1 & VirtualDub 1.6.11

We compressed the same video as in TMPGEnc in Fast recompress mode, via DiVX 6.1 CODEC, in one path with an average bitrate of 1500 Kbits /s, b-frame and with best quality encoding performances. The video preview display is activated during this test.


This time, DDR2-800 only allows the AM2 to come back to the same level as the S939, whereas performance losses due to the DDR2-667 are more significant because of the dual channel for the Sempron, which means that the latter are faster with the Socket 754.

Far Cry

Far Cry

Here are results with Far Cry. The scene used was outdoors in map training.


Game performances are always quite influenced by memory speed, particularly latency. This is the reason why this is type of application benefited most from the integration of the memory controller to the Athlon 64 when it was released in 2003. The impact of the AM2 is very significant, because DDR2-800 makes it possible to come back to the same level as the DDR-400 on Socket 939. In DDR2-667 it is much slower and without it, it’s dramatic. With an AM2 4400+ performances are at the same level as the 4200+ S939. Once more, the Sempron P-Rating isn’t justified and the change to single channel DDR-400 2-2-2 to dual channel DDR2-667 4-4-4 reduces performances.

We might think that with this type of platform we will more easily find DDR-400 in 3-3-3, but it’s important to point out that DDR2-667 will also often be replaced by DDR2-533. Things are equal, let’s see what happens now.

Pacific Fighters

Pacific Fighters

With Pacific Fighters, results are not as good since DDR is always in the lead. The Sempron AM2 P-Rating needs to be reconsidered based on practical performances recorded.

Conclusion

Conclusion

When DDR2 was added to Intel’s platform in June 2004, we really weren’t convinced because gains were non existent. At the time however it was DDR2-533 and one of AMD’s reasons not to change to DDR2 was that DDR2 had to mature in terms of pricing and performance.

Now that AMD made the change, we understand why the father of the Athlon wasn’t in a rush. Indeed, latencies due to DDR2 technology and its implementation by AMD are extremely penalizing for K8 architecture and the integrated memory controller. In the end, only DDR2-800 makes it possible to reach performances comparable to the DDR-400’s! The change doesn’t seem to be advantageous while waiting for a more efficient implementation.


Of course, manufacturers are more often choosing DDR2 as DDR begins to be more and more expensive. It is important to compare similar things, however, and if you have to buy DDR2-800 with an equivalent level of performance, the price advantage is no longer in a favour of the “new generation” memory.

This situation is even more problematic when you take a look at Sempron processors. Indeed, they now have a joint Socket to the Athlon 64, which is a good thing even if it would have been simpler to release them on Socket 939 months ago. The P-Rating chosen by AMD is completely unrealistic, however. Gains due to dual channels are completely non existent (or even more) by inherent latencies to the DDR2 and the +200 P-Rating for the Sempron isn’t justified unless the comparison is made between a Sempron Socket 754 with PC-2700 and a Sempron Socket AM2 with PC2-6400 (sic).

In the end, the release of the Socket AM2 leaves us more than perplexed. We are still looking for its interest for the final user (except for being the only platform which could have been made with the Socket 939). The release of « Energy Efficient » processors with reduced TDP is a good thing, but here again it could have been made without a new Socket. A few weeks from the release of the Core 2 Duo, is the AM2 what AMD and the Athlon 64 needed? Probably not…