Meeting First Socket 939 Processors: AMD Athlon 64 3800+ and Athlon 64 3500+

Today, new processors from AMD come to us in their new form-factor, Socket 939. What benefits does the new dual-channel controller of the Athlon 64 bring to us and what’s so interesting about the Socket 939 platform? Find it out with us!

by Ilya Gavrichenkov
06/01/2004 | 07:59 AM

Today, on the 1st of June 2004, one thing happened that we had long been waiting for. Advanced Micro Devices, Inc. released its new family of Athlon 64 processors, equipped with a dual-channel memory controller. This event is also accompanied with the transition of the Athlon 64 to the new processor socket, called Socket 939, which seems to have the potential of becoming a “stable platform” and live a long life as a solution for high-performance and mainstream computer systems. The rejuvenated Athlon 64 has acquired much more appeal to the end-user. On the other hand, the arrival of Socket 939 processors doesn’t necessarily mean that the currently-available mainboards for this platform are equipped by the latest fashion in technology. For instance, the AMD admirers won’t have a chance of using expansion cards with the new PCI Express interface which will show up after June 21, in response to Intel’s proposed launch of new chipsets with support of this bus. This fact means that the Athlon 64 platform is not yet rock-solid, but is rather still evolving.

 

There are precedents in the past – CPU architectures live and die. First, every processor grows through its childhood, being polished off and perfected. Then, it matures and the platform develops confidently, but the lifecycle anyway ends with a fade-out period when the manufacturer tries to squeeze the last juices out of the obsolete architecture. For example, Intel’s Pentium 4 architecture had its child years in the Willamette core and in Socket 423; it matured in the 0.13-micron Northwood core. Today, this architecture seems to be declining as there appear various senile diseases and it is harder to push the performance bar further.

This “law of CPU evolution” may be equally applied to the Athlon 64. AMD has been adjusting the architecture “while running”, and this is an obvious indication of the Athlon 64’s having been in its infancy up till now. Just consider: in the short period since the launch of the first models of this family (i.e. since September 2003), the Athlon 64 used two different sockets, Socket 940 and Socket 754, and two different cores, ClawHammer and NewCastle, and these two cores went through several steppings. In other words, AMD has been trying to make up its mind as to the most profitable and competitive configuration of the Athlon 64, improving the tech process on the way to reduce the manufacturing costs. Today, we have a hope that this kindergarten is over. The new processors of the Athlon 64 series use the new Socket 939, which is going to have a long lifecycle, and the processors themselves have obtained definite properties which, I hope, will be implemented in CPUs to follow. At least, I hope the Socket 939 platform won’t repeat the fate of the Socket 754 one and live a little bit longer.

Right now, AMD ships three new processors for the new socket: two models of the Athlon 64 family with 3500+ and 3800+ ratings and the Athlon 64 FX-53 targeted at extreme gamers. Today, the company also released the Athlon 64 3700+ for older Socket 754 systems. In this article we will examine the new Socket 939 processors of the Athlon 64 family in more detail. The other newcomers will be covered in our upcoming reviews.

Athlon 64 for Socket 939: What’s New?

AMD times a few minor innovations in the architecture to the launch date of Socket 939 CPUs, but before proceeding to them, I’d like to say a few words about the newly-baked CPU socket first. These are Athlon 64 processors in the new form-factor:


There’re no differences on the lid, save for the marking


Socket 939 CPUs have a different pin configuration than Socket 940 and Socket 754 ones

It is clear that the different pin configuration of Socket 939 makes these processors incompatible with the older Socket 940. Moreover, there are significant differences in the functions of the same legs of Socket 939 and Socket 940 processors. Thus, Socket 939 and Socket 940 and, moreover, Socket 754, are not compatible with each other. The new Socket 939 processors require new mainboards with the appropriate CPU socket. The cooler retention mechanism has remained the same, though, and you can use a cooler for a Socket 940/754 processor in a Socket 939 system.

So, the Athlon 64 processor has grown extra legs, standing now on 939 pins instead of 754. This is because of the dual-channel 128-bit memory controller that came to replace the older single-channel 64-bit one. Unlike in Athlon 64 FX and Opteron processors, used in Socket 940 systems, the memory controller of the Socket 939 Athlon 64 supports ordinary (non-registered) memory modules. This makes the entire platform cheaper (non-registered modules cost less) and faster (registered modules have bigger latencies).

Besides the support of dual-channel memory access, the AMD engineers improved the memory controller of Socket 939 CPUs for better compatibility with different memory modules. They introduced a special 2T DRAM Timing, which lowers the bar the memory controller of the Athlon 64 sets for DDR SDRAM modules. Thanks to that, the Socket 939 Athlon 64 can work with four memory modules in the DDR400 SDRAM mode. You should be aware, though, that the highest performance, with the most aggressive 1T timing, is only achievable when you plug in a couple of identical DDR400 SDRAM modules. When four memory modules are in use, the memory controller of the Athlon 64 can only support DDR400 SDRAM with the slower 2T timing. Moreover, if the four installed modules are two-sided, the speed of the memory subsystem will be dropped to DDR333 even with 2T timing.

We carried out a brief test session to check out the influence of the 2T timing on the performance, using a new Socket 939 Athlon 64 3500+ processor (2.2GHz, 512KB L2 cache). Its memory subsystem worked with DDR400 SDRAM (2-3-2-6), the DRAM Timing being set to 1T and 2T. We left the rest of the memory subsystem settings identical in this test:

1T DRAM Timing

2T DRAM Timing

Sandra 2004, Memory Bandwidth Int, MB/s

5906

4900

Sandra 2004, Memory Bandwidth Float, MB/s

5832

4900

Sciencemark 2.0, Memory Bandwidth, MB/s

5692.29

4525.71

Sciencemark 2.0, Memory Latency, cycles

96

107

PCMark04, Overall score

4525

4438

PCMark04, CPU score

4178

4134

PCMark04, Memory score

5392

4734

Quake3 (four), 1024x768

419.3

403.1

Unreal Tournament 2004 (dm-rankin), 1024x768

112.33

108.52

Far Cry, 1024x768

68.65

68.26

As you see, DRAM Timing is an important parameter, affecting the overall performance of the system. The owners of Socket 939 systems should pay attention to the fact that the use of more than two memory modules may bring about a performance reduction.

The memory controller is not the only thing that changed with the transition to Socket 939. Athlon 64 processors have acquired a faster HyperTransport bus, which can now work at frequencies up to 1GHz, providing a 25% bandwidth growth. Thus, the bandwidth of the HyperTransport bus in Socket 939 systems is now 4GB/s, into each direction. On the other hand, this acceleration of HyperTransport is unlikely to bring any serious profits, because all the busses attached to the chipset don’t sum up to match this gigantic bandwidth. This is confirmed by the results of the mini-test we carried out with an Athlon 64 3500+, using 1GHz and 800MHz HyperTransport.

1000MHz HyperTransport

800MHz HyperTransport

PCMark04, Overall score

4525

4526

PCMark04, CPU score

4178

4173

PCMark04, Memory score

5392

5393

Quake3 (four), 1024x768

419.3

419.2

Unreal Tournament 2004 (dm-rankin), 1024x768

112.33

112.31

Far Cry, 1024x768

68.65

68.56

The two versions of the HyperTransport bus show nearly the same performance. The difference between them fits into the measurement error range. Thus, I can’t say the acceleration of the HyperTransport is in any way profitable to current computer systems. Some advantages may show up on the transition from the AGP 8x to the PCI Express x16 interface. Having dedicated read and write channels, the PCI Express x16 may provide a total bandwidth of up to 8GB/s as HyperTransport bus that works at the 1GHz frequency.

Talking about the HyperTransport bus, I should note that Socket 939 processors, unlike their Socket 940 mates, are not supposed to work in multiprocessor systems, because they have only one HyperTransport link, to the chipset.

The New NewCastle Core

Switching to Socket 939, AMD also transfers the Athlon 64 family to the new processor core, known under the NewCastle codename. Compared to the ordinary ClawHammer, this core has a twice-smaller L2 cache (512KB). The NewCastle core has already been used in some modifications of the Athlon 64 for Socket 754, but now the ClawHammer core (1MB L2 cache) will only be used in expensive Athlon 64 FX family processors. The purpose of this transformation is evidently in making the Athlon 64 cheaper to manufacture.

Really, the ClawHammer core with its 1MB of L2 cache has a die size of about 193 sq. mm. The reduction of the amount of L2 cache memory allows reducing the die size to 144 sq. mm.

By using the Wafer utility written by Rick C. Hodgin, we can estimate the number of dies in one 200mm wafer like those used at the Dresden Fab30 which is now producing Athlon 64 processors.

As you see, one 200mm wafer contains either 144 ClawHammer cores with 1MB L2 cache, or 194 NewCastle cores with 512KB L2 cache. Thus, the use of the NewCastle core allows increasing the yield of processor cores by 34% from one wafer! This means that the manufacturing cost of one NewCastle core is about 25% lower than that of a ClawHammer core. In other words, if the manufacturing cost of a ClawHammer processor is estimated at $100, the new NewCastle-core processor will cost AMD about $75. On the one hand, AMD will increase the profitability of its Athlon 64 production, and on the other hand, this may become a step to an invasion of the Athlon 64 into mainstream and low-end market sectors, because AMD can make cheaper modifications of its 64-bit processor. According to some sources, we may soon see an Athlon 64 with a PR of 2600+, selling for less than $150.

Well, notwithstanding the use of the cheaper NewCastle core in Socket 939 processors, their price will be very high initially. For example, the official price of the Athlon 64 3500+ is $500, while the Athlon 64 3800+ is evaluated at $720. Such high prices are not because of manufacturing costs, but rather because AMD thinks that Intel offers no worthy alternatives today. That’s why the pricing of the new Athlon 64 in the Socket 939 form-factor will become more modest with time, but only after Intel has introduced new and faster processors.

Another fact needs emphasizing: AMD doesn’t seem willing to exert any effort in promoting Socket 939 systems for the time being. At first, the production volume of such processors will be rather small, while junior CPU models won’t come out for the new socket. This situation will only start changing in the fourth quarter when new chipsets and new mainboards will bring PCI Express and DDR500 SDRAM to the Socket 939 platform.

New Processors: Characteristics

I don’t want to be garrulous, so I just give you a table with the formal characteristics of the newly-announced processors: Athlon 64 3500+, Athlon 64 3700+, Athlon 64 3800+ and Athlon 64 FX-53.

Athlon 64 FX-53

Athlon 64 3800+

Athlon 64 3700+

Athlon 64 3500+

L1 Cache Size

64KB data + 64KB instruction = 128KB Total

L2 Cache Size

1MB (exclusive)

512KB (exclusive)

1MB (exclusive)

512KB (exclusive)

CPU Core Frequency

2.4GHz

2.4GHz

2.4GHz

2.2GHz

Memory

Integrated 128-bit wide memory controller

Integrated 128-bit wide memory controller

Integrated 64-bit wide memory controller

Integrated 128-bit wide memory controller

Types of Memory

DDR400, DDR333, DDR266, DDR200 SDRAM

HyperTransport Links

1

HyperTransport Spec

1GHz (x2, DDR)

1GHz (x2, DDR)

800MHz (x2, DDR)

1GHz (x2, DDR)

Packaging

939-pin organic micro-PGA

939-pin organic micro-PGA

754-pin organic micro-PGA

939-pin organic micro-PGA

Process Technology

130nm (.13-micron) Silicon on Insulator (SOI)

Approximate Transistor count

105.9 million

68.5 million

105.9 million

68.5 million

Approximate Die Size

193mm2

144mm2

193mm2

144mm2

Nominal Voltage

1.5 V

Max Thermal Power

89 W

Although the maximum thermal design power (TDP) for Socket 939 processors is set to 89W, like with Socket 754 CPUs, AMD demands a certain reserve from the mainboard makers, with a proposed TDP of 105W. It is expected that the heat dissipation of Socket 939 processors will grow up considerably with the launch of models on the 90nm tech process. Thus, it is quite possible that the future Athlon 64 will be as hot as the notorious Prescott-core Pentium 4.

Now let’s view the info about the Athlon 64 3800+ and the Athlon 64 3500+, which the CPU-Z utility provides:

The Socket 939 processors return a CPUID value which says they are based on the DH7-CG revision of the core. That is, this is the same CG core stepping that we described at length in our Athlon 64 FX-53 review. I was not at all surprised to meet this stepping once again: such cores are only capable of reaching a clock rate of 2.4GHz.

I should confess the Socket 939 modifications of the Athlon 64 and the Athlon 64 FX bring some confusion into the lines of AMD’s 64-bit processors. The market now is full of numerous models, quite different in their characteristics. In order to clear this mess up somewhat, I filled up the following table which lists the basic specs of all the members of the Athlon 64 series:

 

CPU Core

Socket

CPU Frequency

L2 Cache Size

Memory Controller

Athlon 64 FX-53

ClawHammer

Socket 939

2.4GHz

1024KB

Dual-Channel

Athlon 64 FX-53

SledgeHammer

Socket 940

2.4GHz

1024KB

Dual-Channel, Registered memory required

Athlon 64 FX-51

SledgeHammer

Socket 940

2.2GHz

1024KB

Dual-Channel, Registered memory required

Athlon 64 3800+

NewCastle

Socket 939

2.4GHz

512KB

Dual-Channel

Athlon 64 3700+

ClawHammer

Socket 754

2.4GHz

1024KB

Single-Channel

Athlon 64 3500+

NewCastle

Socket 939

2.2GHz

512KB

Dual-Channel

Athlon 64 3400+

ClawHammer

Socket 754

2.2GHz

1024KB

Single-Channel

Athlon 64 3400+

NewCastle

Socket 754

2.4GHz

512KB

Single-Channel

Athlon 64 3200+

ClawHammer

Socket 754

2.0GHz

1024KB

Single-Channel

Athlon 64 3200+

NewCastle

Socket 754

2.2GHz

512KB

Single-Channel

Athlon 64 3000+

NewCastle

Socket 754

2.0GHz

512KB

Single-Channel

Athlon 64 2800+

NewCastle

Socket 754

1.8GHz

512KB

Single-Channel

Testbed and Methods

Today, we will test two processors for the Socket 939 platform: AMD Athlon 64 3800+ and AMD Athlon 3500+. Two other newcomers, the Athlon 64 FX-53 for Socket 939 and the Athlon 64 3700+ for Socket 754 will be discussed in our upcoming reviews. Basing on the price tags on the two processors, we took the following CPU models from the opposite camp to compare them with: a Pentium 4 3.4GHz on the Northwood core and a Pentium 4 3.4E on the Prescott core. Besides that, from the Intel camp again, we took a Pentium 4 Extreme Edition 3.4GHz. A couple of older processors from AMD were also thrown into the heap: an Athlon 64 3400+ for Socket 754 and an Athlon 64 FX-53 for Socket 940.

A short remark about the prices: the Athlon 64 3500+ only fits into the same price niche as the Pentium 4 3.4GHz and 3.4E, while the price tag of the Athlon 64 3800+ ($720) is similar to that of the Athlon 64 FX-53 and has no alternatives on the Intel part, because Intel sells its Extreme Edition of the Pentium 4 at a price much higher than $900.

We used the following hardware in our tests:

We ran our tests in Windows XP SP1 with DirectX 9.0b installed.

Synthetic Tests of the Memory Subsystem

We haven’t yet met a dual-channel memory controller from AMD with support of faster non-registered modules of DDR SDRAM, so we first examine its performance in synthetic tests. We use the ScienceMark 2.0 utility which offers a good toolset for exploring the memory subsystem. First of all we measured the bandwidth and the latency of the memory subsystem in platforms with different Athlon 64 processors (Socket 939 Athlon 64, Socket 940 Athlon 64 FX and Socket 754 Athlon 64). To be able to compare the results correctly, we made all the processors work at 2.2GHz. Moreover, we added the results of Pentium 4-based systems, in which the processor was clocked at 3.4GHz.

Athlon 64 3500+

Athlon 64 3400+

Athlon 64 FX-51

Pentium 4 3.4

Pentium 4 3.4E

Pentium 4 XE 3.4

ScienceMark 2.0, Memory Bandwidth, MB/s

5692

2977

5504

4322

4615

4371

ScienceMark 2.0, Memory Latency, cycles

96

101

112

254

247

254

ScienceMark 2.0, Memory Latency, ns

43.6

45.9

50.9

74.7

72.6

74.7

As the table suggests, the memory controller integrated into Athlon 64 processors shows a miraculous performance. As you know, the AMD64 architecture puts the memory controller into the CPU die and clocks them both at the same frequency. This trick brings such benefits as higher memory bandwidth and lower latencies compared to Pentium 4-based systems.

The memory controller of the Athlon 64 3500+ works with dual-channel non-registered memory and it is really faster than controllers of both Athlon 64 FX-51 and Athlon 64 3400+. In comparison with the first of these two CPUs, the Athlon 64 3500+ wins due to the use of non-registered memory. In the second comparison, it wins due to the dual-channel architecture with the bank interleave feature. Well, I should remind you that the memory controller of the new Athlon 64 for Socket 939 can only show such high results if you use the 1T timing.

To confirm the highest performance of the memory controller of the new AMD processors, I’d like to show you the numbers I got in the memory test of SiSoftware Sandra 2004:

Athlon 64 3500+

Athlon 64 3400+

Athlon 64 FX-51

Pentium 4 3.4

Pentium 4 3.4E

Pentium 4 XE 3.4

Sandra 2004, Memory Bandwidth Int, MB/s

5906

3069

5601

4966

5003

4959

Sandra 2004, Memory Bandwidth Float, MB/s

5832

3067

5542

4970

5002

4956

Really, Sandra 2004 says the practical bandwidth of the memory subsystem of the Athlon 64 3500+ is 92% of the theoretical maximum and that’s very good. For example, the i875P-based system enjoys only 78% of the theoretical bandwidth. In fact, only Athlon 64 CPUs with a single-channel memory controller can boast a higher efficiency, reaching to 96% of the theoretical memory bandwidth.

Performance

Before proceeding to the tests, let me remind you once again that the new Socket 939 processors from AMD differ from their older Socket 754 mates in two respects: they feature a dual-channel memory controller, but have a smaller amount of L2 cache memory (512KB instead of 1024KB). At the same time, AMD, the manufacturing company, states that Socket 939 CPUs are better than Socket 754 CPUs of the same frequency and that’s why they come with a 100+ higher performance rating in their names. So, AMD reasons that the dual-channel memory controller contributes to performance more than a larger L2 cache.

Gaming applications

Gaming applications come first, because a majority of users who buy top-end CPUs for desktop systems have one thing in mind – gaming! – and other things just follow.

The use of a dual-channel memory controller does help the Socket 939 Athlon 64 3500+ processor to be slightly faster than its Socket 754 Athlon 64 3400+ mate, which works at the same frequency. The Athlon 64 3800+ is also better in this test than the Athlon 64 FX-53, which works at the same frequency, but with registered memory. So far, we meet no surprises. The Athlon 64 CPUs outperform the top-end Pentium 4 models in Quake 3, but lose to the Pentium 4 Extreme Edition with its 2MB L3 cache.

The Athlon 64 processors have notched considerably better results in Unreal Tournament 2004 than their competitors from Intel. As for the relations among Socket 939, Socket 940 and Socket 754 models, we see processors of the same clock rate to be on about the same performance level.

The new processors from AMD take the top positions in Aquamark3. The Athlon 64 3500+ even outperforms the Athlon 64 FX-53 that works at a higher clock rate. The latter CPU has a dual-channel memory controller, but its memory subsystem is slower due to the use of registered DDR SDRAM modules. Evidently, high memory bandwidth matters much in this test.

The CPU test from this benchmark produces another picture. Thanks to their Hyper-Threading technology, the processors of the Pentium 4 family are strong enough: the both Pentiums 4 with 3.4GHz clock rate find themselves ahead of the Athlon 64 3400+ and also ahead of the Athlon 64 3500+. Anyway, the faster Athlon 64 3800+ (2.4GHz clock rate) only loses to the Pentium 4 Extreme Edition.

The Athlon 64 processors of all the modifications have no rivals in Halo. The Pentium 4 and its Extreme Edition cannot reach their level of performance. As for the standings inside the Athlon 64 team, it is all quite predictable. The Athlon 64 3500+ with a dual-channel memory controller and 512KB L2 cache is faster than the Athlon 64 3400+ with a single-channel memory controller and 1MB cache. The Athlon 64 3800+ outruns the Athlon FX-53 with a twice larger L2 cache (1MB), but a less fast memory subsystem.

The AMD team is victorious in Far Cry. Note a curious fact: the Socket 939 processors with their most efficient memory controller provide a substantial fps rate gain in this game.

Overall, the Athlon 64 processors look preferable to Intel’s CPUs in gaming applications. This is especially true about the new Socket 939 CPUs whose improved memory subsystem brings about a nice performance gain in games. By the results of the gaming tests, we may say that the Athlon 64 3500+ and the Athlon 64 3800+ are true to their ratings. For example, the Athlon 64 3500+ beat the Athlon 64 3400+ in all the tests, although the latter has the same clock rate and a larger cache!

Office and digital content-creation applications

Traditionally, this section of our review is dedicated to the results of the Winstone tests.

The situation is practically the same in both test suites of the Winstone family. The Athlon 64 processors take the lead, followed by top-end Pentium 4 models. Note, however, that the L2 cache size is most important in this test: we see the new Athlon 64 3500+ losing to the Athlon 64 3400+, while the Athlon 64 3800+ is slower than the Socket 940 Athlon 64 FX-53.

Data Compression and Media Encoding

The processors of the Athlon 64 family are much faster than all the Pentium 4s in the test of data compression with the popular WinRAR 3.3 utility. You may have noticed the things that matter in this test: a large L2 cache and an efficient memory subsystem with low latencies. At the same time, the faster memory controller of the Socket 939 processors doesn’t save them from defeat they take from their older mates with a larger L2 cache. According to this benchmark, the Athlon 64 3400+ outperforms the Athlon 64 3500+ and also the Athlon 3800+, which works at a higher clock rate. Thus, the use of a dual-channel memory subsystem in the new CPUs from AMD doesn’t make up for the reduced L2 cache in data compression tasks.

The size of the cache and the speed of the memory subsystem have a small impact on the speed of encoding MP3 files. The clock rate and the efficient micro-architecture are important here. That’s why the Athlon 64 3500+ and the Athlon 64 3400+ as well as the Athlon 64 3800+ and the Athlon 64 FX-53 match each other in this test. Note also that the top-end Athlon 64 CPUs lose in this test to the Pentium 4 processors on the 0.13-micron Northwood core.

The Pentium 4 is traditionally good at encoding video into the MPEG-2 format. The Athlon 64s, even with their new 128-bit memory controller, cannot catch Intel’s CPUs. By the way, the memory subsystem speed and the amount of the L2 cache is of little effect in Mainconcept MPEG Encoder. That is, the two pairs – Athlon 64 3500+ with Athlon 64 3400, and Athlon 64 3800+ with Athlon 64 FX-53 – go neck and neck in this test.

We used Xmpeg for encoding video into the MPEG-4 format. The performance of the Athlon 64 processors complies with their ratings, while the eldest, the Athlon 64 3800+, is even faster than the fastest Pentium 4 CPUs from Intel.

The NetBurst architecture and the Pentium 4 processor win the test of encoding video in Windows Media Encoder. The Athlon 64s lack frequency to compete here. The new dual-channel memory controller in the Athlon 64 3500+ and in the Athlon 64 3800+ has increased their performance in comparison to older Athlon 64s, but this is not enough for an efficient work in Windows Media Encoder.

Overall, it is impossible to name a processor architecture, most suitable for data encoding. Depending on the tools and formats you use, the results may vary in favor of the Athlon 64 or the Pentium 4. Anyway, I should note that the Pentium 4 is often faster at encoding streaming multimedia data than the Athlon 64.

Adobe Photoshop

I guess Adobe Photoshop CS 8.0, the popular image-processing software suite, needs no introductions. We paid special attention to tests in this program. For our tests, we used a slightly modified PSBench 7 with a 100MB image.

The final performance rating is the geometric mean of the times it took to perform various typical operations. In the diagram below, the results are given in seconds. That is, a smaller number means higher speed.

And the following table shows the speeds of applying various filters of Photoshop CS 8.0 on the tested systems. The results are given in seconds:

Athlon 64 3400+

Athlon 64 3500+

Athlon 64 3800+

Athlon 64 FX-53

Pentium 4 3.4

Pentium 4 3.4E

Pentium 4 XE 3.4

Rotate 90

0.2

0.2

0.2

0.2

0.2

0.1

0.2

Rotate 9

3.8

3.6

3.4

3.4

3.8

3

3.8

Rotate .9

3.5

3.4

3.1

3.2

3.5

2.8

3.5

Gaussian Blur 1

1.1

1.1

1

1

1.2

1

1.2

Gaussian Blur 3.7

3.5

3.4

3.1

3.3

2.7

2.3

2.7

Gaussian Blur 85

3.9

3.8

3.5

3.6

2.9

3.2

2.8

Unsharp 50/1/0

1.9

1.8

1.6

1.7

1.9

1.5

1.9

Unsharp 50/3.7/0

4.4

4.2

3.9

4

4.1

2.8

4.1

Unsharp 50/10/5

4.4

4.2

3.9

4

3.3

2.9

3.3

Despeckle

5.3

5.4

5

4.8

3.1

3.4

3

RGB-CMYK

11.1

11

10.1

10.1

6.9

7

6.9

Reduce Size 60%

1

1

0.9

0.9

0.7

0.7

0.7

Lens Flare

7.9

7.8

7.1

7.3

6.1

6.2

6.1

Color Halftone

10.8

10.7

9.9

9.8

14.5

13.3

13.6

NTSC Colors

3.1

3

2.8

2.8

3.2

3.4

3.1

Accented Edges

20.4

21.2

19.5

19.1

22.3

27.8

21.7

Pointillize

41

40.8

37.2

37.5

32.7

35.2

32

Water Color

43.5

44.2

40.6

40

43.3

57.8

43.1

Polar Coordinates

14.4

14.5

13.3

13.3

10.6

11.4

10.3

Radial Blur

71

72.9

68.5

68.7

60.4

48.5

60.4

Lighting Effects

4

4

3.6

3.7

3.4

3.3

3.4

The Prescott-core processor from Intel took more first places than any other CPU. Although the use of the dual-channel memory controller helps the Athlon 64 in the Socket 939 form-factor to show a higher performance, it is still not enough to shine brightly in Photoshop.

3D rendering

This section of our testing session is dedicated to final rendering as performed in the popular 3ds max 6 suite.

The Pentium 4 models endowed with Hyper-Threading technology are the best at rendering one frame. This technology helped the Intel processors to speed up exactly in such types of tasks. The Athlon 64, regrettably, cannot boast any technologies and its results are somewhat worse. Note also that the amount of L2 cache memory and the speed of the memory subsystem don’t affect the result seriously. This explains why the Athlon 64 3500+ and the Athlon 64 3400+, as well as the Athlon 3800+ and Athlon 64 FX-53 run neck and neck in this tests.

When we go over from rendering one frame to rendering a frame sequence, we get quite difference numbers. We rendered 30 initial frames of the Ape.max animation, coming with 3ds max. Again, the amount of L2 cache memory and the speed of the memory subsystem don’t affect the result much, but all the Athlon 64 family processors feel much more confident here. For example, the new Athlon 64 3800+ is faster than the top-end Pentium 4 models at rendering this animation.

Scientific calculations

We used the popular Mathematica 5.0 program for estimating the system performance in scientific applications. This program is extensively used for numerical and symbolic calculations.

It is a trivial thing to say that computational load is among the fortes of the Athlon 64, and this test gives another confirmation. The Athlon 64 is really far ahead of the Pentium 4 processors. At the same time, the reduction of the L2 cache, which happened to the Athlon 64 on the transition to the Socket 939 platform, negatively reflected on the performance. The new Athlon 64s on the NewCastle core lag behind their predecessors on the ClawHammer core that work at the same frequency.

Software Development Tools

Besides our traditional tests, today we run one more benchmark, which we tried in our previous processor reviews. Namely, we measured the speed of compilation of projects in Visual C++ .NET, a popular software development environment. For our measurements, we took the source code of the Emule client, to which we added the source code of several libraries, necessary on the compilation stage: crypto51, CxImage, zlibstat. We measured the compilation time in two modes: “Debug” (this version included debugging information into the code) and “Release” (building the final product, optimized for the execution speed and the amount of the resulting code).

The two following diagrams show the results in seconds, so a smaller number is better.

Processors from AMD can be honestly recommended to scientists as well as to software developers. By the way, you may note that the processors with higher amount of cache memory are faster at compiling software projects.

Overclocking and Temperature

We didn’t have high hopes about the overclockability of the new Socket 939 processors. As I mentioned above, they are all based on the same CG core revision as the Socket 940 Athlon 64 FX-53 we had tested earlier. Without extreme cooling methods, we had only reached a little over 2.6GHz clock rate for that processor. Considering also that the Athlon 64 3800+ is the last CPU model on the 130nm core, I may venture a supposition that Socket 939 Athlon 64 models should speed up to 2.5-2.6GHz and no more, without any special cooling.

Moreover, we should keep in mind the fact that the memory controller of the Athlon 64 is situated in the same chip as the CPU itself, and making a more sophisticated controller inevitably leads to a higher heat dissipation of the processor at large. This means that Socket 939 Athlon 64 processors with a dual-channel memory controller will overclock no better than their Socket 754 counterparts based on the CG core stepping.

Before proceeding to the results we achieved at overclocking our Athlon 64 3800+ and 3500+, we should note the fact that these processors, like the Athlon 64 in the Socket 754 form-factor, have their frequency multiplier locked from above. The new Athlon 64 FX-53 in the Socket 939 form-factor, however, doesn’t have any multiplier lock, as this processor is targeted at computer enthusiasts. Thus, we had to speed up our Athlon 64 3800+ and 3500+ samples by increasing the frequency of the clock generator. Fortunately, the modern chipsets from VIA and NVIDIA (K8T800 Pro and nForce3 250, respectively) employed in mainboards for Socket 939 processors can clock the AGP/PCI busses asynchronously. That is, you may leave these busses at their normal frequencies during overclocking. This way, even if you overclock a CPU with a locked multiplier, you won’t render your peripheral devices non-operational.

We tried to overclock our processors, an Athlon 64 3500+ and 3800+, on an ASUS A8V Deluxe mainboard, using a Thermaltake Silent Boost K8 (A1838) cooler. For better overclocking, we also increased the Vcore by 10%, from the nominal 1.5v to 1.65v.

The maximum frequencies at which the processors kept stable were rather low, as we had expected. Particularly, the Athlon 64 3500+ with 2.2GHz frequency overclocked to 2.55GHz, where we arrived by increasing the FSB frequency to 232MHz. The other processor we tested today, the Athlon 64 3800+, sped up from 2.4GHz to 2.58GHz, the FSB clock rate being 215MHz. Thus, the frequency reserve of Socket 939 processors turned to be somewhat poorer than that of Socket 940 Athlon 64 FX-53 processors. That’s natural, though.

The temperature is an important factor with regard to the reviewed processors, because their integrated memory controller contributes to their heat generation. Of course, the Athlon 64 3500+ and 3800+ models are based on the NewCastle core with only 512KB of L2 cache. Anyway, this fact shouldn’t compensate for the memory controller’s heat, because the transistors of the cache memory don’t add much to the die’s heat dissipation.

To clear up this situation, we measured the temperatures of the reviewed processors under a load and when they were idle. We made our measurements using the thermo diode integrated into the CPU core and reading its data with the hardware monitoring tools of the mainboards.

Athlon 64 3400+

Athlon 64 3500+

Athlon 64 FX-53

Athlon 64 3800+

Temperature, Idle, C

40

38

38

41

Temperature, Burn, C

60

58

60

64

The Socket 939 Athlon 64 3500+ turns to be colder than its Socket 754 counterpart, the Athlon 64 3400+. It seems strange at first sight, but there’s no mystery after all. The Athlon 64 3400+ we used in our tests was based on the earlier C0 core stepping, which had a lower frequency potential and higher heat dissipation. If we compare the Athlon 64 FX-53 and the Athlon 64 3800+, we may come to the conclusion that the dual-channel memory controller in the Socket 939 CPU generates a lot of heat.

By the way, all Socket 939 processors, including Athlon 64 3800+ and 3500+, and also the Athlon 64 FX-53, feature Cool’n’Quiet technology for reducing the temperature and heat dissipation under small workloads.

Conclusion


New processors from AMD for the new socket, in a new box

Launching the new Athlon 64 processors in the Socket 939 form-factor, AMD put the AMD64 architecture on a new level. Starting selling these CPUs, AMD solves several problems at a stroke:

First, Socket 939 becomes a “stable platform” with a lifecycle stretching to 2006. Thus, AMD makes a step towards end-users who want to have low-cost upgrade opportunities.

Second, the new processor socket offers dual-channel memory access to the owners of the Socket 939 platform. I can’t say that the two channels give the Athlon 64 a great advantage in speed (the performance gain from enabling the second memory channel is 3-5% in average). Well, no one promised any performance breakthroughs from the transition to Socket 939, but the improvements in the memory controller allow users to flexibly configure the memory subsystem and use four two-sided DIMM modules in their systems, while Socket 754 processors only supported two two-sided memory modules.

Third, AMD achieves a 25% reduction of the manufacturing cost of Socket 939 processors by cutting their L2 cache in two. This move will bring in profits and will also allow manufacturing cheap Athlon 64 models to ensure their popularity in the market.

The performance level of today’s Athlon 64 processors is astonishing. I won’t claim that top-end Athlon 64s are always faster than top-end Pentium 4s, as the NetBurst architecture is very good in some applications. However, across a majority of tasks, including games, the Athlon 64 provides an excellent performance.

Thus, the new Athlon 64 models in the Socket 939 form-factor may become a good choice for a high-performance computer system. Moreover, the AMD64 architecture has its main trump up the sleeve yet – support of 64-bit applications. At the same time, it’s early yet for AMD to claim victory. First, the next models of the Athlon 64 with a higher frequency won’t come out soon, as they need 90nm tech process to be manufactured. Second, we haven’t yet seen Intel’s processors with x86-64 architecture. Third, Intel is going to introduce its new platform, the i925/i915 chipsets and new faster CPUs on the Prescott core. All this may shatter AMD’s now-firm positions in the market. Let’s not anticipate, though.