First processors of the AMD64 architecture appeared yet in April this year. It was the time when AMD launched its Opteron 200 server models. They could be used in uni and dual-processor configurations. Unfortunately, the original clock speeds weren't that high (1.4..1.8 GHz), but thanks to the unique architecture the Opteron scored good results. By autumn the row was extended at the expense of new clock speeds and new series. Today AMD offers three series for single (series 100), dual (200) and four or eight-processor (800) systems. The maximum clock speed for the Opteron CPU is 2 GHz (xx6 models).
But the market was waiting for new and inexpensive solutions so that everyone could afford it. We were excited with rumors about speeds, socket, L2 cache size and even new names of desktop processors. Finally, in September AMD unveiled its market penetration plans.
So, on September 23 AMD officially presented its new models for desktop PCs:
Besides, they announced the processors for notebook PCs (DTR) rated 3000+ and 3200+, but the only difference from the Athlon 64 is that they don't have a cap on the die, that is why we will leave them aside. Just remember that the Cool‘n'Quiet (clock speed and voltage dynamic management technology) can be used with all processors of the AMD64 architecture. It depends only on whether a given motherboard supports it. The Mobile Athlon 64 can be used only in DTR systems as they consume up to 89 W. For the Opteron it makes 85 W for low-level models and 89 W for 2.0 GHz and over (it also relates to the Athlon 64/Athlon 64 FX). That is why for all processors of the AMD64 architecture the power consumed is defined by the clock speed.
Before we start our examination I'll recommend you to look through our previous materials on the AMD64 architecture:
Since the Opteron processors were much spoken about earlier, we'll examine new products by describing their difference since the cores are almost identical.
The Athlon 64 uses the Socket 754 and has a single-channel integrated memory controller with the DDR400 support (not registered!). It replaces the Athlon XP which will be gradually leaving the market. Although its performance index equals the maximum one of the predecessor (and the clock speed is even lower), I hope that the considerable changes in the architecture will let the Athlon XP 3200+ work faster.
The Athlon 64 FX has fewer differences - at the moment of announcement it differed from the Opteron only in the clock rate which was 2.2 GHz for the FX-51. Formally, the DDR400 memory support also differs, but it's actually not accounted for :). This processor is positioned as a high-end desktop model. But taking into account that it's fully interchangeable with the Opteron (in single-processor systems), such positioning can be easily ignored by smart users :).
Although both sockets have the same distance between the pins equal to 1.27 mm, the Socket 754 doesn't belong to the Socket 940 subset as its pins make a square of 29 x 29 against 31 x 31 of the Socket 940. And in contrast to the famous pair of i865/i875 and i848 the manufacturers will have to develop a different PCB design for these products.
But both sockets use the same cooler mounting system.
The base the cooler is mounted on consists of two parts: a metallic mount and a plastic frame which are located on different sides of the mainboard and fastened with two screws. The cooler is attached to the frame with two strong clips.
The coolers we used had a copper base and copper fins. The design is identical to the well known models Thermaltake Volcano 7+/11+. By the way, judging by the number of trade marks on different parts of the box cooler we can assume that exactly this company helped AMD to develop coolers for the new processors. The size of the different models differed a little. The box version of the Opteron 240 (which works flawlessly with faster processors including Opteron 146) has the base of 55x75x5 mm and 46 pins of 12 sq cm. The Delta fan of 70x70x15 mm of the AFB0712HBB has a built-in temperature sensor for rotation control (the maximum is 4300 RPM). Thermaltake's version had different parameters: the base of 65x60x4 and 36 fins of18 sq cm, the same fan but without a temeprature sensor. Apart from copper versions there was an aluminum one with a copper cylinder inside. The Zalman CNPS7000-Cu can also be installed (but it attaches with screws, which is inconvenient for frequent replacements).
The cooler is meant to blow on the memory chips located near the processor, but one of the versions used had fins located along the long socket side and therefore it can't be used this way (at least, with the boards tested).
All the coolers are quiet (for the Delta the specified noise level is 38.5 dBA at the maximum rpm). That is why AMD's solutions have it all right although the number of transistors is twice more compared to the Athlon XP (105.9 M against 54.3).
Below you can see a summery table of the old and new processors for desktop
PCs. The Opteron is taken in there mostly for comparison with the Athlon 64 FX.
But the price for models of the the 100 series is not that high - $250 and over.
* depends on the chipsetIn spite of the official data there is a little point to be mentioned - the Opteron does support the DDR400 memory (we checked it for the earlier revision (B3) and for the latest one (C0))! The matter is that there were no registered modules for such clock speed in April, and validation of memory for server systems is not a quick process. Let's consider that AMD simply played safe.
I think that the clock speed will be increased: for the previous architecture the maximum clock speed reached was 2.2 GHz, and the Athlon 64 FX kicks off from this figure. That's why I hope we will soon get more efficient processors. But the revolution itself did take place already. The next grave transition will be towards the 90nm technology.
The processors are very similar, though the package is the Athlon 64 is identical to the latest green Athlon XPs with the organic base, and the Athlon 64 FX and Opteron have a ceramic one. All are covered with a metallic cap.
Let's try to guess what the marking means using the information provided. Note that the information is not strictly official, and changes are possible in future.
So, the processors are:
The first letter stands for brand: O - Opteron, A - Athlon 64. The second means the sphere of application: S - Server, D - Desktop. There are only OS and AD combinations, but who knows - maybe AMD will release a server Athlon 64 as well? :)
The third letter defines "Power Limit". But we don't know more about it and all the processors tested have "A" here, that is why we can't distinguish them by this parameter.
The forth item is model number. For the Opteron it consists of three figures: the first is the series number, the second equals 4 and the last one defines clock speed - from 1.4 GHz for "0" to 2.0 GHz for "6". The Athlon 64 and Athlon 64 FX have for figures of the performance index which correspond to a name of a certain model.
The next is the package version: A - 754pin OuPGA with a cover (for Athlon 64), B - 754pin OuPGA without a cover (Mobile Athlon 64), C - 940pin CuPGA with a metallic cover of the Opteron and Athlon 64 FX.
The next letter indicates the core voltage. It makes 1.55 V for the first Opteron model we tested (letter C), and for the rest it's 1.50 V (letter E). Letter are used only odd up to Y which corresponds to 1.00 V.
The 7th symbol defines the CPU working temperature. "O" corresponds to 69 degrees, "P" to 70 deg. Next letters in the alphabet relate to a higher temperature up to "Z" which corresponds to 105 deg.
The last figure indicates L2 cache size: 1 - 64 KB, 2 - 128 KB, 3 - 256 KB, 4 - 512 KB, 5 - 1 MB. As you can see the processors based on the AMD64 architecture have it at least 1 MB.
The last two letters define stepping, revision, socket, the number of coherent HT buses and so on. If the letter is over AI, the stepping is C0 or higher.
The most important letters are the first three as they define the processor type (server or desktop) and model index which indicates performance in the units known only for the manufacturer :)
The company plans on selling the new solutions at the following prices: Athlon 64 3200+ at $417, Athlon 64 FX-51 at $733 (the mobile versions will sell at $417 and $278 for the 3200+ and 3000+ respectively). The prices are on the level of high-end desktop processors but it's still far from desirable "$64 for 64 bits!". On the other hand, the prices will be falling down. The sales volumes will also depend on their test scores.
If you remember, AMD published the list of applications it used for rating their models yet at the moment of releasing its Athlon XP 3200+. This time they used another original approach giving a code name to the desktop processor (FX-51) instead of rating it.
Here is the modern version of the list of applications used for performance
The new version has such popular applications added as media content encoding and archiving. On the other hand, there is a surplus of synthetic tests like SYSmark and Winstone. We know that any modern processor working at the speed of 2 GHz ensures good performance in modern office applications. Sure, one can get 1000 messages a day with attachments and anti-virus scanning (for the 2GB mail base as well), but in this case it's not hardware that should be upgraded, and the synthetic tests included in the pack do not deal with such situations.
3DMark with "D3D Software T&L" are not needed either because if a user buys such an expensive processor but uses a cheap video card, it simply means that he is not going to play games.
The situation around some games like QuakeIII is not clear as well - is it worth buying a new processor only to lift fps from 220 to 290? :). By the way, in the benchmarking manual AMD suggest you "Select "Preferences" to "Speed"". On the one hand, it's clear that we test not a video card but...
What is left is MP3 encoding (but who needs it faster than 5-10 minutes per disc? :), MPEG2 conversion (but I don't understand why to covert from RAW AVI if one minute takes over 1.5GB). "MPEG2 to MPEG4" is as slow as it was before.
They'd better use more rendering and computational tasks. Probably the company uses such applications mostly for workstations. That's right because we perfectly know what home powerful PCs are mostly used for :) but the positioning of the Athlon 64 FX can be easily shifted toward the entry-level workstations if it proves to be efficient enough in such applications.
64bit applications and Windows XP for AMD64
In spite of the figure 64 in the headline 64bit applications won't be used that soon in desktop PCs. The enthusiasts can try them in some Linux versions but the 64bit mode will start spreading out only with the release of Microsoft's Windows for this platform. At the moment the company is working on two OS versions - server and desktop. The beta versions are already released, and we had an opportunity to try the prerelease Windows XP for AMD64.
The basic Microsoft Office XP, VirtualDub with the DivX codec, and Far ran successfully, in contrast to graphics applications. In spite of the "full compatibility" the Quake 3 and Return to Castle Wolfenstein failed to start (the games couldn't configure the graphics subsystem the right way). Serious Sam: The Second Encounter and Unreal Tournament 2003 Demo worked flawlessly. As to the speed, 3D applications depend much on video card's drivers. In this case NVIDIA's Detonator 50.30 released in May this year demonstrated the 30% drop as compared to the Windows XP Pro with the driver v45.23. Probably, the main problem will be the driver porting to the new platform (the drivers have to be 64bit). The OS hides them so well that one can find them only manually in the explorer. The searching program in the explorer and Far manager finds nothing. NVIDIA's driver version is also doubtful judging by the figure "50.40" and date "August 8 2003" found in the properties.
Console applications shouldn't have problems under this OS version. There can be problems only with programs that use a 16bit code (for example, in libraries) and those which use special system drivers, for example, to access hardware resources (one of such programs is the utility that provides information on the processor, mainboard and memory (cpu-z) - it couldn't show all the information under the Windows XP for AMD64). The fact that the SPEC CPU2000, some tests of which are very sensitive to the memory speed, shows almost the same scores under the Windows XP for AMD64 indicates that win32 applications (not graphics) in the new OS perform not worse (at least) than in the 32bit version.
The chipsets for AMD64 processors have one distinguishing feature: in case of desktop usage the speed doesn't depend on them: memory chips in such systems connect directly to the processor, and the only component that consumes a lot of data - video card - has its own big and quick memory. That is why data flow mostly outside the chipset. Besides, standard 100BaseTX controllers need only 10 MB/s, and hard drives (for desktop PCs) are only nearing the data rates of 70-80 MB/s though the interface development is aimed at 150 MB/s.
In case of workstations we also have 1Gbit LAN controller and RAID arrays on hard drives, but it's a different case.
Another interesting property of the chipsets is their universality and scalability. As they address processors only via the HyperTransport bus, the manufacturers can count on long and reliable operation of their models due to the positive experience with the SocketA. The fact that any chipset (at least, formally) can work both with one and with two processors makes it possible to position the same products on several markets.
But the first generation of the desktop processors has one disadvantage: they support only one HT bus. If you remember, the AMD8000 chipset has excellent expandability as most chips have two HT buses and can be connected consecutively (though the output bus is only 8bit). Since the HT in the current edition supports the exchange speed up to 6.4 MB/s there will be no bottlenecks for six PCI-X buses, twelve PCI 2.2 64bit/66MHz buses or 48 ordinary PCI 32bit/33MHz ones.
Unfortunately, current solutions not from AMD are deprived of such features and they can be used only in desktop PCs. That is why the developers have to think out something new to proceed to the next performance level.
Apart from NVIDIA's and VIA's products tested today ALI and SiS also offer solutions for the new AMD processors (ALi 1687 and SiS 755). At the moment they are dual-chip solutions, but the companies also plan on making single-chip ones. Chipsets with the PCI Express and 3GIO support are also expected. ATI also promises its chipsets including a version with integrated graphics.
One of the first non-AMD chipsets for the AMD processors was NVIDIA nForce3 150. This single-chip solution combines a bridge for the AGP and PCI buses and all controllers standard for the south bridge:
It's planned that the next version indexed 250 will include a Gigabit LAN controller, 2 PATA and 4 SATA ports. Current boards use external chips for the SATA and Gigabit Ethernet.
In the tests we used mainboards based on this chipset: ASUS SK8N for Socket 940 and GigaByte GA-K8NNXP for Socket 754.
Since our subject is the new processors, we will show only the brief specs
of the boards.
Memory size and speed are limited in case of the Athlon 64 processor because of unregistered modules used. Thus, only 2 modules can be used at 400 MHz and the maximum size is limited by 2 GB.
As usual, first versions always have all possible components installed as the developers consider that first buyers are not short of money. SK8N and GA-K8NNXP can be currently bought at $200. It's pretty much for the mass market. But we will soon get new versions without FireWire and SATA controllers at lower prices. Besides, the pricecuts will be also fostered by the competition on the market of mainboards for the new AMD processors.
VIA also offers its chipset named VIA K8T800. You might remember the phantom named K8T400M (or even K8M400 with the integrated video controller) the mass production of which wasn't started. While AMD was preparing to launch its desktop processor VIA released a new version of the chipset :) (though I think they simply renamed the old version).
In contrast to the nForce3, it looks classical - the north and south bridges are connected with the 8X V-Link 533 MB/s bus (1GB/s according to other sources). The VT8237 chip (also used in the KT600) is used as a high-end south bridge; it supports:
The Hyper8 technology, one of its advantages, stands for the support of the HyperTransport bus between the processor and chipset with 16bit/800MHz in both directions.
For the nForce3 based boards it equals 8bit/600MHz in one direction and 16bit/600MHz in the other. But such seemingly big difference is actually unimportant because the only component that consumes a lot of data in case of the K8T800 is the AGP video controller, and the AGP bus is not loaded much in real operation. Probably, it will get important in future for workstations and servers with PCI-X and PCI Express buses. Since the bit capacity and clock speed of the HT bus can be changed in the BIOS in the K8T800 based board, we carried out the express test in Return to Castle Wolfenstein and SPECviewperf and revealed no difference in performance in these modes.
In the tests we used ASUS K8V Deluxe and MSI K8T Neo boards for Socket 754.
The scores are almost equal. On the diagrams you can see scores of the ASUS. But
remember that the tests were carried out with the beta version of the BIOS, and
the scores may change in future.
The boards look like typical samples of high-end mainboards. Both use external Gigabit LAN adapters, audio 5.1 controllers allow connecting AC via optical and coaxial digital-outs. The number of storage units supported is impressive - 6 for the south bridge plus an external ATA/RAID controller reserved.
The ASUS has a special slot for the proprietary wireless card (supplied with the Deluxe version) of the 802.11b standard (11 Mbit).
Software & drivers:
The testing technique differs from the one we used before. Besides, we changed the video card, that is why the scores can't be compared to the earlier obtained ones.
We didn't use the whole list of applications offered by AMD; we ran only games, media content encoding and archiving as most CPU-intensive applications for desktop PCs.
All the tests were run three times at least, and the median was used for calculation of the final scores.
The demo scenes (checkpoint, Grand Cathedral, botmatch-antalus, flyby-antalus) were played at different resolutions with the parameters set to Quality in the game. In the video card drivers we just disabled the buffer vertical sync.
The performance much depends on resolution, and therefore, on a video card.
Only in the botmatch-antalus scene the fps didn't fall down when the resolution
was increased. The scores are compared in 1024x768 as the gap between the contestants
is greater in 800x600 while it's much narrower in 1600x1200. If we enable AA and
anisotropy the gap can disappear at all.
In this pretty old game the palm of supremacy always belonged to Intel. But
the situation changed with the release of AMD's 64bit processors. The new CPUs
clocked at 2 GHz go on a par with the Pentium 4 3.2, and Athlon 64 FX improves
its scores by 10% in proportion to the clock speed and takes the lead.
This game works better with AMD's products. If earlier we could witness parity between the Athlon XP 3200+ and Pentium 4 3.2, now the new processors easily beat the rest. Athlon 64 FX-51 is a leader again.
Look how the scores depend on a resolution. The next two diagrams show scores
of only the Athlon 64 FX and Pentium 4.
The RtCW is not a problem for the ATI RADEON 9800 Pro, nd the scores hardly depend on a resolution. The breakaway of the Athlon 64 FX is 10 to 6% depending on a resolution.
In case of the SSAM2 the scores are nearly the same in 1600x1200, but in 800x600
the gap reaches 30%.
In this game the scores are similar to Serious Sam: The Second Encounter. But the data scattering in flyby is less than 10%, while in the more complicated benchmark botmatch the leader outscores its competitor by 25%.
We also tested two fastest systems with the NVIDIA GeForce FX5900 Ultra video
card (driver version 45.23).
The Athlon 64 FX-51 still outdoes the Pentium 4 3.2 by 7.5% in RtCW and 26.7% in UT2003 botmatch.
Media content encoding
As before, we use encoding of audio data into MP3 and video into DivX, but this time the settings and program version differ.
The first test uses Lame 3.93 codec and three variants of settings:
All of them generate files of the same size at the average bitrate of 192
kbps. The WAV file, 71 minutes long, was used as the source one (rewritten from
The encoding speed much depends on the clock rate , and the Athlon XP 3200+ easily beats all AMD's new processors clocked at 2.0 GHz and even Athlon 64 FX-51. Intel's processor clocked at 3.2 GHz takes the lead with the breakaway of 10%.
The DivX encoding (codec 5.1) was carried out from a movie trailer recorded
in MPEG2 (time 2:25, resolution 720x576) in VirtualDub (with MPEG2 reading supported,
v1.5.4) with the crop, deinterlase and resize supported.
The Pentium 4 3.2 is a leader again. But this time the Athlon 64 FX-51 is close on its heels. The Athlon XP 3200+ looks pretty weak here. We can suppose that it's caused by the lack of SSE2 support, but we don't know whether the DivX codec supports SIMD and therefore we can't assert that the problem is exactly in this support. Like in case of the lame, the scores hardly depend on the memory speed.
The program used are WinRAR (v3.20) and 7-Zip (3.09.01 beta). The maximum compression settings are -m5 for WinRAR and -mx9 for 7-Zip.
The 7-Zip shows one of the best results in compression but the time of operation
is the cost of it. Below you can estimate efficiency in the maximum compression
mode (input to output files correlation) and times of its operation in
seconds. The zip format is represented with the win32 console version of
the pkzip v2.50 from PKWARE.
You can see why we excluded the zip archiver - its speed depends on a hard
drive rather than on a processor. Besides, its compression degree is lower.
This is the only test where the difference in performance of the Athlon 64 on different chipsets is well noticeable. The tandem with the nForce3 gives the best scores, maybe because this configuration uses the Sil3512 SATA controller, or maybe the reason is different.
The Pentium 4 3.2 yields a little to the Athlon 64 FX-51.
This test depends both on the memory speed (7-Zip takes over 300MB of RAM) and on the CPU clock. It looks like the integrated controller of AMD's CPUs will be preferable due to shorter delays. The Athlon 64 on the nForce3 shows a good result again and almost catches up with the leader.
Have a look at the summary table:
The AMD Athlon 64 FX-51 shines in the games and outscores its closest competitor Intel Pentium 3.2 by 10% and over. But remember that the scores much depend on a video card used.
In the MP3 encoding benchmark the Intel is beyond any competition due to the higher core clock. The memory speed is not very influential.
The MPEG2 to DivX encoding is a more complicated problem, and the core clock and the CPU/memory bus bandwidth are both important. The Athlon 64 FX comes very close to the Pentium 4. And the other processors from AMD have better scores than the previous Athlon XP.
In the archiving the Athlon 64 FX also outscores its competitor. In case of the 7-Zip it was the integrated memory controller that ensured very short memory access delays bettering the overall performance.
If we compare the Athlon XP and Athlon 64 with the same rating of 3200+ the latter will look better in all tests but for the MP3 encoding. But with the release of the 2.2 GHz model (I hope it will happen soon) this problem will be done away with.
As to NVIDIA's and VIA's chipsets for Athlon 64, they have similar scores in all benchmarks except WinRAR. But do consider the scores of the K8T800 as preliminary.
We were generally right when supposed last time how AMD's new processors could perform. They look good, but not that good as everybody would want. The potential of the new architecture can be estimated with these samples, but end-users are interested in real scores rather than in abstract debates. It's hard to say whether the Athlon XP core is used up but AMD had to show something new. And I think they succeeded.
We tested the new processor not in all applications, but it's enough for the start. Next time we will test it in professional programs and various synthetic ones.
In closing let's try to guess where AMD took such an interesting processor as Athlon 64 FX-51 which is so similar to the postponed Opteron 148. This is one of the versions, but I think it looks verisimilar.
Starting from April the Opteron line was developed in the regular way - the clock speeds were increased, the new series were released. At the same time, the company tested performance of the Athlon 64 processor which differed from the Opteron in a single-channel memory controller, and we can't say the processors were developed separately. Utilization of unregistered memory modules is natural for the desktop processor. The only unclear aspect was why the first Athlon 64 was clocked at 2.0 GHz. It wasn't enough to beat the Pentium 4 3.2. Besides, it had the sinlge-channel memory controller in contrast to its competitor. But anyway, the Athlon 64 3200+ beats the competitor in the games and archiving; only in the MP3 and DivX encoding the speed is lower.
Still, AMD needed a bright victory. That is why using a version of the server processor of 2.2 GHz and with the dual-channel memory controller and making sure that 400MHz registered modules are produced in quantity AMD launched a new brand - Athlon 64 FX which differed from the Opteron in clock rate and memory speed and from the Athlon 64 in the dual-channel controller.
It won't hamper sales of the Opteron especially because AMD can also lift the clock speed of them up to 2.2 GHz. And fixing the price a bit higher than that of the Pentium 4 3.2 the company remains on the desktop PC field.
However, I don't understand the situation concerning registered modules and this processor. Many expected AMD's desktop high-end models use usual modules. But if it wasn't so, we could have not waited so long and the processor could have competed against the Opteron 100 due to the higher clock rate and cheaper memory. Most users associate registered modules (which are actually needed for large memory volumes) with the market of workstations and servers. But it would be strange to suppose that the memory controller of the Athlon 64 FX and Opteron needed redesigning for the ordinary modules support as the Athlon 64 has no such problems. That is why it's again an unexplainable market trick.
The further life of the Athlon 64 FX is vague. On one hand, AMD can't stop increasing the clock speeds, on the other hand, the Opteron row is almost completed: x46 will be followed with x48 and further extensions needs another marking. The FX-51 will be followed with the FX-53 with the higher clock speed. A release of a desktop processor identical to the server one but with a higher clock rate (and only for uni-processor systems) would mean weakening the attack on the workstation market.
I doubt that AMD has technical problems with the release of processors with higher clock rates and two or three HT buses for multiprocessor configurations. But we can't count on a transition of the mass market to registered modules either.
In the view of it AMD willprobably release the Opteron clocked at 2.2 GHz which will remain the fastest server processor of the company until the transition to the 90nm process. The Athlon 64 FX will reach 2.6 GHz or a bit higher and will become a flagship among AMD's desktop processors. Taking into account that it uses registered memory, the volumes won't be that large. And the Athlon 64 will successfully replace the current Athlon XP.
Kirill Kochetkov (firstname.lastname@example.org)
Write a comment below. No registration needed!