AMD\'s Opteron 250 vs. Intel\'s Xeon 3.6 GHz 듀얼 웍스테이션 BMT
서원택
2005-03
2005-03-27 17:55:18
조회 18789
추천 190
이글의 출처는 http://www.tomshardware.com/cpu/20040927/index.html 입니다..
그림이 보였다가 안보였다 하네요...
AMD\'s Opteron 250 vs. Intel\'s Xeon 3.6 GHz in a Workstation Duel of the Elite
<목차>
• AMD Is Winning
• AMD Opteron Processor
• One, Two, Four... Or How About Eight CPUs?
• 64 Bits: Windows XP X64 Edition Free
• The World, But At What Price? Opteron Prices In September
• Chipset Platform: AMD 8100 + A8151 + A8131/8132 + A8111
• What\'s Next?
• Dual Core, Baby!
• Intel Xeon Processor With Nocona Core
• An Overview Of Models
• Heat Loss Decreases With Demand-based Switching
• Psst! EM64T Is Here
• Intel\'s Ideal Platform: E7525 + 6300ESB + 6700PXH + ICH5R + IOP332
• Not Downward Compatible
• Platform Alternative I: Intel 875P + 6300ESB
• Platform Alternative II: Intel E7505 + P64H2 + ICH4
• Platform Alternative II: Intel E7505 + P64H2 + ICH4, Continued
• Test Hardware
• NVIDIA Quadro FX For AGP And PCI Express
• A Board For The Opteron: Tyan K8W
• Three Xeon Boards
• Mains Adaptors And Heatsinks
• Test Setup
• Benchmark Results
• SPECapc Applications Benchmarks
• THG Application Benchmarks
• Audio & Video Encoding
• SPECopc Viewperf (Synthetic)
• SiSoftware Sandra 2004 (Synthetic)
• Conclusion
<본문>
AMD Is Winning
For a long time, the workstation sector was stagnating. Now, Intel\'s Xeon, code-named Nocona, has finally found its way into the retail channels four months after its unveiling. While the device offers a 3.6 GHz clock speed and a technically sophisticated platform, numerous difficulties held back its commercial launch, including an unusual number of steppings that crept into the chipsets.
But all\'s well that ends well: the Xeon, based on a 90-nm process, and its workstation chipset the E7525/Tumwater, is expected to offer superior performance compared to competing workstation processors. Of course, when we say \"competing products,\" we mean the Opteron, which AMD has had available with speeds up to 2.4 GHz in the form of the Opteron 250 since the beginning of the summer.
With out a doubt, Intel pulled out all the stops to upgrade the new platform. As a result, significant ingredients have been added, which are already spiking the desktop platform based on socket 775, such as PCI Express graphics and DDR2 memory. However, the latter only works with 200 MHz (DDR2-400); DDR333 can be used as an option. A new PCI-X hub called 6700PXH now provides new bus segments for up to 133 MHz - until now the discriminating customer had to rely on chipsets from the competition.
Now the Xeon processors, based on the Nocona core, work with 200 MHz system speed (FSB800), offer a 1 MB L2 cache and handle SSE3. One of these days, the CPUs are also supposed to learn how to handle IEST (Intel Enhanced Speedstep Technology), which should lower the speed dynamically to 2.8 GHz under minimal loads and thus reduce heat loss.
To note, there is the issued of enhanced memory 64 technology, or EM64T for short, which Intel, after a quiet mention at IDF in the spring, is now very inconspicuously postponing. That sounds suspiciously like what the company has had to do with the Pentium 4 Prescott, which in fact is what the processor is. We are assuming that both processors start with the same manufacturing process, but follow different validation routes and finally get a different package.
The Opteron, on the other hand, still works with AMD\'s reference chipset, the 81xx, which consists of a total of three chips for AGP, PCI-X and I/O. PCI Express does not recognize this chipset, and the nForce4 chipset will soon need to take that into account.
In the case of AMD too there is a close link to its desktop processor line. The Athlon64 FX and the 200-series Opterons, for example, are the same, except for the RAM used (the latter needs registered DRAMs). The DDR400 memory controller is integrated into the processor, affording greater independence from the chipset; all it needs to worry about is I/O and graphics. The recipe is spicy, so it\'s not surprising that the Opteron really lets it rip when every processor can access its own RAM.
The status quo is now such that the differences between the Intel and the AMD platforms could not be greater, which at the same time means that the test of strength between the two processor fighting-giants is just that much more interesting.
AMD Opteron Processor
The Opteron 250 from AMD
In the case of AMD, the clock rate seems to be stagnating - it has not increased since the Athlon64 FX 53, the Athlon64 3800+ and the Opteron 250. The maximum is 2.4 GHz, and the Texas manufacturer of the 130 Nm version won\'t say more than that. For comparative purposes the AthlonXP 2800+ was clocking 2.25 GHz as early as the end of 2002.
Unlike the desktop processors, the Opterons continue to work with HyperTransport channels with speeds of 800 MHz (3.2 GB/s), while the desktop models based on socket 939 have been accelerated up to 1 GHz (4 GB/s). The latter would have been advantageous particularly for servers and workstations, but AMD still has not taken this route. However, this is somewhat the case with Intel as well, because it settles for DDR2 400 in the E7525. After all, safety first.
With two CPUs it really takes off - assuming the right applications. Accordingly, WindowsXP also displays two load curves.
The processor socket for the Opteron has 940 pins, whereas the new Athlon64 processors work with Socket 939. One significant difference, besides the different HyperTransport speed, is the support of registered DRAMs. Incidentally, the Athlon64 FX 51 also used this socket.
One, Two, Four... Or How About Eight CPUs?
Overview of model numbers and HT channels
The way the Opteron processors are named is only confusing at first glance. If you look closely you will detect an easily comprehensible system, which first and foremost determines the number of HyperTransport channels. We will deliberately dispense with the \"HT\" abbreviation here, to prevent any confusion with Intel\'s HyperThreading.
Every HyperTransport channel is designed to be bi-directional and works with 16 bits. A 200 MHz speed (quad-pumped, the clock speed adds up to 800 MHz) yields a bandwidth of 3.2 GB/s - or 1.6 GB/s per direction. The channels are used to respond to either an I/O device (e.g. PCI-X bridge, AGP) or another Opteron processor. Each Opteron basically has a dual-channel memory controller.
Opteron 100 Series Opteron 200 Series Opteron 800 Series
HyperTransport channels 2x 200 MHz quad-pumped 3x 200 MHz quad-pumped 4x 200 MHz quad-pumped
Total bandwidth I/O 6.4 GB/s 9.6 GB/s 12.8 GB/s
Applications Single-processor Single-processor
Dual-processor Single-processor
Dual-processor
Multi-Processor
( up to 8)
A system with two Opteron processors.
A 4-way system.
This is what an eight- or multi-processor system looks like. However, this is intended exclusively for the server world.
One, Two, Four... Or How About Eight CPUs?, Continued
The Opteron System works with DDR400 with a CAS latency time of 2 cycles.
AMD supports DDR400 memory, which must, however, be designed to be registered to keep the signal load low for the memory controller - this is the only way that permits the addition of numerous memory modules, which ultimately allow many gigabytes of RAM. ECC (Error Correcting Code) memory can be used - but it is not absolutely essential.
Technically the Opteron has its finger on the pulse of the current developments - MMX, SSE, SSE2, and AMD\'s 3DNow! are all supported, as is x86-64, the 64 bit instruction set. Each Opteron has a 64 KB L1 cache and 1 MB L2 cache.
CPU-Z correctly detects our Opterons in WindowsXP without prompting.
64 Bits: Windows XP X64 Edition Free
Even in 64 bit mode, it was possible to start the test system with the x64 edition of Windows XP from the word go. The installation runs just as it does with the extremely well-known 32 bit variation.
Windows XP Professional x64 edition with Service Pack 1.
1, 2, 4... Or How About 8 CPUs?, Continued
The 360 day beta version of Windows XP x64 arrives with Windows 5.2.
An Opteron based on AMD64 should be able to work with 64 bits - and it can, but Microsoft is still not at that point, except for the server version. The 64 Bit version of Windows XP, called the x64 edition, has been in the beta stage for months. That\'s why the use of corresponding applications (if even available) is only possible within a limited scope.
AMD warmly recommends the Panorama Factory in the 64 bit edition for testing the AMD64 processors.
The current price list of the Opteron models for server and workstation
Chipset Platform: AMD 8100 + A8151 + A8131/8132 + A8111
Unlike Intel\'s Xeon, AMD cannot rely on at least two chipsets. The reference logic thus comes once again from AMD itself, although it would rather stay out of this business.
Because the chipset was launched together with the Opteron in April 2003, it can no longer be ranked at the top, technically speaking. But that\'s not so tragic now, because the only real downside is the lack of a PCI Express bridge for future graphics cards or other expansion cards. So AMD-based workstations will have to settle for AGP graphics cards for the time being. But considering the test results, this doesn\'t seem to be much of a problem. With regard to the data throughput rate, the PCI Express graphics still cannot make use of the speed advantage. The card manufacturers will eventually switch to PCI Express, with the result that new and correspondingly faster models will no longer necessarily be available in an AGP version.
But back to the main subject: the A8151 is needed as an AGP interface, while the A8131 represents the PCI-X bridge (two separate buses). Finally, A8111 is the I/O hub that gives access to all of the other interfaces like PCI, ATA/133, USB 2.0, sound and a low-pin-count interface for a Super I/O chip.
Because all the modules communicate via HyperTransport, they can in principle be combined in any way as long as HyperTransport tunneling is supported. Devices weighed down by I/O, like the A8151 for AGP and the A8131 for PCI-X, should however each be connected to the processor via its own HyperTransport channel.
Block diagram of the S2885 Thunder K8W from Tyan
The A8151 controller only holds the AGP interface.
Chipset Platform: AMD 8100 (A8151 + A8131/8132 + A8111), Continued
The A8131 provides two PCI-X channels that correspond to PCI-X 2.0.
The A8131 or A8132 chips each provide two PCI-X buses. Because PCI-X is downwards compatible to PCI 2.2, the 64 bit wide slots can also be used for 32 bit expansion cards. Both chips work with 600 MHz, which reduces the total bandwidth of 2.4 GB/s in the A8132. The A8131, moreover, only works with 8 bits, which allows a maximum of 1.2 GB/s. The A8132 also supports PCI-X 2.0, which allows clock rates of 200 and 266 MHz for individual devices.
The A8111 controller is the I/O hub of the Opteron system.
Block diagram of the A8111 I/O hub.
The 8111 is comparable to the Southbridge of any desktop chipset. And here\'s why: the demand for bandwidth is also much lower. It won\'t come as a surprise, then, that this chip is connected by HyperTransport to the A8131 or the A8132 with only 8 bits and 200 MHz. This allows for 800 MB/s, i.e. enough for common applications.
What\'s Next?
In the future, AMD would like to withdraw from the development of its own chipsets in the enterprise workstation/desktop segment. As with the first Athlon (Slot A) the manufacturer thinks it has done its time here, because the Opteron has proven its mettle many times over. In the case of the very first Athlon, the switch to Socket A worked well, because within months ALi, NVIDIA, SiS and VIA were offering suitable chipsets - and still do today.
In the workstation segment, AMD is putting its hopes on NVIDIA, because the nForce 4, successor to the nForce 3, shouldn\'t just be fast and well-equipped but also have two independent PCI Express interfaces in the top version, which will support the introduction of SLI solutions with two graphics cards. With graphics workstations in particular this step should allow for a decent increase in performance.
Manufacturers are reckoning on getting the first sample of suitable boards even before the year is out, and in comparison to earlier chipset launches, the mood is good. Tyan, for example, will extend its involvement in the Opteron area and hopes to add a quad CPU workstation based on the Opteron to its lineup soon.
Dual Core, Baby!
Probably the best news of the summer, however, involves the option of easily replacing the Opteron processors currently in use with future dual-core models as they appear. AMD admits that in the case of dual-core Opterons there are performance losses of about 10% compared to single-core dual systems, because the main memory of other CPUs has to be accessed more often. However, the possible increase in performance compared to earlier processor upgrades is apparently unprecedented.
This also means that AMD will initially not convert to DDR2 memory, or at least only with a few CPUs. While we think it is unlikely that dual-core Opterons will be designed exclusively for DDR400, a DDR2 version could follow in due time.
The memory controller integrated into all AMD64 processors demands, above all, low latency times to work as fast as possible. Here a DDR2 memory only makes sense if the increased bandwidth offsets the disadvantage of longer latency times with faster clock rates - which we think will only happen with DDR2-800.
Intel Xeon Processor With Nocona Core
The new Xeon processors with Nocona core are most easily recognizable by the absence of condensers.
When comparing Prestonia and Nocona, quite a bit has actually been accomplished. The retention module has disappeared because more efficient cooling elements have become necessary as a result of increased heat loss. These, in turn, are often so heavy that hefty jolts in transit can inflict damage on the motherboard. So Intel has found the solution in screwing the cooling element directly to the mounting plate of the motherboard, and thus in effect to a part of the casing.
This method has advantages when installing the heatsink, for where previously you could very easily slip with a screwdriver, the attachment is now much less dangerous.
As already mentioned in the introduction, the Xeon Nocona corresponds by and large to the Pentium 4 Prescott: 90 Nm fabrication, FSB800 (200 MHz FSB clockspeed) as well as a maximum of 3.6 GHz. Certainly the most significant innovation in the application of the server is, in all sincerity, the FSB capacity, which has increased by 50%, and which now, instead of 4.2 GB/s (FSB533) amounts to a regular 6.4 GB/s (FSB800).
We tested the Xeon with 3.6 GHz and FSB800.
The Intel Xeon still also works with Socket 604.
One of the changes: The heatsinks are now directly screwed to the rear panel of the casing.
HyperThreading As In The Past
The new Xeon also still has Hyper Threading technology, whereby WindowsXP Professional or Windows Server recognizes four virtual processors.
In addition, with regard to the FSB clockspeed, Intel attaches great importance to the fact that the Hyper Threading introduced two years ago also profits from it: the bandwidth of the data bus of every virtual HT-CPU now amounts to 1.6 GB/s, instead of 1.05 GB/s.
An Overview Of Models
Overview of the present Intel Xeon Models with Nocona Core
Heat Loss Decreases With Demand-based Switching
In contrast to the Desktop P4 processors, Intel is now delivering the first Nocona Xeons with DBS. This is based on what is known as demand-based switching, which Intel will also deliver in the form of IEST or Intel Enhanced Speedstep Technology from now on. It is a well-known principle: If only a small amount of processing has to be done, the fast processors lower their clock speed (from 3.0 GHz), if required, by up to 2.8 GHz, while at the same time the core voltage is also decreased slightly. As a result, heat loss is decreased and energy is saved without any loss of performance.
If, on the other hand, high performance is required, heat loss remains close to the maximum. So no savings can be made on the heatsinks on the basis of these new features. Even more aggravating is that at the moment there is no easy way to distinguish the DBS-capable processors from the conventional Nocona Xeons. The only thing that can be done is to get a read-out of the CPUID.
DBS occurs without the assistance of the operating system while the machine is in operation, and calls for a correctly programmed ACPI Table. The manufacturer of the motherboard is responsible for that. So if corners have been cut in this respect, it is quite probable that the feature cannot be used. However, if it is running, then it will continue running - to our knowledge, DBS cannot be disabled.
Psst! EM64T Is Here
The Nocona core supports the EM64T enhancement, which is compatible with AMD64. The server version of Windows is already available as a 64 bit edition.
The second interesting feature of the Nocona processor is the 64 bit enhancements, which Intel groups together under the heading Extended Memory 64 technology, or EM64T for short. As the Itanium II, Intel\'s \"proper\" 64 bit model, absolutely demands its own Windows, along with adapted 64 bit applications, Intel has preferred to accept the fact that EM64T is compatible with AMD64. As a result, to be specific, the immense number of 32 bit applications can continue to be operated - a not entirely insignificant factor. Certainly, the technology is implemented in a different way at Intel compared to its competitor AMD, but this should not make any difference with Windows. The activation of the 64 bit mode does, in practice, have one noteworthy disadvantage: the floating-point unit is thereby rendered inoperable (even for AMD).
However, according to Intel, the main reason for introducing EM64T (x86-64) is not the higher operating speed, which is made possible by the processing bandwidth, but the possibility of addressing up to 281 terabytes of RAM. Up to now, as a general rule, 4 GB has been the limit, though up to 12 GB was already feasible with some chipsets.
Expanding memory of this magnitude has, however, been impossible to date, as firstly, memory modules of over 2 GB per DIMM are very unusual and, secondly, only 50 of the 64 data lines are external with Xeon, not to mention the limited number of DIMM sockets. In that respect, there remains a technically feasible limit of 39 terabytes.
Both Intel and Microsoft are currently displaying great ambition when it comes to introducing 64 bit technology on a broad front. Intel, according to its own statistics, has already sent 4,000 systems to developers for this purpose. At the same time Microsoft is providing the 64 bit edition of Windows XP Professional for download free of charge as a 360 day demo version. Well, if that is not a reason for changing over!.
ISSE3
The video encoder TMpgEnc offers SSE3 support.
The third and last new feature of the Xeon Nocona compared to the Prestonia is Streaming Extensions 3 (SSE3), which, above all, can be used in the areas of audio and video, as well as in the scientific field, and, according to Intel, can ideally yield up to 50% more capacity.
Xeon price list for servers and workstations.
Intel\'s Ideal Platform: E7525 + 6300ESB + 6700PXH + ICH5R + IOP332
The E7525 now offers a x16 PCI Express Interface instead of AGP, so that future high-end graphics cards can be used.
As the launch of the 775 chipset has demonstrated, the highlights this year are PCI Express and DDR2 memory. This is now followed by the corresponding technology for the workstation platform. The code name of the corresponding E7525 chipset is Tumwater, while the server version E7520, alias Lindenhurst, has a different PCI-Express configuration (three x8 PCIEs instead of one x16 and one x8). In contrast to the desktop chipsets, the professional versions have a total of over 24 PCIE circuits. According to Intel\'s specifications, the x16 PCIE port of the E7525 cannot be split up into smaller units.
Furthermore, Intel is making the PCI X-Bridge 6700PXH available, which offers a PCI X-Bus for up to 133 MHz clockspeed. The connection to the E7520 or E7520 can either be made via x4 PCI Express ports or one x8 port, so that either 2 or 4 GB/s can be achieved as required. PCI X-slots are longer than conventional PCI slots, and offer 64 bit bus width and up to 133 MHz clockspeed.
There is still a regular Southbridge, but this is somewhat technically inferior to the ICH6 of the 915/925 chipsets: only the ICH5-R is used, which, in contrast to the ICH6, only offers two serial-ATA ports instead of four. For the workstation market this is less important, however. In its place, the 6300ESB can also be used, which has a PCI X-port. Both Southbridge components are still connected via the hub interface 1.5.
The block diagram for the Intel E7525 chipset.
System driver of E7525 with the 6700PXH under Windows XP.
Not Downward Compatible
PCI-Express interface block diagram of E7525 chipset
The E7525 is manufactured in 90 nm, contained within a flip-chip BGA-casing (ball grid array) and connected to the motherboard via 1077 balls. The integrated CPU interface now operates at a clockspeed of 200 MHz (FSB800), whereby slower clockspeeds such as 133 MHz (FSB533) are no longer supported. Consequently, an older Xeon Prestonia, although it is pin-compatible, cannot be operated on an E7525 system.
The two-channel memory interface is now capable of working with DDR2 400. Here a conservative path has been taken, as DDR2 533, as registered DRAM and ECC, is still not a Jedec Standard. Moreover, it is still possible to use DDR333 or DDR266, but non-registered DIMMs cannot be operated any more - this was still feasible with the previous chipset, E7505.
DDR2-400 memory modules of Infineon, registered, with ECC - recognizable by the suffix \"x72\", which stands for the increased number of memory bits.
2 GB random access memory is the required minimum for workstation applications in most cases.
Not Downward Compatible, Continued
On the X6DA8-G2 circuit board of SuperMicro there are a 16x and a 4x PCIe slot
The 6700PXH is a new Southbridge for use via PCI Express.
Platform Alternative I: Intel 875P + 6300ESB
The 6300 ESB is manufactured in 250 Nm and offers new features.
You don\'t necessarily need an enterprise chipset in order to run two Xeon processors, because since the introduction of HyperThreading, the Northbridge components have to be suitable for multi-threading operation. Consequently, desktop chipsets, such as the 875P (see above), are basically also good for Xeons. For the previous generation of Xeons, Intel had already developed for this purpose an \"Enterprise Southbridge\" by the name of 6300ESB (code name Hance Rapids), which restored the 875P to PCI-X-slots. Naturally, the connection to the 875-chipset achieves the usual 266 MB/s of the hub interface, so that fast PCI-X equipment running at over 66 MHz does not make any sense.
In comparison to the ICH5, the 6300ESB only offers four USB 2.0 ports, as well as a stripe size limited to 64 KB when using the integrated RAID 0.
Interestingly, the 875P chipset currently offers the best system performance for the Nocona Xeons (ca. 4.1 GB/s), as Dual-DDR400 is supported, whereas the E7525 supports a maximum of Dual-DDR333 or DDR2-400(3 GB/s). Also, the latter so far has proven to be faster in practice.
This graphic shows all interfaces of the 6300ESB Southbridge.
The well-known ICH5R-Southbridge lives on at Xeon.
Dual Nocona with the Canterwood Chipset on the DH800 from Iwill. This platform supports only AGP for the graphic interface and is combined with the 6300 PXH Southbridge.
Block diagram of E7505-chipset
The workstation platform used by Xeon to date is called E7505 and is marketed under the code name Placer. As with the AMD system, the platform consists of three components: the E7505-Northbridge, P64H2 as PCI-X-Bridge and ICH4 as Southbridge.
It operates at FSB533, or 133 MHz FSB clock speed, with four data transfers per cycle (quad-pumped). The memory interface operates synchronously at 133 MHz, making DDR266, but is designed for two-channel operation. A total of 16 GB RAM is supported by the chipset; the modules must be registered DRAMs. Many board manufacturers disable this feature, however, with 4 GB being the maximum that can be fitted. The ECC support is also optional. In contrast to the new E7525, E7505 offers an AGP-8X slot and no PCI-Express components at all.
The connection of the PCI-X-Bridge P64H2 completes Intel\'s hub interface in accordance with Specification 2.0. Unlike the usual Version 1.5 in the desktop area, the revised interface can now transmit a full 1 GB/s instead of 266 MB/s, thanks to the 16 bit bus width and 133 MHz. This is especially important for connected PCI-X-equipment, such as ineffective memory controllers.
The E7505-Northbridge is manufactured in 130 nm.
Our E7505 platform is equipped with four 512 MB Corsair modules, which operate with a CAS latency of 2.0.
Platform Alternative II: Intel E7505 + P64H2 + ICH4, Continued
The P64H2 provides for PCI-X-interfaces.
The P64H2 (82870P2) offers connection possibilities for a maximum of three PCI-64 cards and one PCI-X card. PCI-X 2.0 with high clock speeds is not supported, unlike in the case of AMD\'s Opteron platform.
The ICH5 (82801ER) is the I/O hub controller of the E7505 platform.
The Southbridge ICH5 (82801ER), which is also used for desktop applications on the Canterwood and Springdale platforms (875/865), is likewise used for the E7505 platform. In this case, the connection is made via hub interface 1.5 at 266 MB/s.
Test Hardware
In order to make a realistic workstation test possible we use current high-end OpenGL graphic cards. NVIDIA is currently the leader in this sector, which is why we have resorted to suitable cards from the Quadro FX model range. Further information on this topic can be found in the article NVIDIA, ATi Bring Five OpenGL Workstation Graphics for PCI Express to Market.
For the AGP tests with the Xeon together with E7505 and Opteron plus AMD\'s 8000-chipset, a Quadro FX 3000 was used, whereas the PCI-Express System based on the Xeon Nocona and E7525 was equipped with the Quadro FX 3400.
When selecting suitable graphic solutions from among the AGP and PCI Express models respectively, a problem arose: for both types, there are no versions which are technically the same, which would thus make a fair comparison possible. We have thus decided, with a heavy heart, to use the fastest possible model in each category.
NVIDIA Quadro FX 3000 (AGP)
The Quadro FX 3000 of NVIDIA/PNY for AGP 8X. There is still no faster model than the NV35GL with 400 MHz clockspeed.
Quadro FX 3000 cards based on the NV35GL can be purchased for around $ 1,200 in retail stores. While the chip is operated at 400 MHz, it brings the DDR memory up to 425 MHz.
The Quadro FX 3000 works with AGP 8x.
NVIDIA Quadro FX 3400 (PCI Express)
The Quadro FX 3400 of nVidia/PNY for PCI Express.
A PCI Express model from INVIDIA, which was specially developed for the workstation market, currently costs around $1,300. The card is based on the NV45GL, which operates at 350 MHz. Due to the architectural differences between NV35GL and NV45GL, there is a similar difference in capacity as between GeForce FX and GeForce 6-Series.
In addition to which, GDDR3 memory is already used with the FX 3400, which can handle 425 MHz.
The Quadro FX 3400 operates at 16x PCIE lines.
A Board For The Opteron: Tyan K8W
The Thunder K8W (S2885ANRF) from Tyan is intended for dual-Opteron systems and makes it possible to assign each processor its own RAM.
AMD sent us the Opteron Platform with a K8W (S2885ANRF) by Tyan. It was unfortunately not possible to launch the Solidworks-2003 benchmark with this circuit board, as the application could not address the parallel port dongle. Tyan sent us a replacement board, which did not bring any improvement, however.
Solidworks 2003 misses the dongle and refuses to cooperate.
Xeon Board: MSI Master-LS2 For Prestonia
MSI Master-LS2 (MS-9121).
We selected the MSI Master-LS2 for the Xeon with the Prestonia core, as this motherboard emerged as the test victor in our last roundup of dual Xeon motherboards.
Xeon Board: Supermicro X6DAE-G2 For Nocona
The X6DAE-G2 with E7525-chipset is the flagship of Supermicro.
For the new Xeon Nocona with 200 MHz FSB (FSB800), the X6DAE-G2 motherboard, based on the E7525-chipset from Supermicro, arrived at our lab complete with RAM shortly after the introduction of the processors.
Xeon-Board With 875P: Asus NCCH-DL
The Asus NCCH-DL based on Intel\'s 875 chipset.
As the memory interface of the Canterwood/875P offers noticeably better performance than that of Tumwater/E7525, Asus developed the NCCH-DL board. This motherboard is considerably cheaper, but it does not offer features such as support for DDR2 and PCI Express, and so takes its place at the low end of the workstation sector. A PCI Express model from NVIDIA, which was specially developed for the workstation market, currently costs around $1,300. Iwill and MSI also have similar boards in their portfolio.
Mains Adaptors And Heatsinks
With a power supply weighing 4.8 kg, the new Xeon is well provided for. Supermicro sent this gem along with it.
At 760 watts, the SP762-TS power supply from Ablecom, which is marketed by SuperMicro, offers sufficient capacity for the dual Nocona - and certainly for any other system that is not to be listed in the top server categories with four or more processors.
The new Xeon heatsinks weigh almost precisely two kilograms together.
Test Setup
Intel Processors (Socket 604)
200 MHz FSB (Dual DDR2-400) Nocona 3.6 GHz (3600 MHz 12-16/1024 KB)
133 MHz FSB (Dual DDR266) Xeon 3.20 GHz (3200 MHz 12-8/512/2048 KB)
Xeon 3.06 GHz (3066 MHz 12-8/512/2048 KB)
AMD Processors (Socket 940)
200 MHz FSB (DUAL DDR400) Opteron 250 (2400 MHz 128/1024 KB)
Intel Processors (Socket 775)
200 MHz FSB (Dual DDR2-533) Pentium 4 3.6 GHz (3600 MHz 12-16/1024 KB)
Memory
AMD Opteron
(Socket 940) 4x 512 MB - DDR400 (200 MHz)
Corsair CMX512RE-3200LL XMS3200Rv1.1
Timings: CL 2.0-3-2-6-11-16
Voltage: 2.5 V
Intel Xeon
(Socket 604) 4x 512 MB - DDR400 (200 MHz)
Corsair CMX512-3200LL XMS3205v1.2
Timings: CL 2.0-2-2-5
Voltage: 2.5 V
Intel Nocona
(Socket 604) 4x 512 MB - DDR2-400 (200 MHz)
Infineon HYS72T64000GR-5-A
Timings: CL 3.0-3-3-11
Voltage: 1.8 V
Intel Prescott
(Socket 775) 4x Infineon DDR2-533 (266 MHz)
HY64T64000GU-3.7-A
Timings: CL 3.0-3-3-12
Voltage: 1.8 V
Motherboard
Intel Xeon Platform
(Socket 604) MSI Master-LS2 (MS-9121), Rev. 1
Intel E7505 Chipset
BIOS: V1.2B4 (01/05/04)
Integrated Broadcom 5721 GbE single port with ASF 2.0
AMD Opteron Platform
(Soket 940) Tyan Thunder K8W (S2885ANRF), Rev. 1.01
AMD 8151
BIOS: V2.02 (05/14/04)
Broadcom BCM5703CKHB PCI-X GB LAN
Intel Nocona Platform
(Socket 604) Super Micro X6DA8-G2, Rev. 1.1
Intel E7525
Bios:
Intel FW82546GB PCIe GB LAN
Asus NCCH-DL, Rev. 1.02
Intel 875E
Bios: 1001 (06/21/2004)
Intel
Intel P4 Platform
(Socket 775) Gigabyte GA-K8ANXP-D, Rev. 1.0
Intel 925X Chipset
BIOS: F2 (06/07/2004)
Marvell Yukon 88E8053 Gigabit (PCI-E)
Common Hardware
Graphics Cards Nvidia/PNY Quadro FX3000 AGP
GPU: NV35GL (400 MHz)
Memory: 256 MB DDR (425MHz)
Nvidia/PNY Quadro FX3400 PCIE
GPU: NV45GL (350 MHz)
Memory: 256 MB G-DDR3 (450 MHz)
Sound Card Terratec Aureon 7.1 Space
96.00 kHz sample rate
Hard Drive (AMD System) Promise FastTrak S150 TX2plus (Bios: 1.00.0.37)
2x SATA Maxtor 6Y080M013530B (RAID 0)
80 GB / 8 MB Cache / 7200 rpm
Hard Drive (Intel System) Intel FW82801ER ICH6FR
2x SATA Maxtor 7B250S00 (RAID 0)
250 GB / 16 MB Cache / 7200 rpm
DVD/CD-ROM MSI MS-8216D 16x DVD
Software
AMD Chipset
A8111 and A8131 X 2 1 0 1 Driver Pack
AGP 8.1.9
EIDE 8.2.2
SMBus 2.0
System Management 1.8.1
High-Precision Event Timer 1.0.0
IOAPIC 1.8.0
Intel Chipset Chipset Installation Utility 6.0.1.1002
Nvidia Graphics (AGP and PCIE) Detonator 61.76
DirectX Version: 9.0c
OS Windows XP Professional 5.10.2600, Service Pack 2
Benchmarks Used
We have used many programs for desktops and workstations. This time, it was of course crucial to include the SMP support, in order ideally be able to use all test platforms to full capacity.
For the workstation sector, the SPECapc benchmarks are of great significance, as these cover a large number of graphics applications within the scope of professional use.
SPEC Benchmarks
SPECapc Solidworks 2003 Service Pack 5.1
SPEC Benchmark Mode
SPECapc 3DSMAX 6 Service Pack 1
MAXtreme 6.00.07.00 Driver
SPEC Benchmark Mode
SPECapc Maya 5 SPEC Benchmark Mode
SPECopc Viewperf Version 8.01
1280 x 1024 - 32 Bit
SPEC Benchmark Mode
Video
Mainconcept MPEG Encoder Version: 1.4.2
1.2 GB DV to MPEG II
(720x576, Audio) converting
Auto Gordian Knot
DivX 5.2 Pro
XviD 1.01 Version: 1.51
Audio = AC3 6ch
Custom size = 100 MB
Resolution settings = Fixed width
Codec = XviD and DivX 5
Audio = CBR MP3, kbps 192
Microsoft Movie Maker Version 2.0.3312.0
416 MB DV to WMV
TMPGEnc Plus 3.0 Express
Pegasys Version 3.0.4.24
1.2 GB DV to MPEG I
Size: 720x480 (DVD NTSC)
MPEG-1 Audio Layer II (192 Kbit)
Studio 9
Pinnacle Version: 9.1.2.15
Rendering - DVD Compatible
no Audio
Applications
Newtek Lightwave Version 8.0
Render First Frame = 1
Render Last Frame = 60
Render Frame Step = 1
Rendering Bench
\"variations.lws\"
Show Rendering in Progress = 320x240
Ray Trace Shadows / Reflection = on
Ray Trace Transparency / Reflection = on
Depth Butter AA = on
Multithreading = 8 Threads
Cinema 4D XL 8
Maxon Computer Version 8.503
Rendering in 1280 x 1024, \"ship_dirt\"
3D Studio Max 6.0
Discreet Characters \"Dragon_Charater_rig\"
Pixel: 1024 x 768
Rendering Single
Synthetic Benchmarks
SiSoftware Sandra 2004 Version SP2 2004.10.9.133 (BETA TEST)
CPU Test = MultiMedia / CPU Arithmetic
Memory Test = Bandwidth Benchmark
Audio
Cool Edit Pro
Syntrillium, Adobe Version 2.1
Amplitude Normalizing
Cache Size: 81920 KB
2.6 GB wave Audio file
Nuendo 2.0
Steinberg Version: 2.2.0.33
VST Multitrack
ASIO Driver: Aureon sky/space (96.000 kHz)
Clock Source: Internal
Resolution: 32 bit (float)
Sample Rate: 96.000 kHz
File of type: Wave File (.wav)
Channels (Stereo Splite)
Microsoft Media Encoder 9 Version: 9.00.00.2980
436 MB AVI File convert to WMV
Windows Media server (streaming)
Benchmark Programs In Detail
The SPECViewperf 8.01 is a synthetic benchmark, which was derived from real applications.
3D Studio Max 6 SP1 from Discreet is one of the best-known solutions for 3D modeling and rendering
Solidworks 2003 SP5.1 from Solidworks
Maya 5 of
Benchmark Programs In Detail, Continued
Lightwave 3D V8.0 Discovery Edition von Newtek
Cinema 4D XL 8 von Maxon Computer
3D Studio Max 6.0 von Discreet
Benchmark Results
Solidworks 2003
3DSMAX 6
Maya 5
Newtek Lightwave
Cinema 4D XL 8
3D Studio Max 6.0
Mainconcept MPEG Encoder
Auto Gordian Knot
Microsoft Movie Maker
TMPGEnc Plus 3.0 Express
Pinnacle Studio 9
Cool Edit Pro
Steinberg Nuendo 2.0
Microsoft Media Encoder 9
SPECopc Viewperf (Synthetic)
SPECopc Viewperf, Continued
SiSoftware Sandra 2004 (Synthetic)
Conclusion
AMD can consider itself lucky, because due to the dual channel memory controller that is part of each processor, the dual Opteron has a nice advantage, despite having a clock speed that is 1.2 GHz slower. When simply comparing AMD and Intel, the astute observer will also notice that the results for the Xeon Nocona are, to a large extent, due to the 875P-chipset, which we also included. The reason behind that is the memory controller, because even though the E7525, alias Tumwater, supports DDR2-400, these memory modules are worlds behind DDR400 in terms of the timing possible. The new memory thus currently only delivers theoretical advantages, which is why the increase in FSB to 200 MHz does not achieve much either in many cases: the Xeon clearly suffers from a lack of memory performance.
Moreover, AMD also benefits from the fact that every Opteron processor ideally has its own area of RAM. The theoretical total bandwidth amounts in this case to 12.8 GB/s, whereas the DDR2-400-interface of the E7525 can only deliver 6.4 GB/s, and this bandwidth is also missing from the two Xeon processors.
In the area of video, AMD clearly asserts itself, while Intel is marginally ahead in 3D. And even so, only if we include the 875P platform in the evaluation. If we only compare the Opteron 250 and the Xeon Nocona, including E7525, it is the Opteron that wins out by a small margin, in spite of significantly inferior graphics hardware. The test with Solidworks is, unfortunately, still outstanding, as their copy protection dongle thwarted our efforts. The results for the applications that you use should in any event be examined in detail, as all we can do here is collect and evaluate the results.
Intel, in turn, has medium-term advantages when it comes to PCI-Express components or graphics performance, as the latest expansion cards and the fastest OpenGL boards will be appearing especially for PCI Express. Meanwhile, things are still undecided regarding the \"64-bit\" factor, because benchmarks will only really make sense once 64-bit applications become available.
