Intel Canterwood D875PBZ 14th April 2003
Before we get started we need to explain the format of our reviews over the past six months. Readers will have noted that an entire review is laid out on one long page instead of 15-20 short pages. While it takes longer for a single page of this length to load we have designed the layout so that text appears first and graphics last so you can start reading before completion of loading. We decided not to split over 15-20 pages as many people find it annoying to have interruptions once they start getting into a review. It also means you can use the search option in the menu bar to go straight to parts that are of interest. We'd like our readers to let us know which method they prefer by providing feedback in our Forums. In any case, by the time you finish reading this paragraph the page should have mostly loaded.
It's now exactly five months since Intel launched their Hyper-Threading (HT - not to be confused with Hyper-Transport) Technology and the time has come to push the boat out further. The P4 3.06GHz CPU has a multiplier of 23 and this is really at the point where the Northwood core begins to run into latency issues. Future designs such as Prescott will alleviate this but the best way of reducing the multiplier while ramping up the frequency is to increase the Front Side Bus the CPU uses. With the widespread availability of DDR400 memory Intel has opted for an 800MHz FSB by supporting Dual Channel DDR400. Today they are simultaneously launching the Canterwood chipset and the 3GHz P4 with 800MHz FSB (they couldn't really have one without the other although Canterwood is backwards compatible with 533MHz it would rather defeat the purpose)..
Having both CPU and memory running at the same frequency means no timing issues that in other chipsets have resulted in poorer performance with DDR333 memory than with the faster DDR400. Although Intel will be releasing the 865 (Springdale) chipset in the near future in a variety of flavors as it's expected to be the mainstream choice, today's launch is their flagship chipset, the 875 Canterwood which is aimed at enthusiasts, power users and the workstation market. Several new technologies appear for the first time with the D875PBZ motherboard including Communication Streaming Architecture (CSA), native support for S-ATA in the Southbridge (ICH5/ICH5R). We'll also explore how Performance Acceleration Technology (PAT) works.
Let's get the CPU out of the way first.
This is the new 3GHz Pentium 4 processor with an 800MHz Front Side Bus (FSB). The CPU multiplier used is 15 rather than the 23 used in the 3.06GHz Pentium 4 with a 533MHz FSB. It supports the same Hyper-Threading so we will have to wait until Prescott is released to see any fundamental improvements to this technology - naturally we will publish a new Hyper-Threading Technology Guide when this happens. The official designation for this CPU will be announced very shortly by Intel. Also of interest is the heat sink and fan combo that Intel sent us.
This is a large and heavy beast with twisted fins to increase the surface area that dissipates heat. We're not sure what aerodynamic advantages there are to this design. Here's a side view.
The fan is not contained allowing it to be as large as possible and care should definately be taken to avoid finger damage. In operation the fan is surprisingly quiet and stays around the 2000rpm mark. The overall impression is of an elongated orb type cooler that should suffice for any air-based cooling needs.
The Canterwood Motherboard
This is after all the main thrust of our review. As with all Intel boards the build quality is very high but don't expect any overclocking capability, the board is designed to run at spec reliably and the failure rate of Intel's own boards is extremely low. We had no stability problems at all with this board except for a small intermittent glitch with the onboard RAID which is detailed below.
The following are the manufacturer's specifications:
This seems fairly comprehensive so lets take a look at the board itself:
The DIMM slots are well positioned and both channels clearly marked. Details in the instruction manual clearly show how to position memory and under what circumstances the board can operate in dual channel mode. Our benchmarks showed single channel mode to be worse under every test so we will omit those results and concentrate on dual channel performance although it's useful to know that those with only a single DIMM can use the board and upgrade to dual channel performance at a later date.
The ATX power connector is well positioned and will not obstruct the air flow of the CPU although it may be awkward to access the floppy connector. This is not the case with the 12V 4-pin connector which could have been better positioned as it may obstruct exhaust fans on some cases.
While its not apparent from the perspective of the above photo, the fins on the Northbridge heat sink are quite tall. We prefer passive heat sinks because those little 40mm fans used spin at high speed and can create a whine that can be more annoying than the noise of the CPU fan.
We find the usual connectors here including the onboard gigabit LAN connector. Although this is a high end board we do not find features like onboard audio. This is not necessarily a bad thing as most users will have their own sound cards. Six USB 2.0 connectors should be enough for anyone. Onboard audio and integrated graphics are much more likely in the forthcoming Springdale based boards which are targeted more towards the entry level market.
The board we received had onboard RAID referred to as SOFT-RAID in the BIOS. This immediately gave the impression of a cut-down solution but our tests have shown it to be slightly faster than our Promise TX2000 with the same type of drives. This controller is only supported in Windows XP (as is Hyper-Threading) so another reason to upgrade from Windows 2000 etc.
A useful feature is the ability to migrate from a single drive to a RAID0 configuration without backing up your data, reformatting and re-installing. Experience has shown us that no matter how reliable this process it's always a good idea to back up first. The onboard RAID only works with S-ATA and not IDE so you will need to obtain S-ATA drives which are still scarce at the moment. We would have liked to see more S-ATA channels available although 2 should be enough for most people. The serial power cables we received did not fit very well and occasionally when rebooting we received a POST message stating that one of the drives was missing although resetting would cure this. The chip in the above photo is the new ICH5 which has native support for S-ATA so no bridging is required.
Communication Streaming Architecture (CSA)
Data arriving through a PCI network adapter must traverse several buses before it is presented to the user. These buses operate at different speeds. Most interconnect speeds exceed a Gigabit per second. The 32-bit/33 MHz PCI bus's maximum speed is 1.06 Gbps which cannot support full-duplex Gigabit speed through the network interface.
CSA is a new communications architecture that creates a dedicated link from the Memory Controller Hub (MCH) to the network interface, offloading network traffic from the PCI bus. CSA, based on HubLink architecture, provides a throughput of 266MB per second, making it twice as fast as the PCI bus interface.
Direct attachment to the MCH affords a dual benefit. Memory read and write operations have lower latencies because data moves directly from the network interface to RAM. CSA also reduces the number of traverses across the HubLink interface to the ICH by half, further reducing the latency of the network transmission. Network-sourced traffic is sent from the network controller via CSA to the MCH and RAM, then to the ICH via the HubLink interface and local storage. This architecture eliminates the latencies inherent to PCI-based network traffic.
Readers may be wondering why there is such a push toward Gigabit ethernet when 100Mbps meets their needs. Intel is looking to the future and Gigabit is set to become the standard for networking (and not just for connecting servers but for the home too).
Performance Acceleration Technology (PAT)
Since this is the enthusiasts board of choice (more likely because it's aimed at the high-end and workstation market) Intel has incorporated what they refer to as PAT into their 875 Canterwood chipset. Basically this optimizes memory timings to give some extra performance.
Firstly to enable Dual Channel Operation the following conditions must be met:
The following conditions do not need to be met:
Optimal configurations for highest performance:
As you can see, there are a lot of factors to take into account when choosing memory to get the most out of Canterwood. It's not as bad as it seems but it would be worth buying memory in pairs.
Here's a summary of the system and the reference systems we compared it to:
The above descriptions are accurate with the following exceptions:
The memory used was DDR333 for the DDR machines (DDR400 for this board) and PC1066 for the i850 system. DDR400 was not used (apart from the Canterwood D875PBZ board) as it gave worse results due to memory timing issues.
This clearly shows the 800MHz FSB and that Hyper-Threading is supported.
Let's start with the synthetic benchmarks.
This test is largely dominated by the CPU bus and so the higher FSB makes no difference here. The dual CPU bus of the Dual Athlon system leads
Now the full might of having a synchronized 800MHz FSB is apparent dwarfing the other P4 platforms.
Little difference between P4 platforms here but the double width AMD 760MPX CPU bus really shines in this test and easily bests all rivals.
Same situation here. The 3GHz P4 Is slower than the 3.06GHz P4 clock for clock.
PC Mark 2002
This is a series of tests and is more comprehensive than any of the Sandra benchmarks.
And in comparison to other machines:
The most notable thing is the very high memory score as expected but how does this translate into practical use?
3D Mark 2001
We repeated this several times as we just couldn't believe this score. It is absolutely astounding on a Ti4400 graphics card and wouldn't look out of place on a Radeon 9700 Pro. No doubt gamers will take delight in having a Canterwood Rig.
The D875PBZ is head and shoulders above the competition, beating other P4 rivals despite running slightly slower.
Unreal Tournament 2003
Now for some real world benchmarks starting with UT2003 Flyby.
Again the Canterwood board blows away the competition.
The same results are repeated here with an even bigger lead at the highest resolutions. It's interesting to see that all the systems are CPU limited at lower resolutions differently (i.e. some platforms are more efficient). This is also the case with the bandwidth limitation at the highest resolution. Until Doom 3 is released we just don't have any games that can really stress today's systems.
Let's turn to an area where we know that fast CPUs will make a difference - Audio and Video encoding. This is becoming more and more popular and is very computationally intensive with long processing times (relatively speaking that is, this field is not for those that complain about how long their Outlook Express takes to load).
For consistency we will use Jet Li's The One as our test matter. It is not interlaced and contains a mixture of action types and is not too long. There will be three tests all using Divx 5.02. Audio will be encoded separately. I will try and keep my commentary to a minimum as all configuration information is shown in the images below.
and here are the results
For the first time ever we have an encoding speed in excess of 100 frames per second showing the Canterwood board t be ideal for video encoding.
AviSynth and VirtualDub
No serious Divx encoder uses Xmpeg alone and it's just used by the media for benchmarking purposes so let's get serious. We ripped our source material to hard disk and created a DVD2AVI project file using forced film (it was 99% film). Loading this into Gordian Knot we first saved an .avs file with no changes at all (720x480) and no filters of any sort. This was loaded into VirtualDub with the following CODEC parameters :
After encoding we got these results:
The D875PBZ is even faster here earning the applause of millions of Divx users around the world.
This is all good stuff but how about a real-world test? To simulate a realistic test we added a neutral bicubic resize filter in the .avs file and used the following CODEC parameters (including two popular Pro settings) which are designed to total 700MB (when the audio is muxed in):
Which resulted in the following.
The addition of computationally heavy filters put less emphasis on the memory bandwidth and more on raw CPU speed, changing the situation dramatically. The Dual MP2600+ is well in the lead due to the extra processing required for the filters and Pro settings. The Canterwood is still the best single processor solution.
What about audio? We took the AC3 track from the above sample material and used HeadAC3he to convert it into Vorbis format so our final muxed file could have Ogg containment. There isn't space here to go into the advantages of Ogg Vorbis over MP3 and AVI so let's just say that Vorbis sounds about the same as MP3 for half the file size or twice as good for the same file size (that is subjective though).
Since it is more meaningful to show throughput than time taken (which depends on the length of the source) we display the results thus:
This is another test that only uses a single processor and is CPU intensive but the D875PBZ scores very well.
Intel have raised the bar yet again. While AMD talk about the imminent release of their 64-bit processor Intel have added Hyper-Threading technology and now an 800MHz FSB product range with features like CSA and PAT. CSA in particular may go on to help more than just the network connection. Already there are signs that Intel want to overhaul the AGP bus. They have not seen the lack of real competition as a chance to sit on their laurels, rather they are increasing their lead and market share, something bound to increase further with this release.
In the next few weeks many manufacturers will be releasing their own motherboards based on the 875 chipset. The 875 (and 865 very soon) chipsets along with 800MHz FSB CPUs are a significant milestone.
Given the test results we have seen there is no reason for anyone buying a Pentium 4 system to not go for an 800MHz FSB (there will undoubtedly be 865 solutions for all pockets).
Intel should be applauded for driving technological change at a relentless pace instead of taking a planned obsolescence approach. They definitely know that in the long run those with the best technologies will triumph.
We would like to thank INTEL UK for the review sample P875PBZ Motherboard and P4 3GHz CPU.
We would also like to thank MICROSOFT UK for the review sample Windows XP Pro and their technical assistance.
All trademarks are the property of their respective owners.