We lifted the veil of secrecy off the future last week when we issued the first test results of DDR2, a new memory type. (The same article also contains a description of basic differences between DDR2 and its predecessor.) However, this pales before an official announcement of a new Intel platform (June, 21) which will surely bring a lot of changes into our lives. Well, let me introduce you the i925/915 Express family supporting Socket 775, DDR2, and PCI Express.
Intel 925X/915P/915G Express
The schematic below contains key features of the new chipsets:
There are no differences between i925X and i915P in terms of architecture, and i915G is, in fact, the same i915P with integrated graphics. The release of the i865/875 and i848P lines was obviously aimed at supporting new Pentium 4 CPUs with an 800-MHz FSB and also at implementing the dual-channel memory mode. The situation with i915/925 is somewhat more complicated. It doesn't just provide a ground for systems based on Pentium 4 with the new LGA775 socket, it also supports multiple new technologies which suddenly appeared all together and will stay for long in our lives. Well, let's take them one by one.
Support of CPUs with a 533/800MHz bus frequency. First of all, it is only the i915 series that supports 533 MHz, while the top i925X doesn't condescend to old-time Pentium 4 models or new Celeron D ones. In fact, this is exactly what we saw when i875P was announced: in contrast to i865x, that particular chipset didn't support 400-MHz buses. Second, it is sort of implied that the new chipsets will only work with new CPUs made for Socket 775. But that is not true: because the bus interface hasn't changed, it is possible to make i9xx-based boards for Socket 478. However, they will hardly become widespread. As for the details of the new socket, you can find them in this article. All new chipsets support Hyper-Threading, and that actually brings us to the end of the processor part.
Dual-channel DDR333/400 and DDR2-400/533 memory controller. Here, it's all about market positioning too. Officially, i925X won't work with DDR (no compomises with the past). When you select memory frequency, you'll see no lowering multipliers, such as DDR400/DDR2-400/533 at FSB 800 MHz (1:1 and 4:3) or DDR333/400 at FSB 533 MHz (5:4 and 3:2). And you can use DDR2 only with CPUs on 800-MHz buses (although there is contradicting information on that account, so we'd better wait for real boards). The qualitative parameters of the memory subsystem haven't changed: up to 4 modules/4 GB; i925X support ECC, i915 do not.
Since i865/875 times, the organisation of the dual-channel mode has become more handy for users. The so-called Flex Memory technology enables to install three DIMMs preserving dual-channelness if both channels have the same total memory size. For example, stepwise upgrade with no performance loss can now look as follows: 256/256 MB > 256+256/512 MB > 256+512/256+512 MB. The system can surely cope with an asymmetrical slot filling in different channels, but then performance will fall (as is the case with i865/875).
One of the burning issues about i915x is if they can juggle DDR and DDR2 simultaneously. Problems are likely to arise with onboard layout of connectors for both memory types. At least, Intel has no Design Guide for a case like this. There are samples with DDR and DDR2 slots, but their stability is dubious. And of course, only one-type DIMMs can function simultaneously (there is a set of pins, common for both types, that is responsible for interaction with memory in the chipset, and an external onboard commutation is needed). Thus, you won't be able to make incorporating memory upgrade and therefore, such hybrids will hardly be of any use. On the other hand, you can install your current (or newly-bought) DDR modules into a future board that has soldered connectors of this type. It is an essential point, as DDR2 gives no real performance gain in today's and tomorrow's computers, such modules cost much more, and reduced heat-emission can only be interesting in the case of small-size (SFF) systems.
To increase performance, Intel chipsets use a number of memory controller optimisations which are supposed to make them superior to rival chipsets. There are two noteworthy innovations here. First, the memory arbiter tries to perform packet refresh of the memory pages according to the current access algorithm. It takes up to 15 percent of operating time to regenerate dynamic memory, so it is essential that it is refreshed within interaccess periods. Evidently, an i9xx arbiter uses a sort of a smart analysis of the current frequency and access character and can also employ specific DDR2 features. The other innovation consists in an advanced data rearrangement in the memory. Obviously, the memory controller divides the linear address space into banks according to DIMM configuration and access mode. Thus, it can provide more open pages for integrated video if display resolution and refresh rate are increased.
And certainly, Intel didn't forget about the legend for i925X: it can't be considered a top chipset just because it has minimal support of older componentry, and ECC support can hardly be regarded as a marketing advantage too. i925X performance exceeds i915 insignificantly and only due to PAT which now has no special name, though sometimes you can come across the word «Turbo» in the documentation. The fact that many board manufacturers enable PAT on i865x chipsets has definitely compromised this term. As for acceleration, it results, as usual, from a faster data pass through the chipset due to less matching buffers on intrachipset ways (the so-called trajectories) which are used for certain combinations of FSB/memory frequencies. Naturally, the question arises if PAT can be enabled on i915. Well, time will tell, but we have reason to believe that it can't. (Our permanent readers may remember Intel's patient for a technology that allows to block overclocking on the manufacturer's chips.)
PCI Express x16 bus for external videoaccelerators. Intel sees no ground for compromises in this respect, and the company has rejected AGP altogether. Indeed, it is too difficult for a chipset to juggle both interfaces due to a great difference in electric parameters, while a separate pinout unit for AGP 8x is too much, considering MCH has over 1200 pins as it is (see our latest article with comparative results for more details). Alternative ways include VIA chipsets (though we have seen no real working boards with two graphic interfaces) and a PCI/AGP slot transformation (some manufacturers have come up with this idea but realisation details are yet unknown). And again, time will show how stable, productive and hence attractive such boards will be.
Integrated i915G graphics. The third generation of Intel's modern integrated graphics is officially named GMA (Graphics Media Accelerator) 900. GMA 900 differs from its i865G predecessor in an increased core frequency (333 vs. 266 MHz), more pipelines (4 vs. 1), hardware support of DirectX 9 (vs. 7.1) and OpenGL 1.4 (vs. 1.3), as well as in some minor details. We don't think it's really interesting to take a closer look at each of the points considering it is a low-end accelerator. But Intel officials also stress an almost twofold excellence over Intel Extreme Graphics 2 (i865G) in a number of tests, and that is what we'll check as soon as we get a corresponding motherboard. Until then, we can only mention that despite a formal support of DirectX 9, GMA 900 is only suitable for old games. Those willing to buy an i915G-based board will probably be more interested in the support of higher display refresh rates (up to 85 Hz in QXGA (2048x1536)) and two separate images (vs. one in previous generations of integrated Intel graphics).
ICH6(/R/W/RW) south bridge and a connection to the north one via a DMI bus. We'll dwell on the functionality of the new south bridge family a bit later. For now, we'll note that the bridges in i915/925 are connected via a special DMI (Direct Media Interface) bus which is an electrically modified version of PCI Express x4 (~2 GB/s). Why so low one may ask. Well, Intel's interbridge (interhub) bus has never set records in bandwidth (e.g. it was 266 MB/s in i865/875 vs. ~1 GB/s in SiS chipsets). However, Intel system logic sets have never seen fixed falls in periphery interaction performance. Evidently, the company's engineers are competent in assessing the real need for bandwidth. As for ICH6 variants, there are 4 of them: minimal, with a RAID function, with a wireless network (Wi-Fi) organisation function, and a combination of the latter two. All the details are given below.
4 PCI Express x1 ports. It is certainly no accident that the full name of the new chipsets is Intel 9xxx Express. The use of PCIE instead of PCI is a more important event than a transition to DDR2, to say nothing of a socket replacement. A single-channel PCI Express variant is designed for work with standard desktop periphery, so new boards will have several PCIEx1 slots (500MB/s bandwidth) and probably soldered controllers connected to the south bridge via this interface. Theoretical aspects can be read in our presentation of PCI Express, and practical ones are soon to come as the industry has already embarked on a transition to this bus.
Matrix Storage: support of Serial ATA devices with RAID and AHCI (RAID only in versions ICH6R/RW). Now there is support of 4 SATA ports for 4 devices with a possible RAID 0 or 1 organisation from any two discs and a separate RAID 0 or 1 organisation from any other two discs. But it's not only quantitative aspects that have changed since ICH5 times. In the spring of 2003, HDDs with full-spec SATA controllers were practically absent from the market, and normally, interface bridges on the HDD control board were used. Thus, there was no realising such functions as NCQ (Native Command Queuing) and others from the SATA AHCI (Advanced Host Controller Interface) set. In 2004, HDDs that realise this basic functionality have become available, so the new generation of Intel south bridges offers AHCI support.
For users, the most important results of the SATA disc working via the ICH6 SATA controller are acceleration of RAID and even a separate disc, as well as Hot Plug support. But an AHCI-aware driver (Intel Application Accelerator 4.0 and higher) is needed for this, as the SATA driver integrated in Windows XP has no such support. The difference in this case can reach 10 percent in tasks that are sensible to the disc subsystem speed. Hot Plug itself is not so important for users (you don't often remove the HDD on the run, do you?), but now you can fully realise RAID 1 (mirroring). RAID 1 is mainly used to detect a fault, disable the fault disc, install a new one, and continue working (information is doubled automatically). In ICH5R times, one had to switch off the computer and restore RAID in BIOS.
RAID Migration and Matrix RAID are two other interesting ICH6R functions. The former enables a transparent extension of the current one-disc system to RAID 0 or 1 straight from Windows, and you don't have to digress from your PC activity during array organisation. Matrix RAID forms RAID 0 and RAID 1 on two discs simultaneously. For this, each disc is divided into two parts that are then combined into different arrays one by one. As a result, we can get a small but fast disc (RAID 0) for storing the swap file and frequently used applications and a safe one (RAID 1) for documents, archives, etc.
High Definition Audio: a new standard for integrated sound. The basis of the future standard (and HDA is very likely to become one) is a support of 24-bit 8-channel sound at the sampling frequency of 192 kHz. As we remember, the AC'97 specification mentioned 20-bit 6-channel sound at 96 kHz. HDA's advantages are obvious especially considering that it supports all actual audio formats (including Dolby, DTS, and DVD-Audio). The functioning scheme has undergone no fundamental changes since AC'97 times, only a clock oscillator of the south bridge is now suggested to use instead of an oscillator integrated into the codec. In theory, it increases sound processing quality due to a stable and accurate oscillator frequency.
All AC'97 codec manufacturers have easily shifted to the new-standard products, so there is no shortage of componentry. By the way, ICH6 not only supports HDA, but can also work with AC'97 codecs (though not simultaneously as the board must realise only one type of integrated sound). Thus, we may occasionally come across a new model with old music. A single universal Microsoft driver is regarded as HDA's strength, but in practice, proprietary drivers are unlikely to be rejected, as each codecs has its own particularities and extended functionality features, and a software interface (mixer, equaliser, etc.) is already written for the codec's own hardware.
Jack Retasking is a curious feature of the HDA standard. It retasks an audio connector according to the device plugged into it. The current version of AC'97 specification (v2.3) mentions Jack Sensing which identifies the type of the connected device (due to cardinally different impendance of, say, a microphone, headphones, and active speakers). Sounds good, but in reality, it is not so useful to have a program telling you that you have connected speakers to the Line In. You can sooner find it out without software assistance, just by hearing no sound from the speakers. And in general, the best thing to do is to consult the user manual. But it's quite another thing if a smart computer can receive (and preamplify) the microphone signal with the help of a new module codec scheme and output the sound to the headphones irrespective of the connector their plug has been fitted into. Thus, we'll no longer have to reach for the back panel of the PCM trying to see what is wrong through layers of cables and dust.
Wireless Connect, an organisation of a wireless network. Here we have the most complicated situation. In principle, a non-supplied extension card together with the ICH6W south bridge realise the functionality of a wireless LAN controller standard IEEE 802.11b/g (Wi-Fi). However, it is unclear what ICH6W is charged with in this liason (we only know that it stores firmware and is responsible for ACPI modes). The format of the extension card is uncertain too: it seems to be a PCI, but there is information that it is a PCIEx1. Its reference design is ready, but again, we know nothing about the shipment time. Considering the fact that you pay a little too much for ICH6W (compared to ICH6) and will have to buy the only variant of the extension card (made by Intel, that is, not inexpensive either), we think it's better to choose a complete solution of some other manufacturer. The advantages of Intel's variant might only include convenient software that enables even a beginner to organise a wireless network using the mainboard controlller as the access point.
South bridge supporting other peripherals. The term USB 2.0 wasn't so popular a short time ago. Now a support of 8 ports of this standard is taken for granted and is not regarded as something new. The expansion of PCI Express and Serial ATA hasn't caused (and won't cause in the nearest future) a full rejection of old peripheral interfaces, so up to six PCI Bus Master devices can work with ICH6, whereas the number of Parallel ATA channels has been reduced to one (for two devices). It is strange that Intel hasn't delivered a proportionate reaction to the spread of Gigabit Ethernet, but anyway, the only noteworthy thing about the company's new south bridge is a Fast Ethernet (10/100 Mbit/s) MAC controller. And may we not mention other legacy and low-speed peripherals, such as SMBus and LPC buses.
We've done with the chipset's functionality, but ther is one more important thing about i915/925-based mainboards, namely, the power scheme and the form factor. Well, Q3 and serial products of the new BTX form factor are soon to come, but there seems to be no fever in this connection. In general, the transition to BTX won't be obligatory and across-the-board, as the standard is mostly important for small-size systems. Concerning the market of desktop products, ATX is still reckoned with and will retain its position if it introduces timely modifications.
The current one is power-related once again. Because actual norms allow to make graphic cards for PCIEx16 with an over-75W heat emission, +12V is necessary to be supplied to the board via an additional conductor (6 A). The power connector has been specially enlarged for motherboards: the old 2x10-pin socket has been made into a 2x12 one, and four new pins (+3.3, +5, +12 V, ground) are placed on the side, so an ATX 2.03 PSU connector can be used. But certainly, it will only work if you don't use too power-consuming components. However, manufacturers can quite resort to the already-used solution of soldering a socket for the HDD power connector (in fact, the socket will move from AGP graphic cards onto motherboards with PCIEx16).
We decided not to compare potential rivals theoretically as only SiS had officially released the specs of its first PCI Express supporting chipset (SiS656) by the moment the article was written. And as for a practical comparison, it will come much later than that as retail stores sell no DDR2 memory, or PCIE videoaccelerators, or PCIEx1 extension cards for the moment. Due to these reasons, it was the first time we suffered acute shortage of hardware not of time when carrying out our tests.
The table below contains brief characteristics of the tested boards.
The GeForce PCX 5900 (PCIEx16) card that opened our testing couldn't make it to the end due to certain reasons, and we had to replace it with an ATI's middle-level PCI card (Radeon X600XT). As a result, the systems' readings can't be directly compared in the games and i925X has to skip the game tests.
We will start with the low-level memory test and use RightMark Memory Analyzer. Interestingly, in read without data prefetch, i875P shows much higher performance than modern i915/925 with DDR2 that has higher bandwidth. If we use prefetch, both new systems are limited by FSB bandwidth (6.4 GB/s; the nominal FSB frequency of i915P-based boards is overstated by about one percent). i875P is two percent slower, probably because 6.4 GB/s limit this chipset too.
Chipsets with DDR2 have a higher memory write speed, and i925X takes the first place here. The test of maximal memory write speed with the influence of the CPU cache excluded (forward data save) shows that all the chipsets are roughly equal.
We're measuring memory latency using pseudo-random access as it prevents the hardware prefetcher to overstate the result and doesn't allow the results to lower due to D-TLB misses (that take place in the case of truly random access). However, the algorithm is not effective enough if you're working with a Pentium 4 as the CPU reads two 64-byte lines to the cache at each memory address. The solution is to increase the step size at read to make it larger than the standard size set by RMMA.
Now it's fair. These results have an absolute value and can serve for platform comparisons. However, the "relation of forces" hasn't changed. i925X/i915P-based boards show very close results (i925X is a little better), while i875P confirms the well-known fact that DDR has smaller latencies.
Archivers are the most sensitive applications to memory latencies. The results matched our expectations: i875P is the leader, i9xx are roughly equal and lose about 5-10 percent to Canterwood.
Video coding speed measured according to our methodology,
depends little on memory performance and mainly responds to CPU architecture
and frequency. We're only giving you the results of MPEG2 coding carried
out with the help of Mainconcept MPEG Encoder, one of the best programs
of its class. The difference between the rivals was just one percent,
so we don't think any comment is needed here.
Chipsets usually have a minor influence on performance of rendering programs (just like video coding). And because we haven't yet seen visible performance differences between the rivals, we can't expect to find them here. The maximum we have is three precent in Lightwave.
When we nearly lost all hope to see an intrigue in today's competition, Photoshop results suddenly gave it to us. We made several repetitions of the test only to find that the i875P platform really loses much to its rivals. Its setback can only be accounted for by the interference of some dark forces. Or else, PCI Express accelerates Photoshop no worse than Pentium III used to do with the Internet some time ago.
It should be borne in mind that the difference between videoaccelerators can have an impact on game test results, even if the former are only different in the interface bridge. Anyway, our testing revealed a certain difference between them. But on the whole, we saw no sensations and i915P didn't look faster.
Because our today's testing was somewhat limited, the article proved to be a more theoretical one. Thus, we won't be trying to estimate percentages or declare winners. New technologies are infiltrating into desktop systems, and we should definitely take notice of them. Well, we hope our presentation of the first Intel PCI chipset family was helpful to you. In the future, we'll test many other boards, estimate stability and convenience of mixed solutions, reveal the real difference between the chipsets using some of them with well-tuned BIOS versions
For now, we can only note that the new systems were generally not worse than the fastest Socket 478 which had ideal conditions. So, if you want to feel the progress course on yourself, you won't at least have to sacrifice performance for it, like it was in the times of Pentium 4 with Socket 423.
Write a comment below. No registration needed!