by Ilya Gavrichenkov
08/04/2004 | 12:38 PM
Intel’s recent announcement of the new i915 and i925 chipsets made no revolution in the mainboard market. A month has passed since the official introduction but there are still few mainboards on the new chipsets available. However, it would be rather imprudent to claim that the release of the i915/i925 platform was a total failure. The chipsets of the i9xx family conceal numerous innovations that just cannot be accepted by the user community in a second. It takes time for innovations to permeate and win the market – that’s why the new Intel chipsets haven’t been able to show their best so far.
<%BANNER[article]%>And really, there are enough direct and indirect causes that can hinder the expansion of i925/i915 chipsets. For example, these chipsets work with DDR2 SDRAM, which is available in limited quantities today. They also work with graphics cards with the PCI Express interface and such cards have been rare so far. Then, many users are repelled by the “childhood diseases” of the new platform, particularly with overclocking problems that have already been much discussed. Due to these reasons, Intel’s new platform has a smaller appeal towards the user; many mainboard manufacturers predict a steady demand only at the end of the summer.
I should also mention such an unpleasant aspect of the new chipsets as their limited lifecycle. Revised versions of them, more advanced than the i925/i915 (for example, with support of the 1066MHz FSB), are scheduled to appear in this fall. And in the second quarter of the next year the i925/i915 family will be followed by absolutely different chipsets known now by their codenames: Glenwood and Lakeport.
Anyway, the new chipsets from Intel will remain the most advanced and feature-rich products for the Pentium 4 processor for the time being. At the same time, the old and tested i875/i865 family will most probably live alongside the newcomers for a while. Moreover, our tests suggest that the transition to the i925/i915 platform doesn’t bring any perceptible performance gains today.
That’s why mainboard makers, trying to meet the needs of different categories of users, are going to roll out a wide selection of new solutions, both on the i925/i915 family chipsets and on the i875/i865 logic. As you know, Intel intended the new chipsets to be accompanied with a new CPU form-factor called LGA755. In fact, the reference design of i925/i915 mainboards implies installation of the new socket. However, the manufacturers that can design their products independently – ASUS, MSI, Gigabyte, ECS and others – are going to offer combo solutions: i925/i915-based mainboards with Socket 478 as well as i875/i865-based mainboards for LGA755 CPUs. Thus, the user is likely to find himself/herself confused seeing so many choices.
With all these considerations in mind, and before posting reviews of fresh Pentium 4-supporting mainboards on our site, we decided to compare the performance of the older i875/i865 against the new i925/i915 chipsets under equal conditions. This comparison will be a kind of foundation or reference point for our upcoming reviews. Well, and I’m simply curious to know how much the i925 differs from the i915 in practice!
Let’s look for any differences in the formal characteristics of the investigated chipsets:
So, without taking the different packaging and positioning of the chipsets’ North and South Bridges into consideration, the new i9xx series has the following distinguishing traits:
We already discussed the characteristics of the new chipsets, with emphasis on the i925, in our article called LGA775: New CPUs and Chipsets . Today I’m going to dwell on the i915 that targets the mass market.
The i915P Express is positioned by Intel as a mass-market discrete chipset with support of PCI Express and DDR2 SDRAM. Thus, it costs much less than the top-end i925X Express – we saw a similar difference between the last-generation chipsets (the top-end i875P and the mainstream i865PE). What features will be unavailable for the users of i915P-based mainboards?
The i915P Express works with the memory slower than its top-end mate. As you know, the i875P featured Performance Acceleration Technology (PAT) that improved the performance of its memory controller, and the i925X also has a number of optimizations, missing in the cheaper i915P chipset.
The architecture of the Intel 925X Express chipset allows for a higher speed due to an improved pipelined scheme of working with the memory and a more efficient use of each memory channel. This technique is different compared to PAT. Instead of optimizing the data path from the chipset to the memory and back, Intel chose another strategy. Particularly, the i925X places memory regeneration (maintenance) commands into the standard data stream and regroups the data in the memory to optimize the access.
You may remember that mainboard manufacturers used to “outsmart” Intel and enable PAT in their i865-based products, but now this trick seems impossible. At least at the moment i915P-based mainboards are always slower than their i925X-based analogs under any conditions, while the manufacturers themselves deny the possibility of improving their i915P products to the level of the i925X ones.
Besides the higher speed, the memory controller of the i925X chipset boasts ECC support – the i915P doesn’t. However, this is no great loss. Firstly, ECC is only needed in servers and high-performance workstations and, secondly, an error in the current revision of the i925X leaves ECC only as a formal capability, which you cannot enable in practice yet.
I have been talking about the features that the memory controller of the i915P lacks, but it has something that the i925X has not – compatibility with ordinary DDR SDRAM. Well, it doesn’t mean that i915-based mainboards can work with DDR2 and DDR SDRAM simultaneously. The i915 series chipsets use the same group of pins for communication with both types of memory, so the mainboard can work with only one memory type at a time. Intel warns the manufacturers against developing mainboards with DIMM slots for both memory types – this may cause stability-related problems. That is, the manufacturers are offered to choose a particular memory type for each particular mainboard. Well, you couldn’t have expected anything more from Intel – the company’s support of the older memory type is already a nice display of care about the users.
Then, I should note that the memory controller of Intel’s new chipsets gives more configuration flexibility than the one in the i875/i865 series. As a result, i915P-based mainboards allow you to enable dual-channel memory access (both with DDR SDRAM and DDR2) more often i865-based ones.
Thus, thanks to the realization of Flex Memory Technology, the most advantageous symmetrical dual-channel memory access mode can be used in more cases than with the last-generation chipsets, which required the use of the same number of same-organization modules in each channel. Now, the i925/i915P chipsets only require that you install the same amount of memory into each channel. That is, even if you use one memory module in one channel and two modules in another, you can still enable the symmetrical dual-channel configuration. However, if the amount of memory per channel is different, the performance goes down to the level of single-channel chipsets – keep this fact in mind!
There are no other differences between the i925 and the i915 except those in the memory controller. However, this alone was enough to make the two North Bridges pin-incompatible, although they use the same packaging.
A great drawback of the current i925/i915-based mainboards in comparison with boards on the older chipsets is in the problems you encounter trying to overclock the CPU. The problems take root in the support of the new peripheral bus (PCI Express), which is clocked not like the PCI and AGP used to be. Many i925/i915-based mainboards offer you the option of setting the carrier frequency of the PCI Express bus independently in the BIOS, but this option turns to be ineffective.
As a result, increasing the FSB clock rate (and that’s the way you do overclocking) also leads to a higher PCI Express frequency, causing numerous problems. Particularly, PCI Express x16 graphics cards are the first to stop working at overclocking, since many of them are fastidious about the clock rate of the bus. Then, don’t also forget that Intel uses a bus called Direct Media Interface (DMI) to link the North and South Bridges of its new chipsets, and this bus is in fact an analog of PCI Express x4. Thus, the growth of the FSB frequency on modern mainboards provokes an appropriate growth of the clock rate of the PCI Express x16 bus as well as of the DMI bus and then of all the busses that are realized through the South Bridge. That is, the clock rate of PCI Express x1 grows and the frequency at which all IDE controllers work increases, too. That’s why you may render your hard disk drive (or any other peripheral) non-operational trying to overclock the CPU on a mainboard based on Intel’s new chipset.
Thus, it’s not so easy to do overclocking with the new chipsets. In fact, average mainboards on the i925/i915 start having problems at FSB frequencies of about 220-230MHz. On this background, the i925X-based products from ABIT and ASUS stand prominent – they achieve better results at overclocking. The engineering teams of these two companies found a way to control the multiplier of the carrier frequency of the PCI Express bus. As a result, you can raise the FSB clock rate much higher at overclocking – the record is about 280MHz. To achieve this result, you need to use special components, though. For example, graphics cards on ATI’s GPUs only.
Well, the overclocker shouldn’t spurn i925/i915-based boards yet. Yes, they are not as good at overclocking as their i875/i865-based predecessors, but they may improve. Speaking about the impossibility of independent clocking of the PCI Express bus on modern i925/i915-based mainboards, the engineers forget to add “it’s impossible to do that by modifying the mainboard’s BIOS”. I think the overclocking problems may be solved with a hardware modification of the mainboards.
So, it is possible that i925/i915-based mainboards will be perfected within the next few weeks to become fully suitable for overclocking. As a minimum for that, the R&D departments of the mainboard makers should do some work plus new clock generators should come out that would support independent clocking of PCI Express. Right now, they use the ICS954119 generator, which cannot clock the PCI Express bus asynchronously.
I’m inclined to put the overclocking problems into the category of “childhood diseases”, rather than fatal errors of the new chipsets. However, I can now only give two recommendations to overclockers: wait for the second wave of mainboards on the new chipsets or go for the i875/i865.
This testing session is about the performance of i925 and i915 chipsets with different memory types in contrast to that of i865/i875 chipsets. I took the following mainboards for my tests:
Intel D925XCV. This mainboard is based on the Intel 925X Express chipset. I used this mainboard to examine the performance of this chipset with dual-channel DDR2-533 and dual-channel DDR2-400 memory.
Intel D915PBL is based on the Intel 915P Express; it will help us estimate the performance of the chipset with DDR2 SDRAM.
ABIT AG8. ABIT made its i915P-based mainboard compatible with DDR SDRAM rather than with DDR2. This mainboard will help us evaluate the performance of the Intel 915P Express chipset with dual-channel DDR400 SDRAM.
ASUS P4P800-E Deluxe. This is an i865PE-based mainboard, but it matches the performance of i875-based products due to ASUS’ exclusive HyperPath technology. Note also that this mainboard can work with DDR533 SDRAM with the processor operating at its rated frequency, which makes it a most exciting object for comparison with the i925/i915-based boards.
I used the following hardware for my tests:
I performed the tests in Windows XP SP1 with DirectX 9.0b installed.
There are surprises in the first test from my program already. Futuremark’s PCMark04 says that the last-generation i865PE chipset works faster than the new i925/i915, irrespective of the memory type you use with them. The advantage of the i865PE is rather solid and you cannot explain it by the use of DDR SDRAM with its lower latencies since the i915 found itself on the bottom of the table working with DDR400!
This is the same benchmark, but with a set of tests that are very slightly affected by the speed of the memory or any other system component. The new and expensive i925X with DDR2-533 memory finds itself on the very top, followed by the i915P with the same memory. Note also that the i915P has a nice score with DDR400 SDRAM. The older i865PE chipset doesn’t impress here: it is at the bottom of the table irrespective of the memory frequency.
Business Winstone 2004 puts the new chipsets on top, too. This time the result is easy to explain: the SerialATA-150 controller in the ICH6 South Bridge that the new Intel chipsets come with boasts a higher performance than the ICH5. This is enough for the new chipsets to score good results, notwithstanding the high-latency memory used with them.
Multimedia Content Creation Winstone 2004 is most sensitive to the memory subsystem bandwidth. That’s why we see the new chipsets from the i925/i915 series getting to the top positions in the diagram. As for the i865PE, even DDR533 SDRAM cannot help it here: this chipset doesn’t communicate with this memory in the best way possible, and it also works less efficiently with the disk subsystem
This memory test produces most exciting results: the i865PE is the fastest here! That is, the low latencies of DDR SDRAM help this chipset to outperform the new i915/i925. Note also that DDR533 SDRAM permits to get even better results, although the i865PE hadn’t been originally designed to support this memory type. This test is also indicative of the less careful implementation of DDR400 SDRAM support in the i915P – this chipset works much slower with this memory than the i865PE.
Another synthetic memory test from the SiSoftware Sandra 2004 suite produces similar results.
It is not only in synthetic tests that the last-generation chipsets can take the main prize in. Once again we see the team of the i865PE and DDR533 SDRAM on top in the popular gaming benchmark 3DMark2001 SE. The memory subsystem in this combination evidently has a low latency and a high bandwidth.
I should remind you, however, that such a work mode is not normal for this chipset and not all i865-based mainboards allow clocking memory at 533MHz with the FSB frequency remaining at 200MHz. As for the i865PE’s normal operational mode with dual-channel DDR400, it is a bit slower in this case than the i925X with DDR2-533 SDRAM, and is practically similar to the i915P that works with DDR400 SDRAM.
The i865PE looks less attractive in the newer version of 3DMark. According to this test, DDR2 SDRAM is always faster than simple DDR. The i925X with DDR2-533 are the winning duo, followed by the i915P with the same memory. Probably, 3DMark03 enjoys the new PCI Express x16 graphics interface that has a much higher bandwidth than AGP 8x.
The results of the CPU test are somewhere between the results we got in the two versions of 3DMark. Curiously enough, the i865PE and the i915P are far behind the rest of the participants when working with DDR400 SDRAM.
The results in Quake 3 are similar to those of 3DMark2001 SE. The last-generation i865PE chipset wins once again, but we see that Quake 3 needs low memory latency first of all, so the system with DDR400 SDRAM turns to be faster than itself with DDR533 SDRAM. By the way, the i915P works with the same DDR memory much worse.
Unreal Tournament 2003 draws a similar picture of comparative performance. Note, though, that the i915P works somewhat faster with DDR400 SDRAM and even leaves the i925X with DDR2-533 behind.
Newer game engines are more critical to the memory subsystem bandwidth. Aquamark 3 gives out quite surprising results: the winning combination in many tests, the i865PE with DDR533 SDRAM, is at the very bottom of the results table, while the same chipset with DDR400 SDRAM only outperforms the i915P with the same memory subsystem. According to Aquamark 3, the winner is a system with the i925X chipset, DDR2-533 SDRAM and a graphics card with the PCI Express x16 interface.
Everything returns to the norm if we exclude the graphics subsystem. The low results of the i865PE in the previous test are most likely due to the low bandwidth of the AGP 8x bus compared to PCI Express x16.
The speed of encoding MP3 files depends but slightly on the memory type and the chipset – it is the processor that handles the whole load.
The same goes for the speed of encoding video into MPEG-2 format. However, we can discern a certain advantage of the i925X over the i865PE.
The use of low-latency memory accounts for better results in the test of encoding video into MPEG-4 format. The i865PE takes the two top positions here.
The i865PE with DDR400 SDRAM is again on top in Windows Media Encoder 9. However, working with DDR533 SDRAM, this chipset lands right on the bottom of the table.
It is a well-known fact that archivers first of all need a low-latency memory subsystem. That’s why the oldie i865PE feels no competition in WinRAR. The use of DDR2 SDRAM with the i9xx series chipsets leads to a dramatic performance hit in data compression applications.
We see absolutely the same picture in another archiver, 7zip.
The inverse process – decompression – doesn’t put any serious load on the memory subsystem. The speed only depends on the arithmetical capabilities of the CPU.
The final rendering in 3ds max is another purely processor-related task – the memory subsystem doesn’t practically affect the speed. However, the i865PE fell somewhat behind the new chipsets.
Unfortunately, I cannot call the results in SPECviewperf 7.1.1 correct. The strange scores in this test are probably due to the imperfect drivers of the PCI Express x16 graphics cards. If you’re a user of professional OpenGL applications, you may want to pay attention to this fact.
I will now try to summarize. First of all, the new chipsets from Intel, although rich in new and exciting functions intended for a performance growth in the future (DDR2 SDRAM, PCI Express x16 graphics bus), are no better yet than their predecessors. Across a number of tests, the i865PE with DDR400 or DDR533 memory outperforms any of the new chipsets.
By the way, I also couldn’t see any significant advantages of the top-end i925X over the mainstream i915P. Although these products belong to different price sectors, and the i925X features certain memory controller optimizations, they both have a similar speed working with the same memory.
As you see in the diagram, the i925X is only 1-2% faster in average than the i915P, to 5% at the maximum. Well, you may remember that the mainboard manufacturers eliminated all difference between the i875 and i865, and this didn’t affect the i875P sales. We’ll probably see the same again. Although the i925X and i915P are roughly the same, the former will be used in more expensive PCs.
However, if price/performance ratio matters to you, you shouldn’t go for the i925X. As for the i915P, we shouldn’t forget about its support of DDR400 SDRAM. Intel took this support seriously – it is not just a formal feature. The performance of a system with the i915P and DDR400 memory doesn’t differ much from that of analogous systems with DDR2-533 SDRAM:
It should be noted, however, that the last-generation chipsets are more efficient with DDR SDRAM – the i865PE has an advantage over the i915P nearly everywhere:
This is partially explained by the fact that the memory controller of the i925/i915 is oriented at the asynchronous operational mode. The transition to 533MHz memory doesn’t lead to any considerable gains for the i865PE, while the i9xx series chipsets receive a nice boost by transitioning from DDR2-400 to DDR2-533, notwithstanding the high latencies of the latter memory type.
Well, the first feeling the new chipsets from Intel provoke is that of disappointment. These chipsets cannot reach new performance peaks except when the application is actively using the graphics bus. Adding the yet-unsolved overclocking problems, the i925/i915 seem to have no performance advantages from the user’s point of view. Yes, sometimes they win a test but the overall impression is anyway bad. The rather low speed of the new Intel chipsets is somewhat compensated by such features as an improved IDE controller, high-quality sound, numerous USB ports and so on. They are of little consolation, though, especially considering that the i925/i915 family will be soon replaced with new and, hopefully, faster chipsets with support of the 1066MHz FSB, and then by the Glenwood/Lakeport family.