Intel has sustained almost a three-year pause since it launched its revolutionary family of i9xx chipsets. This family added the following features to desktop systems: a new socket type and power connector, DDR2 memory, PCI Express bus (including a modification for graphics cards), and High Definition Audio. The following two generations of chipsets (i945/955/975 and i965) just increased FSB and memory clock rates and added support for new CPU families (dual-core processors and then Core 2).
Now we meet a new generation of chipsets, which update such important system characteristics as a system bus and memory type, as well as bring a new naming scheme.
Intel X38 Express
We'll start the analysis of the new chipset family with its top representative. It's not yet available in the market and will debut only in Q3 together with the second wave of the new chipsets. Note that top products used to be marked with an increased index number (i915 - i925). And now the top product is marked with the X prefix - Intel uses it to designate any general improvements (not only in chipsets, but in processors, graphics cards, etc). The following block diagram shows the key characteristics of the X38:
Let's enumerate the key functions of its Northbridge:
We can see well that all key characteristics of the chipset have been modified. Let's analyze all innovations one by one.
CPU support. It must be noted here that all 3-series chipsets officially don't support Celeron D, Pentium 4, and Pentium D processors (as well as their Extreme Edition modifications). The lack of support is caused not by the modified characteristics of the processor bus, but by the new motherboard standard - FMB (to be more exact, by VRM), which supports future 45 nm processors, instead of 90 nm (and older) models. There is no direct connection between this chipset and its power supply subsystem, of course. But manufacturers almost always follow Intel design standards. So it's highly improbable that we'll see many Intel 3x models with support for "pre-Core 2" processors. To say nothing of motherboards supporting both Prescott and Penryn.
What concerns the Core 2 support, X38 does it perfectly: this chipset will officially support all existing and future models of Core 2 Duo, Core 2 Quad and Core 2 Extreme processors (including quad-core modifications) with a 1333 MHz bus. All low-end families of the new processors (Celeron 400 and Pentium E2000) will work with the X38. But the top chipset will not support Celeron 400 for marketing reasons.
Memory support. Features of the DDR2 controller have not changed in all new chipsets (no alterations are in view here, all features described in the specifications are already implemented). But motherboards based on Intel 3x will work with DDR3 memory as well. Peculiarities and theoretical performance of the new memory type are already analyzed in our review. This article will touch upon practical aspects only. Here comes the first traditional question - is simultaneous support for DDR2 and DDR3 possible? The situation is not different from DDR and DDR2: Intel does not officially test such combinations, but motherboard manufacturers can do it on their own. Our constant readers have surely heard about several models of combo motherboards. Besides, our today's tests were run on one of such models (but we'll hardly see X38-based combo models). DDR2 and DDR3 memory modules cannot work simultaneously, of course: the motherboard will initiate memory modules of only one type.
DDR3 is very good: it features lower heat release (memory voltage is reduced, so that even DDR3-1066 will dissipate less heat than DDR2-800); a different location of the key notch will not let you mix up DDR2 and DDR3 memory slots on combo motherboards. As you already know, DDR3 is supposed to work at up to 800(1600) MHz. X38 will support one of the fastest modifications - DDR3-1333. But the situation with availability and timings of memory at the time of Intel 3x debut is not very good. DDR3 modules are not available in the market on a mass scale. Even elite manufacturers (such as Corsair) take advantage of this situation and sell modules with mediocre characteristics at fabulous prices. Rational readers are recommended to wait until their prices go down and their characteristics are improved. As for now, analysts forecast that DDR3 will take up 50% of the market only by 2009. It will hardly capture 10% by the end of 2007. And in the practical part of the article we'll see what we are offered at such a high price.
PCI Express 2.0. Intel strikes a pre-emptive blow here. It has finally designed a chipset with support for two full-speed PCI Express x16 interfaces, which have been available in top products from competitors for a long time already (this configuration does not have a noticeable advantage in most cases, but principles are more important), and has implemented a host controller of the second version. PCI Express 2.0 still allows to use old graphics cards, because it has the same slots, compatible both ways. The new features of PCI Express 2.0 will hardly be very interesting for the graphics interface, except for two points. Firstly, performance of each PCI Express lane is doubled, so a single lane (PCIEx1) has a bandwidth of 500 MB/s each way. The total bandwidth of a 16-lane PCIEx16 interface is 16 GB/s. I repeat, there will be no practical advantage of this system in the nearest future.
Secondly, the power applied to the bus is also doubled: PCIEx16 slot of the first version provided up to 75 W, now a graphics card can get 150 W. (I wonder how these additional Watts get into the bus - will X38-based motherboards have a special additional power connector?) However, top accelerators neglect the power supply from PCIEx16. You can use one of the two on-board power connectors at best. It's similar to the situation with PCI Express graphics cards with on-board power connectors (75 W from the bus turned out to be insufficient) that appeared with the launch of i915/925-based motherboards. However, we should thank SLI/CrossFire for that: top graphics cards are first of all designed for tandem configurations - while a single graphics card may theoretically work on the power from the bus, the second graphics card without an on-board power connector will not startup. What concerns joining graphics cards on Intel X38, everything is the same: CrossFire is officially supported. SLI is not supported and won't be in the nearest future.
X38 will come with a new south bridge of the ICH9 family. Functionality of this family will be examined below.
Intel P35 Express
This chipset belongs to the Performance group. The workhorse of the new series has the following architecture:
Let's enumerate the key functions of its Northbridge:
It has fewer new features, DDR3 is the most important of them. Support for processors is limited to the same models manufactured by 65-nm and future 45-nm process technologies. But for the above mentioned reasons (P35-based motherboards are supposed to have a simplified FMB design) Core 2 Extreme models (especially quad-core processors) will not work in P35-based motherboards. The chipset also lacks support for DDR3-1333 memory (it does not have a proper frequency demultiplier). The standard PCI Express x16 interface (of the first version) is used instead of PCI Express 2.0. Like P965 and older chipsets, the P35 does not allow to configure this interface flexibly for CrossFire. However, this fact has never stopped motherboard manufacturers - they design P35-based CrossFire solutions by adding the second slot to the south bridge (connected to PCIEx1 peripheral interfaces). This chipset uses an ICH9-family south bridge.
Intel G33 Express
The main integrated chipset of the new family bears an unexpected name - G33, while its functionality is on a par with the P35. That's because Intel will launch another integrated chipset in Q3 (G35) with an improved graphics core, and its index must differ from the top X38. So, the G33 is a modification of the P35 with an integrated graphics core with the following architecture:
Let's enumerate the key functions of its Northbridge:
I repeat, the only difference of this chipset from P35 is its integrated graphics core.
GMA X3100 integrated graphics processor. Let's hope that video drivers for X3100 will be ready promptly, and we'll finally see everything that was promised to us since the times of X3000 (G965). The new graphics core actually suffered no serious changes. We'll examine the differences, when/if we can run all the tests. Clear Video technology is intended for hardware acceleration and improvement (deinterlacing + color correction) of video playback (including HD). It should also offer digital video interfaces (including HDMI) for video output. We are promised full support for the Aero interface in Windows Vista, of course. GMA X3100 also officially supports HD DVD and Blu-ray playback. We'll analyze this issue after tests of G33-based motherboards.
Intel G31, G35, Q35 Express
Let's say a few words about the other chipsets of the new family. They will all be launched in Q3 2007.
G31 is an entry-level integrated chipset. It's a new chipset only at a stretch. In fact, its functionality is on a par with 945G chipsets, which it's intended to replace. This chipset even uses the old ICH7/R south bridge - it also solves the problem with PATA support, which has been troubling the corporate sector since the times of ICH8. G31 supports Core 2 Duo (but its FSB clock rate must not exceed 1066 MHz) and up to DDR2-800 memory.
G35 is an interesting integrated chipset with an overhauled graphics core. According to Intel, it will become the first [integrated] solution to support DirectX 10. We'll describe the G35 (and its GMA X3500) in due time, of course. In other respects, the G35 promises to be very much like the G965. It will be similar to Intel 3-series chipsets only in its support for 45-nm Wolfdale, Yorkfield, and new Core 2 Duo processors with 1333 MHz FSB (DDR3 is not supported either). The G35 will use the old ICH8/R/DH south bridge.
Q35 (and its simplified modification Q33) will be used for Intel vPro business systems, it's an integrated chipset with disabled gaming features. The most interesting combination will be the Q35 and the ICH9DO (Digital Office) south bridge, which will support such technologies as AMT (Active Management Technology) 3.0, Trusted Execution Technology, and Virtualization Technology. Q35 does not support DDR3 memory either.
Intel ICH9 south bridges
The new chipsets will use updated south bridges. Just like north bridges, ICH9 has a number of evolutional improvements relative to ICH8. It also supports (ICH9R only) one revolutionary technology. Let's briefly enumerate the key characteristics of the new south bridge family:
ICH9R traditionally differs from ICH9 in RAID support as well as two extra SATA ports. Special versions of ICH9DO (Digital Office) and ICH9DH (Digital Home) South Bridges are based on ICH9R. But the former additionally offers Active Management Technology 3.0, Trusted Execution Technology, and Virtualization Technology. And the latter offers Viiv Technology (positioning of both modifications is obvious).
What concerns minor evolutional changes, we can mention the increased number of USB 2.0 (12), eSATA and port splitters (which is important for external eSATA connectors) for chipset-based SATA ports; besides SATA connectors (as well as USB ports starting from ICH8) support the hot plug feature. Intel Rapid Recover Technology may become an alternative to RAID. It allows to create a disk image on another drive, quickly update it ignoring unchanged files, and quickly restore data, if the first hard drive is damaged. South Bridge still has an integrated Gigabit Ethernet MAC, but we haven't seen it used in any i965-based motherboard - it's probably cheaper to use a network controller from Marvell, Broadcom, Realtek, and the like, connected via the PCI Express bus, in usual desktop systems. At the same time, vPro users will surely like functionality of the Intel controller. It would have been strange to expect the return of PATA support after its disappearance in ICH8. And it hasn't happened. Intel closed this issue despite a lot of problems with PATA replacements.
The most intriguing feature in the new series of south bridges is Intel Turbo Memory support (which was also known as Robson Technology in the process of its development). It consists in installing a module with some NAND flash memory (512 MB and 1 GB modifications at first). To all appearances, this module will mostly be installed into a PCIEx1 slot, although there are other options (for example, to the connector for the external USB port). Turbo Memory will benefit Windows Vista users. Unlike USB flash drives, the integrated module can be used by the new Microsoft OS for ReadyDrive and ReadyBoost.
In the first case, we get an opportunity to use the flash drive as HDD cache - linear read and write operations will not gain much here (flash memory is slower than a hard drive), so ReadyDrive will be useful for a regular exchange of small portions of data, which is typical of reading/updating a swap file (flash memory has a noticeably lower access time than a hard drive). Additional advantage consists in the reduction of HDD access events (data are written to a hard drive in packets, when it's idle; what concerns reading data, HDD is not accessed at all if the necessary data are available in Turbo Memory cache), which contributes to power saving - this is a real advantage only for mobile devices.
ReadyBoost expands the available memory volume for preloading and caching data (from a hard drive). Although flash drives cannot compete with system memory in speed, reading data from flash memory with its low random access time rather than from HDD makes it noticeably faster to load applications and files (up to twofold). The problem with Turbo Memory is potentially short service life of flash storage drives. The best models guarantee one million rewrite cycles (probably several millions), so a flash drive may lose its capacity long before your computer ends its service life.
Heat Release. Heat release of the new chipsets is another issue worth mentioning. Even though they are manufactured by the same 90-nm process technology and have a more complex logic, the 3-series chipsets consume noticeably less power than their predecessors: P35 - 14.5 W (P965 - 19 W), G33 - 16 W (G965 - 28 W). Moreover, the difference is so evident that you can easily feel it by touch, although P35-based motherboards do not use larger chipset heatsinks. That's a nice trend, because after hot i965 and scorching nForce 600i we might have expected new standards for chipset cooling.
As we have received several MSI motherboards based on the P35 chipset, including one motherboard that supports DDR2 and DDR3, as well as DDR3 memory modules from Corsair, our today's tests will help us answer two questions at once. First of all we'll compared DDR2 and DDR3 performance on the P35 platform. And secondly, we'll compare both modifications of this platform with other chipsets available on the market. We've selected P965 (to be replaced with P35) and the top chipset of the latest NVIDIA family - nForce 680i LT SLI (we've already found out that there is no difference between nForce 680i LT SLI and nForce 680i SLI in speed or functionality, and we have a motherboard based on nForce 680i LT SLI).
It turned out more difficult to compare two memory types, because presale BIOS versions of the MSI motherboards were not ready for DDR3: BIOS of the P35 Neo Combo did not allow to set normal DDR3 voltages (1.5 V) and timings (they were limited to the standard DDR2 scheme, so main timings couldn't be set higher than 6). And our Corsair modules did not work in DDR3-1066 mode with timings below 7-7-7, so we had to allow the motherboard to set timings by SPD. We also had problems, because the platform was new and various Windows utilities couldn't detect whether timings were set correctly (as well as other memory parameters). Fortunately, the latest version of CPU-Z already supports the P35 chipset and DDR3, so we managed to more or less clarify the issue.
According to CPU-Z, we had the following timings in DDR3-1066 mode (by SPD): 7-7-7-20. As the motherboard did not allow to raise main timings above 6, we used DDR2 memory at 1066 MHz with 6-6-6-18 timings to make the comparison as valid as possible. At the same time, we had no problems using our DDR3 modules at 800 MHz even with 4-4-4-12 timings, so we could compare this configuration with P965 and nForce 680i LT SLI in their standard DDR2-800@4-4-4-12 mode. As we did not have a P965-based motherboard at hand (to run our memory in DDR2-1066 mode), previous generations of chipsets in this mode are represented only by the NVIDIA product (it was almost on a par with the i965 in our tests).
And now before we proceed to test results, let's consider the issue from the theoretical point of view. DDR3 cannot be much faster than DDR2 in equal conditions (at the same frequency and the same timings). Our main hopes on memory performance gain can be pinned only to modes with reduced timings and high frequencies. Indeed, DDR2-800@4-4-4-12 and DDR3-1600@8-8-8-24 demonstrate the same absolute timings. So if memory manufacturers manage to make low-latency modules, DDR3 may become more efficient even in "usual" conditions.
The second possible advantage of DDR3 memory consists in its increased bandwidth, so this memory can operate at higher frequencies. Unfortunately, this performance gain may appear only with future processors, as bandwidth of the 1066 MHz FSB is just ~8.5 GB/s, which corresponds to memory bandwidth of dual-channel DDR2-533! In practice, raising memory clock rate "by one step" in such cases can usually gain some performance, but even DDR2-800 is actually more than enough even for future processors with 1333 MHz FSB. And the existing processors need neither DDR3-1066 nor DDR3-1600.
We'll traditionally start with a low-level analysis of memory potential in our RightMark Memory Analyzer.
Even though the NVIDIA chipset and the i965 demonstrate close results in real applications, they still look very different in the synthetic RMMA test. So let's not pay much attention to this difference.
P35 is noticeably outperformed in read rate by both competitors, but it demonstrates an interesting phenomenon: when memory (both DDR2 and DDR3) operates at 1066 MHz, its performance is higher than in DDR2-800, although nForce 680i LT SLI demonstrates lower results. OK, let's set this fact aside, it does not agree well with our theories. Let's proceed to other comparisons. Now we can only note that DDR3 looks noticeably worse than DDR2 even with the same timings. We deliberately omit the exact value of this difference, because it would be too early to evaluate percentage differences before we ran real tests.
When we test the write speed, we are not interested in peak results with non-temporal store, because they will be identical on processors of the same architecture. The situation with a real memory write rate is similar to that with reading: the new chipset is noticeably slower than its competitors, DDR3 is slower than DDR2 (especially at 800 MHz), and setting memory frequency to 1066 MHz again accelerates the P35 with both memory types, but it slows the NVIDIA chipset.
Finally, the memory latency test. And here goes the first surprise - P35 memory controller has a technology similar to DASP in NVIDIA - it radically reduces latencies (manifold) for pseudo-random reads from memory (within a single page). We definitely have the same caching buffer with prefetch here. Nevertheless, P35 is significantly outperformed by competitors (in this case it's nForce 680i LT SLI) even in such a favorable test of pseudo-random reading from memory. The new memory type is again victorious in the competition between DDR2 and DDR3 on the P35. The difference is especially noticeable in DDR2/3-1066 mode, where DDR3 has higher timings.
Interestingly, raising memory frequency to 1066 MHz again accelerates the system, although the ratio between absolute timing values must have led to the opposite: considering the cycle time, CL4 in DDR2/3-800 corresponds to 10 ns, while CL6 in DDR2-1066 - 11.25 ns (not to mention CL7 in DDR3-1066 - 13.13 ns). Why so? Two possible explanations come up to our mind. First of all, note the matching clock rates of the Core 2 Duo E6600 bus and DDR2/3-1066 memory: perhaps, this synchronous mode provides some advantages. However, no such effect in the NVIDIA chipset suggests an idea that there are also some internal optimizations in the memory controller, which provide some performance gain at higher memory frequencies just like in the i965.
And now let's proceed from theoretical aspects to real tests, and then use test results to evaluate advantages of various configurations.
So, we can draw first conclusions from the real results. On one hand, all mentioned ratios are preserved: the P35 is slightly (up to 7%) outperformed by the P965 and nForce 680i LT SLI; DDR2-800 is faster (by 3%) than DDR3-800 with identical timings on the P35; DDR2/3-1066 on the P35 is faster than the same memory type at 800 MHz (more accurate estimates are impossible here, because DDR2 and DDR3 have different timings), no concessions to much higher timings. On the other hand, the 7% difference can be seen only in one test. P35 does not favor DDR2-800, that's evident. The difference is made even less evident by the fact that DDR2-800@4-4-4 is memory with almost the lowest latencies, while DDR3-1066@7-7-7 is the standard solution. Corsair will certainly offer an alternative with much lower timings soon.
But let's not jump at conclusions, have a look at results in other tests.
We cannot expect surprises from our video encoding test (here is our open test procedure). All contenders look the same here, as usual, because the limiting factor here is a CPU.
What concerns professional 3D applications SPECviewperf, only NVIDIA chipsets perform well here. These results probably have to do with their optimized graphics bus controller, because different memory modes (and even different memory types) affect performance purely nominally.
We don't see anything new in games either. We can only mention that the P35 with 1066 MHz memory becomes an absolute winner in one of Doom 3 modes (for the first and the last time in our today's tests). However, the difference between chipsets in Doom 3 is not big, it does not exceed 3%, and performance losses because of using DDR3 instead of DDR2 on the P35 are even lower - about 2%. The spread of results in FarCry is a tad larger, up to 4%. But all three tendencies mentioned today hold true.
It's difficult to evaluate chipsets that introduce several revolutional features at once. In this case the announcement is smoothed over, because PCI Express 2.0 will appear only in Q3 together with the X38, and the upgrade to the new version of the standard is expected to pose no problems with compatibility. The second innovation, DDR3 memory, did not produce a strong impression on us with its performance. But fortunately, at least the first generation of these chipsets will offer a choice between DDR2 and DDR3, so you can wait till the prices for the new memory go down and its performance will grow. Support for new processors is perhaps the trump card of Intel 3x chipsets. But by the time these new processors are available, they may be also supported by other chipsets, including competing products. At least nForce 600i series chipsets will support 1333 MHz FSB, while no one can declare real support for 45-nm models yet. The new south bridge is quite progressive, it adds some new features. But its main innovation, Intel Turbo Memory, must be tested before we draw any conclusions.
Before we evaluate performance, we should note that we are waiting for a confirmation to the performance level demonstrated by MSI motherboards. Indeed, all our three motherboards demonstrated an absolutely identical performance level with DDR2 (two of them support only this memory type). But when our tests were almost completed, we got a new firmware version for the P35 Platinum, which slightly increased its performance (by several percents). Besides, although we cannot assert that combined solutions are outperformed by "dedicated" models, we are still anxious about it. So it's too early to close the DDR3 performance issue. And if we consider inexpensive (that is mass-scale) processors with a 1333 MHz FSB, the situation may change. Nevertheless, it would have been silly to run so many tests and skip conclusions. Here are our conclusions: considering the above mentioned provisos, chipsets of the new series look slower than the old ones (like i965 or NVIDIA nForce 600i); DDR3 memory may result in a 2-3% performance drop; the P35 works better with 1066 MHz memory regardless of timings.
What concerns global prospects of the new chipsets in the market, the X38 will certainly find its users (a few people who love top solutions), being one of the best chipsets in the market in terms of functionality. Having passed its startup period, the P35 should demonstrate decent performance. And its solid functionality, low heat release, support for promising processors and memory make motherboards on this new chipset a better choice than similar solutions from competitors and the old chipsets from this company. Turbo Memory technology may become another weighty argument for Intel 3x. We promise to examine integrated solutions in another article.
Sergei Pikalov (firstname.lastname@example.org)
July 11, 2007
Write a comment below. No registration needed!