Sign in with
Sign up | Sign in
PCI Express & CrossFire: Scaling Explored
By ,
1. Introduction

AMD’s CrossFire technology might be an excellent way to turn a mundane gaming machine into an FPS-shredding powerhouse, but with multiple Intel-based platforms supporting the feature across several different PCI Express lane configurations, not everyone is equally convinced. Putting aside the fact that many games only benefit from CrossFire at high resolutions and high-quality settings, many users are concerned about whether or not their motherboard can provide enough bandwidth to realize the full potential of CrossFire.

We hear questions like, “should I upgrade my motherboard first ?” and “CrossFire Upgrade or New Build” in our user forums, with a different set of answers each time. It’s time to put some data behind those responses.

Five Core 2 motherboards from ASUS, ready for testingFive Core 2 motherboards from ASUS, ready for testing

With Core 2 chipset development halted in light of Core i7 and recent worldwide financial events causing many buyers to re-evaluate their spending habits, now is the perfect time to analyze how CrossFire scales on various chipsets as a guide for those looking to enable the best possible performance at the right price.

Digging deep into our hardware stash, we found Core 2-compatible motherboards going all the way back to the venerable 975X chipset, along with every generation of Intel LGA775-based motherboard since.

Of course, we had to pick a starting point, so today’s article brings with it all the prior upgrades that owners of high-end systems up to two years old could reasonably be expected to have purchased, such as Intel’s fastest Core 2 Duo processor and four gigabytes of high-speed memory. Follow along as we detail each build and attempt to determine which motherboards are suitable for CrossFire upgrades and which are better retired to platform heaven.

2. Every Generation Of Core 2 Chipset?

Intel 975X

The 975X chipset was ahead of its time, simply because this high-end Core 2 platform burst onto the scene long before the CPU micro-architecture it was designed to complement. Last-minute changes to the Core 2’s specifications caused most early 975X boards to support only previous-generation NetBurst-class processors, and many updated Core 2-compatible models required further modification to support Core 2 Quads. A few manufacturers even delayed the release of 975X-based parts for up to eight months while the new processor architecture was sorted out. With all of the compatibility challenges that faced high-end 975X platform selection, most builders instead selected newer, lesser-featured P965 Express alternatives.

The big advantage Intel’s 975X had over the later P965 was its ability to support two graphics cards from the sixteen PCI Express 1.1 pathways connected directly to the northbridge. This meant that motherboards could support either a single card with full x16 transfers or two cards with x8 transfers. Mainstream chipsets were severely limited in how much bandwidth they could provide a second graphics card, so the 975X became a clear choice for CrossFire and early SLI testing (Nvidia later removed the SLI checkbox from its graphics driver whenever a non-SLI-branded chipset was used).

Asus was on the cutting edge of the change from NetBurst to Core 2-compatible platforms and its P5W DH Deluxe is one of the few 975X-based motherboards that has maintained compatibility with even the latest Core 2 products via simple BIOS updates.

We used P5W DH Deluxe BIOS version 2801 to achieve compatibility with the latest E0 stepping Wolfdale-based Core 2 processors. Though FSB-1333 is officially unsupported by the 975X, Asus added automatic overclocking to its BIOS to assure the processor would function normally, even though the chipset is overclocked.

P965 Express

The formerly-mainstream P965 Express was introduced around the same time as Core 2 Duos reached the market, with better overclocking capability and higher stable memory speeds, putting these motherboards at the top of the list for many enthusiasts. The biggest problem for high-end gaming systems is that this chipset was not designed to support CrossFire technology natively, as this was Intel’s way of differentiating the earlier 975X as its high-end solution.

With all sixteen PCI Express pathways locked together to support only a single graphics card, many motherboard manufacturers decided to use four PCI Express pathways from the new ICH8 series southbridge to supply a second x16 graphics card slot with x4 transfer mode. Though this type of bandwidth reduction is almost unthinkable now, those four pathways were enough to provide good performance to most mainstream graphics cards of the day.

The P5B Deluxe WiFi-AP Edition was one of the most popular P965 motherboards, which is why we reserved this one from our Core 2 motherboard shootout in preparation for today’s article.

Publicly-available BIOS 1236 wasn’t able to utilize 4 GB of memory and E0 stepping processors simultaneously, so Asus supplied us with a patched BIOS version 1238. Owners should feel free to contact Asus support if the BIOS they need isn’t publicly available yet.

P35 Express

Intel began preparing the world for its late-2008 release of Core i7 with the mid-2007 release of its P35 Express chipset. The P35 Express brought with it support for FSB-1333 processors and DDR3 SDRAM, though Intel left support of DDR2 memory intact.

What made the P35 Express special? Not much. Some users claimed it could overclock better than the P965, but that wasn’t necessarily true when comparing CPU speeds using later-model high-end P965 motherboards. The biggest boon to overclockers was that the DDR3 memory supported by some P35-based models allowed the bus speed to be increased beyond the 600 MHz of DDR2-1200.

With no practical performance advantages for using DDR3 with Core 2 Duo processors, we stuck to DDR2 for all our motherboards and chose the Asus P5K Deluxe WiFi-AP Edition to represent the P35 Express. Choosing a single motherboard brand should help to maintain consistent results when comparing different generations of chipset technology.

P5K Deluxe BIOS 0902 is the most recent available from Asus, yet it supported our E0 stepping Wolfdale core in spite of its June 19 release date.

X48 Express

Anyone who asks “why did you skip X38?” probably hasn’t read our X48 Express motherboard reviews, because the only difference between the two chipsets is that Intel approved X48 for 1,600 MHz FSB speeds. In fact, both the X48 and X38 have built-in support for FSB1600, also known as the 400MHz boot strap, since the X48 is, at best, a later stepping of the X38. Intel simply delayed the release of 1,600 MHz FSB-capable chipsets until it could validate enough X38 samples at the higher speed, slapping the X48 label on the best parts. On the other hand, a cynical observer could look at the difference between X48 and X38 chipset pricing and draw other conclusions about Intel’s intent.

The two big features Intel released in its X38 were both aimed at high-end graphics. First and foremost was proper support for AMD’s CrossFire technology, which had languished since the 975X. This was accomplished not by splitting sixteen lanes across two slots, but instead by adding sixteen more lanes for full x16 transfer mode to both cards.

In addition to the added pathways, PCI Express 2.0 transfer rates increased peak slot bandwidth by a factor of two. This resulted in the X38/X48 northbridge having four times the peak graphics data throughput of its 975X predecessor.

Because the X38 and X48 chipsets are functionally identical, we could have used either an X38 or X48 motherboard. Asus supplied us with its DDR2-supporting, X48-based P5E Deluxe.

Anyone who still believes that the X48 doesn’t support DDR2 while the X38 does has been hitting the Intel juice a little too hard, because—as we’ve said over and over again—they’re the same chipset design. Though Intel doesn’t want to promote the DDR2 capability of its X48 Express chipsets, X48-based DDR2 motherboards are also available from ASRock, DFI, Gigabyte, and MSI.

We used P5E Deluxe BIOS version 0205 (06/20/2008) to assure compatibility with our E0 stepping Woldale processor.

P45 Express

Is the P45 Express Intel’s last performance chipset? If we use the traditional description of what comprises a chipset, then probably. The memory controller of Core i7 platforms has been moved to the processor and future development of previous-generation designs will probably be limited to low cost, low power, or integrated-graphics markets.

The P45 northbridge brings back the one thing that made the 975X great—the ability to split its sixteen PCI Express lanes among two graphics cards. Bandwidth-doubling PCI Express 2.0 transfer mode gets the P45 in line with modern graphics needs, while the use of “only” sixteen pathways assures that motherboards can be more easily designed and manufactured. By combining lower manufacturing costs and relatively high bandwidth, the P45 became the first Intel chipset to realize the true potential of CrossFire as a mainstream/performance technology, rather than treating it as a high-end-only option.

We reserved the P5Q-Deluxe from last summer’s 11-Way P45 Motherboard Shootout for today’s comparison, updating it with BIOS 0803 for the best results using Intel’s recent E0 stepping Wolfdale-based processors.

3. Test Settings

Test Configuration
Component Base Settings
CPU Intel Core 2 Duo E8600, (3.33GHz, 6.0 MB Cache)
Overclocked to 4.00 GHz, FSB-1600
Memory 2 x 2.0 GB Crucial Ballistix PC2-6400
Set to 800MHz data rate, CAS 4-4-4-10
Graphics 2 x Gigabyte GV-R487-512H-B
HD 4870 GPU (750 MHz), GDDR5-3600
Hard Drive Western Digital WD5000AAKS, 500 GB
7,200 RPM, SATA 3 Gb/s, 16 MB cache
Sound Integrated HD Audio
Network Integrated Gigabit Networking
Power Cooler Master RS850-EMBA (850W, ATX12V v2.2)
Software and Drivers
Operating System Windows Vista Ultimate SP1
Graphics Driver AMD Catalyst 8.9
Onboard Device Drivers Motherboard Driver DVD

While many of our test games show enormous performance gains from higher clock speed, none have shown significant improvements in the transition from two to four cores. Thus, we used a dual-core processor that usually overclocks beyond 4 GHz on air cooling. Intel’s Core 2 Duo E8600 is probably the best gaming processor owners of LGA775 motherboards can hope for their motherboards to support because it starts out as Intel’s highest-frequency model and is easily pushed beyond the speeds other processors can reliably run.

Overclocked to 4.0 GHz at FSB-1600, our Core 2 Duo E8600 required good RAM to achieve optimal throughput. Unfortunately, the oldest chipset in today’s line-up wasn’t designed to support DDR2-1066, and using the DDR2-1000 setting resulted in severe instability. Without the ability to increase memory speed, we instead settled for tight CAS 4 latencies at DDR2-800 and 2.00 volts. Crucial’s Ballistix DDR2-800 4.0 GB dual-channel kit worked perfectly.

With the support hardware in place, it was time to pick graphics cards. Though the targeted upgrade market of this piece will likely favor lower-cost HD 4850 or HD 4830 based graphics models, the HD 4870 allows us to test the limit of motherboard differences when using the same architecture as those two high-value solutions. A pair of Gigabyte’s GV-R487-512H-B graphics cards suited our testing needs well.

To further examine graphics performance differences, we added Race Driver GRID to our regular selection of benchmarks.

3D Game Benchmarks and Settings
Crysis Version: 1.2.1
Test 1: Low Details, No AA
Test 2: Very High Details, 8x AA
Demo: CPU-Benchmark + Tom’s Hardware Tool
Race Driver:
GRID
Version: 1.20.0000
Test 1: Low Details, No AA
Test 2: Very High Details, 8x AA
Track: Jarama
Supreme Commander
Forged Alliance
Version: 1.5.3599
Test 1: Low Details, No AA
Test 2: Very High Details, 8x AA
Demo: WallaceTX_006_006
Benchmark: Fraps 2.9.4 - Build 7037
Unreal Tournament 3 Version: 1.2
Sound and DirectX10
Texture Details: 1
Level Details: 1
Demo: vCTF-Reflection_fly
Time: 12/60
World in Conflict Version: 1.0.0.9
Test 1: Very Low details (No AA, No AF)
Test 2: Very High details (4x AA, 4x AF)
Demo: Game-Benchmark

Sythetic Benchmarks and Settings
3DMark Vantage Version: 1.02
GPU and CPU scores
SiSoftware Sandra XII SP2 Version 2008.5.14.24
CPU Test = CPU Arithmetic / Multimedia
Memory Test = Bandwidth Benchmark

Because different generations of chipset support different bus speeds, we also included Sandra’s CPU and memory bandwidth benchmarks to gauge the effects chipset overclocking had on performance.

4. Test Results: Synthetics

It would be a huge mistake to skip over our synthetic results and jump straight to games, because our findings here can help to explain some of the unusual gaming results.

Our overclocked E8600 required FSB-1600 to reach 4.0 GHz, but most chipsets don’t support this bus speed natively. Overclocking the chipset itself means that the memory controller will operate at a higher performance level while consuming more power, producing more heat, and increasing the likelihood of data errors.

Sets of chipset parameters for a given bus speed are often called “bootstraps,” and correspond to the clock speed for each supported bus speed. Many motherboard manufacturers have been able to locate undocumented bootstraps in various chipsets, enabling these in an attempt to gain an overclocking advantage over the competition.

  • The 975X supports only 200 MHz and 266 MHz boot straps.
  • The P965 officially supports 200MHz and 266 MHz bootstraps, but our tests have indicated that certain boards may be accessing an unofficial 333 MHz strap.
  • The P35 Express officially supports 200 MHz, 266 MHz, and 333 MHz bootstraps, but our tests have indicated that certain boards may be accessing an unofficial 400 MHz strap.
  • The X38 doesn’t officially have a 400 MHz bootstrap, yet it’s there on most motherboards for everyone to use. That’s because its functionally identical to the FSB-1600 supporting X48 Express.
  • The P45 doesn’t officially have a 400 MHz bootstrap, but once again this unofficial setting is found in the BIOS of nearly every performance motherboard.

To rule out chipset overclocking as a cause for increased CPU performance, we first ran Sandra Arithmetic and Multimedia benchmarks.

With no discernible difference in CPU performance, it’s time to see what northbridge overclocking means in terms of memory performance. For that, we used Sandra’s memory bandwidth benchmark.

The farther a chipset was pushed beyond normal operating parameters, the more memory bandwidth the CPU had—almost certainly the result of overclocking the chipset’s memory controller.

But not every benchmark is affected by memory speed. For example, 3DMark Vantage tries to single-out the performance of the graphics processor and CPU. We expect PCI Express bandwidth to take on a far more significant role than memory bandwidth in these tests. For the following charts, we used the P35 Express to examine single graphics card performance in PCI Express 1.1 transfer mode, and the P45 Express to test a single graphics card in PCI Express 2.0 transfer mode.

As expected, 3DMark Vantage related its overall score almost directly to its GPU score, with CPU performance relatively constant.

But synthetic scores have little relevance to real-world performance, and now it’s time to see how each motherboard performed in actual games.

5. Test Results: Crysis, GRID, And Supreme Commander

Crysis gets no additional performance from CrossFire mode at low details and resolutions, but increasing its resolution to 2560x1600 allowed the high-end X48 Express to shine. Notice that these benchmarks fall almost directly in line with PCI Express bandwidth for a single card.

The CrossFire performance advantage is much easier to see at very high details, with both the X48 and P45 playable at medium resolutions. Adding a second card actually hurt P965 and P35 systems in Crysis, with a performance loss exceeding 50% at 1024x768.

Race Driver GRID had an odd issue where the system would occasionally hang while loading if the graphics resolution and details were set too high. If we waited long enough for everything to load, we would experience a sudden increase of up to 1000% in frame rates, but at 2560x1600 pixels and high details our systems would only load approximately one out of every ten attempts. Thus, while we were able to test low details at 2560x1600, we were forced to limit our high-detail maximum resolution testing to 1920x1200 pixels.

There’s no need for CrossFire with GRID set to low details.

CrossFire provided GRID a big performance advantage at 1920x1200 pixels and ultra-high details, where the bandwidth of the P45 Express chipset’s PCIe x8 2.0 mode slots put it ahead of the 975X’s PCIe x8 1.1 slots.

Supreme Commander Forged Alliance gets a big boost from the 975X motherboard’s overclocked memory controller, but more important was the gain that every motherboard showed from enabling CrossFire. Strangely, the PCIe x4 1.1 secondary slots of the P965 and P35 Express motherboards had little adverse effect in this game title.

Enabling higher visual details has little effect on the leadership positions of various motherboards, with the 975X’s higher memory bandwidth still keeping it in the lead and the x4 secondary slots of the older mainstream boards having very little negative effect on the CrossFire performance advantage.

6. Test Results: Unreal Tournament 3 And World In Conflict

Unreal Tournament 3 doesn’t need the power of two high-end graphics cards to get exceptional frame rates at high-resolutions and low-details, but the CrossFire advantage is still good for bragging rights. Once again, the old 975X wins because of its overclocked memory controller.

Cranking up the details does little to slow down UT3.

World In Conflict shows only the slightest CrossFire performance advantage when very low details are used.

At 2560x1600 and very high details, World In Conflict finally shows the bandwidth advantage of the P45 Express’ 2.0 slots, as compared to the 975X’s PCIe 1.1 transfers. Once again, a single x16 slot beats an x16/x4 CrossFire configuration in the P965 and P35 Express chipsets.

7. Performance Summary And Conclusion

If this were a standard motherboard review, we might simply throw up a single chart to show the overall performance differences between various CrossFire configurations, and it would look like this:

But while the simplified chart does show the problem of running CrossFire on P35 and P965 Express chipsets, it doesn’t really address the needs of individual buyers. The biggest concern will be the level of detail and resolution each gamer requires to reach a gaming-immersion comfort level.

At our lowest settings, the average gamer is best-advised to use a single card.

Increasing visual quality levels while maintaining a low resolution puts emphasis on memory performance, not graphics performance, as indicated by a significant win for the 975X motherboard.

Most middle-market buyers use either an old 1600x1200 display or a more modern 1680x1050 wide screen. At low-quality settings, the CrossFire performance advantage is generally too small to call any upgrade a good value unless a specific game requires the added performance.

Anyone who likes their games to look good and has at least a medium-resolution flat panel will love the performance boost CrossFire gives to our more modern motherboard choices. On the other hand, x4 pathways for the secondary graphics slots of P965 and P35 Express motherboards make any CrossFire upgrade a bad idea overall.

CrossFire really shines at 2560x1600, but anyone who can afford a 30” flat panel probably wouldn’t settle for low-quality graphics details.

Increasing detail levels enables even more value on the CrossFire solution, but anyone who can afford a 30” flat panel probably isn’t in the market for old system upgrades. Thus, the X48 and P45 results should be the only ones that matter at these settings.

Conclusion

Upgrading CrossFire-ready chipsets with a second graphics card is a great way for gamers to extend the useful life of their systems, as it provides nearly universal performance increases at medium to high resolutions and high details. Huge gains found on chipsets as old as the 975X prove that age isn’t a problem for buyers seeking a moderate-cost system renovation. Unfortunately, many other users don’t realize that their chipsets aren’t suitable for a CrossFire upgrade, in spite of any second or third graphics card slots.

Individual benchmarks showed that CrossFire performance is far too inconsistent on P35 and P965 chipsets to recommend it universally. In fact, several benchmarks suffered significant performance losses due to the low bandwidth of the second graphics card slot. Owners of these ill-equipped motherboards must carefully consider which games they’d like to continue playing before taking the plunge.

The easy solution for P965 and P35 Express systems is to simply use a single, more powerful card. At the very least, the issue we encountered where a secondary PCI Express x4 slot’s limited bandwidth hurt performance can be avoided by using a single card, so comments such as "CrossFire doesn’t work right" wouldn’t apply to single-card CrossFire solutions such as the HD 4870 X2. But the point of this article was to judge the viability of adding a second card when a reasonably good card is already installed, and this is an area where P965 and P35 Express motherboards simply cannot be upgraded to compete with modern builds.