Intro:

Finally, there is a reason for fans of AMD and ATI to rejoice.  Though the recently released RADEON HD 3870 video card is not the fastest performing video card on the market, it comes to us at a good price ($219 MSRP) that can't be beat from a price/performance perspective.  NVIDIA's own 8800GT 512MB video card provides an excellent value as well, though the 8800GT's MSRP is higher ($249).  Availability problems have plagued both the 8800GT and RADEON HD 3870 since launch, driving the "actual" price of these cards upward.  In fact, there is only one 8800GT card available on Newegg as of the time of this writing (for $299 plus shipping), and there are currently no 3870's available. 

However, if an when the supply of these cards catches up with the demand, many gamers will turn in their current video cards for a chance to use one of these new models.  Though neither the 8800GT nor HD 3870 is the fastest video card available, they provide unmistakable value in terms of providing high performance at a relatively low price.  Also, two (or three?) of these cards could be combined in a SLI or Crossfire configuration, providing top-of-the-line performance for less than $500, however the motherboard compatibility with these technologies can be limiting.  SLI configurations are only possible on motherboards based on NVIDIA's own chipsets, which eliminates many fantastic motherboards based on Intel's chipsets.  AMD chipsets only support Crossfire, since ATI is a subsidiary of AMD.  Intel's chipsets do support Crossfire, though, since NVIDIA appears unwilling to open up SLI support on non-NVIDIA chipsets.  However, Intel's "mid-range" chipsets might not unlock the full potential of Crossfire performance, as we will see.

So, what's the problem?

Many gamers, like myself, have invested in Intel-based systems since the launch of the Core 2 Duo and Core 2 Quad, many using motherboards based upon Intel's 965 and P35 chipsets.  Motherboards based upon these two chipsets often have two PCI-express X16 slots for dual video card operation.  In fact, many motherboard boxes proudly advertise Crossfire compatibility when two ATI RADEONs are used.  However, due to limitations within the P965 and P35 chipset families, only one X16 PCI-express connection has full access to the chipset.  On the motherboards with two X16 slots, the 2^nd slot actually utilizes a connection to the southbridge chip that provides only X4 PCI-e bandwidth.  Additionally, the southbridge now must act as a relay between the P35 northbridge chipset and the video card plugged into the X4 PCI-e slot.  As can be seen in the diagram below, the P35 chipset shares an 8GB/s link with the x16 graphics slot, and the ICH southbridge chip has six 500MB/s PCI express lanes available.  Four of these lanes can be dedicated to the 2^nd x16 graphics slot, but as can be seen, this provides only 2GB/s bandwidth for the graphics, and there is also only a 2GB/s connection between the southbridge and the P35 chipset.  Furthermore, this connection has to share bandwidth with any hard drives and networking present on the motherboard, making this connection seem paltry compared to the dedicated 8GB/s connection with the primary graphics slot. 

So, the question on many enthusiasts' minds is, how much does this limited bandwidth impact performance of a dual RADEON 3870 Crossfire system?  Intel's own X38 chipset now provides full x16 bandwidth to two graphics slots, but motherboards based on that chip cost at least $200, making a full x16/x16 Crossfire configuration an expensive one. 

Test Setup

I will attempt to analyze Crossfire performance on a P35-based motherboard to see if the x4 PCI-e connection and limited chipset/ICH bandwidth negatively impacts game performance. 

The motherboard in question is ASUS' excellent P5K standard edition motherboard.  This board has been great for me, overclocking the current Q6600 CPU to 3.4ghz comfortably and without issue.    Other relevant hardware is as follows:

• Intel Core 2 Quad Q6600 CPU overclocked to 3.4GHz
• ASUS P5K motherboard based on Intel P35 chipset
• 2GB OCZ Platinum DDR2-800 at 2x378MHz (756MHz DDR)
• Two AMD ATI RADEON HD 2900XT 512MB video cards
• Western Digital Raptor 150GB
• Thermaltake 750W Toughpower PSU
• Windows XP Media Center Edition 2005

I have tested Team Fortress 2, Oblivion, and Enemy Territory: Quake Wars using FRAPS to record speeds during repeatable run-throughs.  I attempted to benchmark Crysis, but unfortunately multi-GPU performance in Crysis is still very much a work in process.  A patch is supposed to be released this month that will enable Crossfire support in Crysis, and I hope to be able to run some tests at that time (if I still have the video cards by then).  As it stands today, the performance between one or two video cards in Crysis is identical.  I have also used the ubiquitous benchmark 3DMark06.  I used two 2900XT video cards instead of the new 3870 cards for three reasons:  One, the aforementioned availability problems.  Two, the 2900's were available for me to run some tests on for free, which is always good.  And three, the 2900s are very comparable to the 3870, so the results should be very close to what you could expect from two 3870 cards. 

Game tests were run with the highest available in-game settings at a resolution of 1600x1200, along with 16x AF and 8x adaptive AA set via the control panel, except for Crysis which was run without AA or AF, and in-game settings to "high".  3DMark06 was run with the standard settings at 1280x1024. 

Game Tests:

First, the benchmarks using the games:

It appears Crossfire is working properly for Team Fortress 2 and Oblivion, with average FPS increases of 73.4% and 63.8% for those two games, respectively.  Enemy Territory: Quake Wars showed an improvement of nearly 50% on average FPS, going from 35.1 to 52.5 FPS.  Crysis showed a moderate improvement of over 40% FPS gain on average.  The largest FPS rate increases occurred for the minimum frame rates especially, with TF2's minimum rate increasing to 74 FPS from 39 FPS with one card, and Oblivion's minimum rate increased from 18 FPS to 30 FPS.  These numbers make a significant difference, as it is during crucial scenes during games (intense fights or scenes viewing long distances from a high point, for example) that the FPS drops to the minimum, and these numbers show the minimum FPS increases the most when using a Crossfire configuration. 

3DMark06 & PCI-e Bus Overclocking

Next, I'll take a look at 3DMark06 results:

I tested two ways here.  The blue bars show the increase in performance due to enabling Crossfire versus a single-card setup.  The Purple bars represent the results after further tweaking on my part. 

In the standard Crossfire setup, the overall 3DMark score only increased by 39.4%, however the game tests increased by an average of 50.9% (the overall 3DMark score might be held back by a lack of CPU score increase). 

The second set of results was obtained after the PCI express bus was overclocked via the BIOS.  A setting of 110MHz was used (as opposed to the standard 100MHz), which resulted in an average game test increase of 53.2%.  It would appear that, in fact, the PCI-e connection of the video cards is somewhat limiting the performance of the cards when in Crossfire since increasing the PCI-e speed by 10% resulted in a performance increase of nearly 5%.  I also attempted to run Crysis using the overclocked PCI-e bus speed, and Crysis showed zero benefit.  Crysis is a new game, and it is likely further driver tweaking is necessary on AMD's part before this title is optimized for Crossfire operation. 

As a last resort, I also plugged in a second Crossfire bridge connector between the two cards to see if the theoretical added bandwidth would improve results.  I saw no difference in performance with the two bridges installed. 
Conclusion

So, what's the answer to the question of how much the reduced PCI-express bandwidth of the Intel P35 and 965 chipsets impacts Crossfire performance?  Honestly, the results are mixed.  In TF2 and Oblivion, Crossfire performance was anywhere between 64% to 90% faster than single card performance, depending on the scenario.  Quake Wars showed about a 50% average FPS increase in Crossfire mode.  The game tests in 3DMark06 showed improvements between 31% and 73%, again, depending on the test.  After taking a look at some submitted results on Futuremark's ORB (Online Result Browser) that included a Q6600 at 3.5ghz, two HD 2900 cards, and an ASUS Maximus Formula motherboard (based on Intel X38 chipset with x16/x16 pci-e connections), it appears the additional bandwidth of the 2^nd PCI-e slot provided an average FPS increase of 3% versus my result with the standard Crossfire configuration.  This 3% advantage could be overcome by overclocking the PCI-express bus as I did above. 

According to my results, the PCI-express bus speed is not a significant limitation at this point with ATI's high end video cards in a Crossfire setup on Intel's P35 chipset (which is representative of the P965 chipset as well).  Even though the secondary X16 physical PCI-e slot only runs in X4 mode electrically, my tests show the negative performance impact of this is around 5%, at most.  In fact, in some cases, like Crysis, increasing the PCI-e bus speed had zero effect. 

However, I never once saw performance gains nearing 100%.  The biggest performance increase due to Crossfire was 90% on the minimum measured framerate while playing TF2.  This means there is a bottleneck somewhere (most likely in the drivers).  It is possible the PCI-e bus would become much more of a limitation if driver overhead was improved.  Also, as video cards continue to become more powerful, it stands to reason that the PCI-express bus would become more of a limitation.  For example, ATI's upcoming 3870 X2 video card will essentially consist of two 3870 GPUs on one video card.  This card will theoretically provide twice the performance of the current 3870, and I would assume will consume more bandwidth. 

As it stands today, I would recommend a second 2900 or 3870 card to anyone considering adding one to their Intel P35 or 965 motherboard if the highest level of performance is desired within that same platform. 

---