AMD’s 790FX chipset

Processors and graphics chips are easily the rock stars of this industry, and for good reason. They’re largely responsible for overall system performance, and that’s even more appealing to enthusiasts than eyeliner and tight leather pants are to hair metal groupies. It’s fitting, then, that the bulk of attention and excitement surrounding AMD’s new Spider enthusiast platform will narrow on its new Phenom processor and Radeon HD 3800 series graphics cards.

If Spider were a rock band, Phenom would undoubtedly be the front man, with the Radeon HD 3870 and 3850 filling in at lead and rhythm guitars. On drums, you’d find AMD’s new 790FX chipset, hidden away not behind a massive array of snares and cymbals, but under a heatsink amongst a sea of other chips on a crowded motherboard.

Like drummers, chipsets spend most of their time outside the limelight; even shirtless and sweaty, they don’t pack much in the way of sex appeal. But they’re the glue that holds everything together, and that counts for a lot if you value system stability. On occasion, chipsets can also bring important new features to the table that will shape system architectures for years to come. This is one such occasion, with the 790FX packing second-generation PCI Express, support for two-, three-, and four-way CrossFire configurations, and the first desktop implementation of HyperTransport 3. Join us as we explore how these features tie the Spider platform together, and how the 790FX’s performance, power consumption, and initial motherboard implementations measure up.

The itsy bitsy 790FX
The heart of the 790FX chipset is a north bridge chip that shares the same name. This chip is built using 65nm process technology, allowing for a chipset thermal design power (TDP) of just 10 watts. AMD is quick to hype the 790FX’s power efficiency, no doubt in part due to the notoriously high power consumption of its nForce rivals, particularly the 590 SLI. We’ll see how actual motherboard power consumption shakes out a little later.

With all of the 790FX’s new features packed into the north bridge, it’s fitting that the name extends across the entire chipset. The most important of these new features is easily a HyperTransport 3 interconnect that allows the 790FX to make the most of AMD’s latest Phenom processors. This interconnect retains HyperTransport’s 16-bit data pathway, but doubles its clock speed to 2GHz, yielding a twofold increase in bandwidth between the processor and north bridge.

Doubling bandwidth is a bit of a theme for the 790FX, whose second-gen PCI Express lanes offer twice the bandwidth of their gen-one counterparts. As with the 790FX’s HyperTransport 3 implementation, PCIe’s second coming doesn’t actually change the width of the data pathways. Instead, the increase in bandwidth comes courtesy of faster signaling rates that give each PCIe 2.0 lane 1GB/s of bandwidth—twice that of their predecessors.

In addition to offering substantially greater bandwidth, second generation PCI Express lanes are also capable of throttling lane speeds to conserve power at idle. PCIe 2.0 retains backwards compatibility with gen one hardware, as well, allowing the 790FX to play nicely with such outdated gen-one graphics cards as the, er, GeForce 8800 GTX.

The 790FX block diagram. Source: AMD

With a total of 42 lanes of PCIe 2.0 at its disposal, the 790FX has connectivity to spare. 32 of those lanes are reserved for graphics cards, with lanes split evenly for dual-card CrossFire configurations. The 790FX will also be compatible with three- and four-way CrossFire setups that are due to be enabled in future graphics driver releases. For these multi-card configs, graphics cards will receive eight lanes of bandwidth each—plenty considering the fact that PCIe 2.0 offers twice the bandwidth per lane of its predecessor.

Of the 790FX’s 10 remaining PCIe lanes, six can be used for additional expansion slots or onboard peripherals. The final four are consumed by the interconnect that links the chipset’s north and south bridge components. Four lanes at 1GB/s per lane should yield 4GB/s of bandwidth here; however, because the 790FX’s south bridge component is limited to first generation PCI Express lanes, interconnect bandwidth tops out at only 2GB/s.

So why doesn’t the 790FX’s south bridge component have fancy pants PCIe 2.0? Because it’s the same SB600 chip that ATI introduced way back in May 2006. AMD is working on a new SB700 south bridge that will apparently make its way into the second wave of 790FX boards due out early next year. However, the SB700 isn’t ready yet, and AMD has elected to launch Spider without it.

With all its PCIe lanes and graphics cards hanging off the north bridge, there isn’t an overwhelming need for additional interconnect bandwidth. Heck, even Intel’s latest X38 Express chipset makes do with a 2GB/s chipset interconnect. The SB600 shows its age in the features department, however, particularly when compared with the nForce 590 SLI.

There was a time when four Serial ATA ports was more than enough for most folks, but that was back when SATA optical drives were few and far between. Serial ATA burners are much more common today, and for users looking to rid their systems of bulky IDE cables, four ports looks a little thin. The SB600’s lack of RAID 5 support also stands out, although it will probably be less of a concern for most users.

Motherboard makers can always add auxiliary storage controllers to help bolster the SB600’s I/O capabilities and RAID support. At the very least, they’ll have to add their own Ethernet controller to make up for the SB600’s lack of integrated networking. Ideally, board designers will opt for PCIe-based GigE controllers that offer full Gigabit throughput and exhibit low CPU utilization. That’s not guaranteed, though; we’ve seen enough motherboards equipped with slow PCI-based Ethernet controllers to know mobo makers are prone to cutting costs wherever they can, even if it means sacrificing peripheral performance.

Gigabyte’s GA-MA790FX-DQ6
First out of the gate

AMD may have officially launched the 790FX today, but Gigabyte’s GA-MA790FX-DQ6 has been selling at several online retailers since last week. The board is currently listed at $269, which is a little above the $150-$250 price range AMD expects for 790FX boards. Competition will probably force lower prices as other boards make their way to market, but given the initial sticker price, it seems unlikely we’ll see the DQ6 slot in at under $200.

Of course, Gigabyte’s DQ6 models are typically its most lavish, and the GA-MA790FX is certainly no exception. The board’s a looker, too, draped in Gigabyte’s trademark turquoisey blue.

With motherboard makers struggling to differentiate their offerings from the competition, Gigabyte recently began pushing the component quality of its high-end motherboards. The DQ6 features solid state capacitors across the, er, board, and Gigabyte says they’re much more reliable than traditional electrolytic designs. The company also boasts that the DQ6’s ferrite core chokes lose 25% less energy than those with iron cores, and that its low RDS MOSFETs run 16% cooler than standard ones.

We’re pleased to see motherboard makers pay additional attention to component quality, but that’s only one part of what makes a good motherboard. Layout is also extremely important, and that’s where we start to see the DQ6 falter a little. Things get off to a good enough start, with Gigabyte placing power connectors along the edges of the board where they won’t interfere with airflow around the CPU socket.

The socket area is also free of taller capacitors that might conflict with aftermarket heatsinks or cooler retention mechanisms. Even heatpipe network that snakes between the board’s passive chipset and voltage circuitry heatsinks manages to stay out of trouble.

Gigabyte’s expansive cooling solution actually extends to the underside of the board, where metal plates back not only the CPU socket, but also the north and south bridge chips. This level of chipset cooling seems a little extravagant given the 790FX’s 10W TDP, but it’s silent and unobtrusive, so we really can’t complain. The additional cooling may help the DQ6 respond better to overclocking, as well.

The DQ6 starts to run into problems when we bust out the big daddy, Scythe’s popular—and surprisingly affordable—Ninja cooler. Towering over lesser designs, the Ninja is one of the largest heatsinks on the market. That creates all sorts of potential for clearance issues, and the DQ6 actually does pretty well to avoid them. However, the board’s DIMM slots are quite close to the CPU socket, leaving barely enough room for standard-height memory modules with this monster installed. You can see in the picture above that the Ninja’s heatpipes make contact with the memory heatsink on the closest DIMM, as well.

Of course, most aftermarket coolers should have no problems clearing the DQ6’s DIMM slots. The Ninja is obviously an extreme case, but then that’s exactly why we’ve included it here. You know, because enthusiasts are all xTr3me, and whatnot.

Layout woes continue as we move down to the DQ6’s cluster of Serial ATA ports. Two of these ports connected to an auxiliary GSATA controller (in purple over on the left in the picture above) are tucked out of the way at the bottom of the board. The four orange ports hanging off the SB600 are somewhat more precariously placed. Longer double-wide graphics cards installed in the third PCI Express graphics slot can block access to these, either making them completely inaccessible or requiring the use of right-angle SATA connectors.

Fortunately, it’s extremely unlikely that you’d ever have a double-wide card installed in the third graphics slot. The DQ6’s first and fourth PCIe x16 slots, colored in blue, are “master” slots that get a full 16 lanes of bandwidth. These are the slots you’d use with single-card or two-way CrossFire configs. The orange x16 slots are only there for users looking to run three or four graphics cards, either on their own to provide additional monitor outputs, or teamed up for CrossFire.

While this x16 slot arrangement ensures that you probably won’t run into SATA port clearance problems, it’s a poor choice if you actually want to take advantage of the DQ6’s support for more exotic CrossFire configs. Having the slots stacked so close together means that three- and four-way CrossFire configs will be limited to single-slot cards, and that doesn’t make much sense at all. Single-slot coolers are typically limited to lower-end models like the Radeon HD 3850. However, if you’re going to go all out with a four-way CrossFire config, you’ll probably want to go with something a little faster and with more memory, like the dual-slot 3870.

Even if you could find a single-slot card appropriate for four-way CrossFire, you probably wouldn’t want to stack four of them on top of each other. Such a configuration doesn’t leave much room for airflow between cards, and without the back plate exhaust ports found on dual-slot designs, things are likely to get toasty fast.

Redemption finds the DQ6 when we move to the board’s port cluster, which has everything but the kitchen sink… and a parallel port. Gigabyte manages to squeeze an old-school serial port into the mix alongside two eSATA ports, coaxial and TOS-Link S/PDIF outputs, six USB ports, and Firewire.

BIOS options
The DQ6’s BIOS offers a good array of tweaking and overclocking options, but it’s clearly a work in progress. We’ve experienced several issues with the BIOS that have yet to be resolved, such as the inability to install or even run Windows Vista when the south bridge is set to AHCI mode. AHCI is necessary for not only hot plugging, but also Native Command Queuing (NCQ), so it’s not a trivial loss. We’ve also been unable to get the board to go past the POST screen with manually set memory timings, even when those timings exactly match the default values used in “auto” mode. These problems have occurred with hardware that works flawlessly on other platforms, and the DQ6 isn’t the only board afflicted.

We also have Asus’ M3A32-MVP Deluxe in our labs, and although it doesn’t have a problem setting manual memory timings, it has similar issues running the south bridge running in AHCI mode. The Asus board has exhibited questionable stability, as well, failing to post after warm reboots and apparently losing its onboard Ethernet controller.

Asus and Gigabyte are usually pretty good about having their BIOSes in order, and the fact that both seem to be struggling with the 790FX is worrisome. Hopefully these issues can be resolved with simple BIOS updates.

If they can, the DQ6 looks well-equipped for prospective Phenom overclockers. In addition to giving users CPU multiplier control, HT clock speeds are available up to 500MHz. CPU voltages scale up to 1.75V, as well, and you can push memory voltages up by half a volt.

Ideally, we’d like to see Gigabyte support a greater range of memory speeds, either via additional bus speed options or by exposing users to a larger set of bus multipliers. It would also be nice if the board’s automatic fan speed control options let users set target and reference temperatures for the system and processor fan headers. In its current form, the DQ6’s BIOS provides no way for users to manipulate how the board’s temperature-based fan speed control behaves beyond turning it on or off.

AMD’s OverDrive utility
If poking around in the BIOS seems like a horribly outdated way to tweak system performance, AMD has whipped up a handy OverDrive system utility for Windows that packs all sorts of hardware tweaking, overclocking, and monitoring capabilities.

Perhaps most impressive, if only for scale, is the app’s vast array of memory timing options. You could get lost in here.

OverDrive should make overclocking a breeze by exposing not only control over bus speeds and CPU multipliers, but also a whole string of voltages. The screenshot above was taken with an Athlon X2 5200+ processor, so you can’t see OverDrive’s neatest feature—independent core multiplier control for Phenom processors. AMD’s latest chips are capable of running cores at different speeds, and OverDrive will allow users to easily tune those speeds to maximize performance without sacrificing stability.

Manual tweaking isn’t for everyone, though, so OverDrive also provides an auto-tuning feature that will probe your system’s potential and settle on the fastest stable overclock. This “Auto Clock” mechanism seems to zero in optimal configurations a lot quicker than Nvidia’s nTune system utility, too.

If you’d rather take over at the wheel, OverDrive provides a full suite of manual stress tests to aid with burn-in and stability testing.

A monitoring panel lets you keep tabs on important system variables, as well. Having per-core temperature readings should be particularly useful when combined with Phenom’s ability to set independent multipliers for each core.

OverDrive isn’t the first tweaking and overclocking software we’ve seen developed by a chipset maker; Nvidia blazed that trail years ago with nTune. However, nTune never seemed to catch on with motherboard makers, who had to implement specific BIOS hooks to expose much of the software’s functionality. OverDrive requires cooperation from the BIOS, too, but AMD says it’s all handled by an easy-to-implement ASM plugin. What’s more, AMD claims that all the tier one mobo makers will offer full OverDrive support, although they admit that some aren’t quite up to speed at the moment.

Specifics on specifications
Enthusiasts are an efficient lot, so we’ve distilled the GA-MA790FX-DQ6’s important specifications into a single chart for your inspection. We’ve covered most of the important bits already, but there are a few stragglers worth highlighting.

First, note that the DQ6 actually features a pair of “GSATA” branded auxiliary Serial ATA controllers. One of these chips is responsible for the board’s additional SATA ports, while the other governs its pair of eSATA connectors. The board also sports a couple of Realtek RTL8111B Gigabit Ethernet controllers that are, mercifully, tied to the bandwidth-rich PCI Express bus.

On the audio front, the DQ6 features yet another Realtek codec, this time the ALC889A. The ALC889A was designed with Vista in mind, and Realtek claims a signal-to-noise ratio of 106 decibels for the chip’s DACs. DTS Connect is also supported, allowing users to run eight-channel audio through a single digital connection rather than a mess of analog wires.

Rounding out the DQ6’s integrated peripheral suite is a Firewire chip from Texas Instruments. The TSB43AB23 only supports the 400Mbps 1394a spec, but given Windows’ spotty support for the faster 1394b standard, that’s a good call on Gigabyte’s part.

Our testing methods
Today we’ve pitted the 790FX against its closest and really only rival, Nvidia’s nForce 590 SLI. Launched about a year and a half ago, the 590 SLI is hardly a spring chicken. However, it’s still Nvidia’s flagship chipset for Socket AM2 processors, a testament to just how little attention chipset makers have paid AMD processors since the Core 2 Duo’s release.

Since we couldn’t get the GA-MA790FX-DQ6 to cooperate with manual memory timings, we’ve had to back both boards off to Gigabyte’s defaults for our Corsair DIMMs, which are a relatively relaxed 5-4-4-12-2T. We’ve also had to run the Gigabyte board in Native IDE mode rather than AHCI, so keep in mind that a lack of Native Command Queuing could impact our storage subsystem tests.

For our nForce platform, we’ve elected to use Foxconn’s C51EM2AA because it’s essentially an Nvidia reference design for the nForce 590 SLI.

All tests were run three times, and their results were averaged.

Thanks to Corsair for providing us with memory for our testing. 2GB of RAM seems to be the new standard for most folks, and Corsair hooked us up with some of its 1GB DIMMs for testing.

All of our test systems were powered by OCZ GameXStream 700W power supply units. Thanks to OCZ for providing these units for our use in testing.

Finally, we’d like to thank Western Digital for sending Raptor WD1500ADFD hard drives for our test rigs. The Raptor’s still the fastest all-around drive on the market, and the only 10K-RPM Serial ATA drive you can buy.

We used the following versions of our test applications:

• SiSoft Sandra Standard 2008 12.34
• WorldBench 6.0 Beta 2
• TCD Labs HD Tach 3.01
• Futuremark 3DMark06 Build 110
• RightMark Audio Analyzer 6.05
• NTttcp
• CASE Lab Euler3d CFD benchmark multithreaded edition
• Intel IOMeter v2004.07.30
• Crysis SP Demo
• Enemy Territory: Quake Wars Demo
• Bioshock
• CPU-Z 1.41

The test systems’ Windows desktop was set at 1280×1024 in 32-bit color at an 85Hz screen refresh rate. Vertical refresh sync (vsync) was disabled for all tests.

All the tests and methods we employed are publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.

Memory performance
Memory subsystem performance doesn’t always track with real-world applications, but it’s a good place to start.

Thanks to AMD’s on-die memory controllers, we don’t see much difference in memory performance here. The 790FX has the lead in Sandra’s memory bandwidth test, but only a fraction of a nanosecond separates the two platforms when we probe CPU-Z memory latency.

Memory controllers don’t always handle four DIMMs gracefully, so we popped an additional two memory modules into each system for another round of tests.

With an additional two DIMMs onboard, the 590 SLI manages to close the gap in Sandra. Latency scores remain close.

The following latency graphs are a little indulgent, so I won’t be offended if you skip them. They show access latencies across multiple block and step sizes, painting a full picture of memory controller performance with each chipset. Yellow represents L1 cache, light orange is L2, and dark orange is main memory.

Same on-die memory controller, same overall latency performance. Move along.

STARS Euler3d computational fluid dynamics
Few folks run fluid dynamics simulations on their desktops, but we’ve found this multi-threaded test to be particularly demanding of memory subsystems, making it a good link between our memory and application performance tests.

The 790FX’s slight advantage in memory subsystem performance extends to Euler3d, where the nForce 590 SLI falls a little off the pace.

Worldbench
WorldBench uses scripting to step through a series of tasks in common Windows applications. It then produces an overall score. WorldBench also spits out individual results for its component application tests, allowing us to compare performance in each. We’ll look at the overall score, and then we’ll show individual application results alongside the results from some of our own application tests.

That’s what you call a dead heat, folks. Let’s see if any of WorldBench’s individual application tests can tease out a meaningful performance difference between these two chipsets.

WorldBench’s multimedia editing and encoding tests show slim, but consistent advantages for the 590 SLI. The biggest disparity comes in the Media Encoder test, where the 790FX falls 16 seconds behind.

Results are considerably closer through WorldBench’s office and multitasking tests. The 790FX steals four seconds in the Firefox test, which has traditionally favored faster memory subsystems.

3dsmax rendering also favors the 790FX by four seconds, but the nForce 590 SLI is a hair faster with the DirectX test.

The performance delta between the 790FX and 590 SLI nearly grows to double-digit-seconds in WinZip, with the 790FX again finding itself trailing.

Gaming

Although the 790FX has a slim lead in Crysis and Quake Wars, it can’t keep up in Bioshock. Interestingly, our Bioshock tests were conducted with FRAPS rather than the in-game demos used in Crysis and Quake Wars. Testing with FRAPs does tend to generate more variance within results, but our scores for the 790FX and 590 SLI were consistent to within two frames per second across five test runs each.

Serial ATA performance
The Serial ATA disk controller is one of the most important components of a modern core logic chipset, so we threw each platform a selection of I/O-intensive storage tests using a Western Digital Raptor X.

IOMeter
We’ll begin our storage tests with IOMeter, which subjects our systems to increasing multi-user loads. Testing was restricted to IOMeter’s workstation and database test patterns, since those are more appropriate for desktop systems than the file or web server test patterns. Keep in mind that because we couldn’t get the SB600 south bridge running in AHCI mode, Native Command Queuing isn’t enabled, putting the 790FX at somewhat of a disadvantage.

That disadvantage becomes clear rather quickly when the 790FX’s transaction rate fails to take off as the number of concurrent I/O requests increases. Performance does scale as the load increase, but not nearly as quickly as the 590 SLI. It is worth noting, however, that the 790FX offers higher transaction rates with extremely low and extremely high load levels; it’s just slower in the meat of the curve.

IOMeter response times track with transaction rates, with the AMD chipset proving slightly quicker at the extremes and the nForce delivering better performance in the middle.

The nForce 590 SLI exhibits higher CPU utilization than the 790FX in IOMeter, but only by about half a percent.

HD Tach
We used HD Tach 3.01’s 8MB zone test to measure basic SATA throughput and latency.

AMD has a slight edge in burst performance, managing 4MB/s faster than the 590 SLI.

That victory is short-lived, though. The 790FX is way behind in HD Tach’s sustained write speed test, offering just over half the throughput of the 590 SLI.

AMD finds some salvation in HD Tach’s random access time test, where the 790FX just edges out the 590 SLI.

The 790FX also exhibits lower CPU utilization, although to be fair, scores are within HD Tach’s +/- 2% margin of error in this test.

USB performance
Our USB transfer speed tests were conducted with a USB 2.0/Firewire external hard drive enclosure connected to a 7200RPM Seagate Barracuda 7200.7 hard drive. We tested with HD Tach 3.01’s 8MB zone setting.

USB performance was the Achilles heel of ATI’s SB400 series south bridge chips, and although the SB600 is certainly an improvement, it’s consistently slower than the nForce 590 SLI. These aren’t trivial margins, either, particularly when both implementations offer identical CPU utilization.

PCI Express performance
We used ntttcp to test PCI Express Ethernet throughput using a Marvell 88E8052-based PCI Express x1 Gigabit Ethernet card.

CPU utilization is a little lower on the 590 SLI, but not by enough to get worked up about.

PCI performance
To test PCI performance, we used the same ntttcp test methods and a PCI VIA Velocity GigE NIC.

Early ATI south bridge chips had serious PCI performance issues, which is why we started testing PCI throughput in the first place. It’s a good thing we’ve kept testing, because the 790F’s PCI Ethernet throughput is a good 75Mbps off the pace set by the 590 SLI. Sure, AMD’s CPU utilization is lower, but it’s pushing fewer bits.

Power consumption
We measured system power consumption, sans monitor and speakers, at the wall outlet using a Watts Up Pro power meter. Power consumption was measured at idle and under a load consisting of a multi-threaded Cinebench 10 render running in parallel with the “rthdribl” high dynamic range lighting demo.

Now that’s a surprise. Nvidia’s nForce 590 SLI is notoriously power-hungry, but our 790FX-based Gigabyte board fares only a couple of watts better. So much for Spider’s power efficiency advantage.

To be fair, this higher than expected power consumption may not be the fault of the 790FX. Asus’ first wave of P35-based motherboards were also a little greedy with their power consumption, but subsequent boards based on the same chipset proved considerably more power-efficient. Let’s hope, for AMD’s sake, that the GA-MA790FX-DQ6 is the exception rather than the rule.

Overclocking
For our overclocking tests, we lowered the CPU multiplier to 5X and dropped the memory bus speed to 400MHz to take our processor and DIMMs out of the equation. We also dropped the HT processor link speed to ensure that it didn’t hinder our overclocking endeavors. From there, we started cranking on the HT clock, managing to push it to a perfectly stable 300MHz before the board started acting up.

Our Gigabyte board was happy to loop Prime95 and the rthdribl HDR lighting demo with a 300MHz HT clock, but any faster, and the board refused to post. Most of the time, it would automatically reboot, returning the system to its default configuration. However, on occasion, it would get stuck in an automatic reboot loop that would reset the system every few seconds until the power was cut manually.

We tried tweaking chipset and HT voltages to coax 310MHz from the board, but to no avail. Not that 300MHz is anything to sneeze at, but for a brand new chipset and a motherboard obviously geared towards overclockers, we were hoping for more.

As is always the case with overclocking, your mileage may vary.

Motherboard peripheral performance
Core logic chipsets integrate a wealth of peripherals, but they don’t handle everything. Firewire and audio are farmed out to auxiliary chips, for example. AMD chipsets also rely on third party silicon for networking, and many motherboards feature additional SATA controllers to complement south bridge Serial ATA offerings.

To provide a closer look at the peripheral performance you can expect from the motherboards we’ve tested today, we’ve compiled Ethernet, USB, Firewire, Serial ATA, and Audio performance results below. You’ll notice that there isn’t much variance from one board to another, but there are a few things worth pointing out.

On the Ethernet front, the GA-MA790FX-DQ6’s Realtek networking chips do a good job of keeping up with the nForce 590 SLI’s integrated GigE controllers.

The Gigabyte board also has a clear advantage in Firewire performance, offering faster burst speeds in addition to higher sustained throughput. Were Firewire more popular, that would almost make up for the SB600’s pokey USB performance.

There’s little difference in Serial ATA performance between the Gigabyte board’s SB600 and GSATA Serial ATA controllers.

RightMark Audio Analyzer yields identical analog output quality scores for both boards. Of course, if you’re really concerned with fidelity, you’ll want to take advantage of these boards’ digital audio outputs.

Conclusions
Spider is an ambitious attempt at an enthusiast platform, and while Phenom and the Radeon HD 3800 series might grab all the headlines, the 790FX chipset is undoubtedly a key component of the overall equation. With support for HyperTransport 3 and second generation PCI Express, the 790FX has the high-bandwidth connectivity needed to fully exploit Phenom processors on one end and Radeon HD 3800 series graphics cards on the other. In fact, the 790FX is the only chipset currently on the market that fully supports all the features packed into new Phenom processors. Couple that with excellent OverDrive tweaking and overclocking software that we’re told will be widely-supported by motherboard makers, and the 790FX looks pretty good.

Except for one or two not so minor problems.

Despite being the first new enthusiast chipset for AM2 processors in more than a year and a half, the 790FX relies on a dated SB600 south bridge chip whose limited connectivity options hardly befit a cutting-edge enthusiast platform. The SB600’s comparatively poor PCI and USB performance only make matters worse. With a new SB700 south bridge due to arrive early next year on a second wave of 790FX boards, it’s hard to get excited about current implementations.

The first crop of 790FX boards certainly has its share of problems, too. Gigabyte’s GA-MA790FX-DQ6 suffers from a number of BIOS-related issues that really need to be fixed before we would even consider recommending the board. And it’s not like this is a pre-release product whose kinks are still being worked out; the DQ6 was selling online before the 790FX officially launched. The DQ6 also suffers from a poor graphics slot layout and alarmingly high power consumption given the 790FX’s 10W TDP. It’s not cheap, either. With street prices currently hovering around $270, the DQ6 costs nearly as much as the fastest Phenom you can buy.

Our problems with 790FX motherboards haven’t been limited to Gigabyte, either. We haven’t yet had the opportunity to test MSI’s riff on the Spider platform, but we’ve found Asus’ M3A32-MVP Deluxe to have numerous BIOS-related issues and questionable stability. The fact that the top two mobo makers can’t get their initial 790FX offerings dialed for the chipset’s launch doesn’t give us a lot of faith in the platform.

I’m not even convinced there will be much demand for a high-end Phenom chipset. The processor’s performance isn’t compelling enough to inspire current Core 2 users to jump ship, and those running Socket AM2 processors will probably be more inclined to drop a Phenom into their existing motherboards rather than going out and buying new ones.

In the end, the 790FX is a bandwidth-rich new north bridge chip held back by a dated south bridge and flaky motherboard implementations. AMD will have to do better if it expects the Spider platform to take off with enthusiasts. In its current state, we simply can’t recommend the 790FX.