Nvidia

You gotta give kudos where they're due. NVIDIA has done an unparalleled job of going from nowhere to dominance in its core logic segment in only a few years. Part of this success has hinged on the company's ability to design amazing software solutions able to capitalize on and popularize the chipsets' underlying functionality. Part of the story is simply the outstanding performance delivered by each of the platform's components.

"Mercury Research published in their Q4 [2004] analysis that we are the leading core logic supplier for K8 products with 48% market share," says Drew Henry, general manager, MCP business for NVIDIA. "That was a growth of 3% for us from Q3. In fact, our only disappointment with the K8 market is that it grew slower than we'd hoped. But now AMD is ramping K8 very, very hard, and it's fast becoming their highest volume SKU."

With the nForce3 family, NVIDIA tackled the Athlon 64/FX platform. NF3 lost the SoundStorm technology with Dolby 5.1 encoding found on the nForce2 but does offer an integrated firewall, robust RAID features (0, 1, and 0+1 with the use of a secondary SATA controller), and AGP 8X. Within the nForce3 family, NVIDIA delivered its usual handful of top-to-bottom SKUs, with the 250 Gb and Ultra (same chip, only one is for 754-pin CPUs and the other for 939) offering such good networking throughput and low CPU utilization that it even blew Intel's vaunted CSA chip out of the water. According to NVIDIA, its GbE implementation still outperforms PCIe chips from the likes of Marvell and Broadcom. By also jettisoning the integrated graphics, NVIDIA was able to squeeze all of this into a single core logic chip, which further cut down on communication latencies.

Today's nForce4 family updates the nForce3 into the PCI Express era, then adds a few notable perks. In addition to 20-lane PCI Express, NVIDIA bumps the NF4 up to 10 USB 2.0 ports, preserves the on-chip GbE as well as the RAID features, and then updates the firewall and RAID functionality. The NVIDIA firewall still looks and feels the same. The key difference is that the "ActiveArmor" implementation in the NF4 keeps all of the packet processing in the chipset whereas the NF3 offloaded this to the CPU via driver software. The difference between the approaches can easily exceed 50% in CPU utilization. Note that the standard model of the nForce4 still preserves this NF3-class approach. Only the NF4 Ultra and SLI models support ActiveArmor. Since we spent a fair amount of space covering NVIDIA's firewall in our recent cover story on security hardware, there's no need to rehash all of that here.

As with the nForce3, the nForce4 supports cross-controller RAID, meaning that both PATA and SATA drives can coexist on the same array. Moreover, NVRAID allows for rebuilding of a RAID 1 spare on the fly, much like a RAID 5, which potentially eliminates downtime in the face of drive failure. Users can morph from one RAID type to another in only a few clicks, and NVIDIA even throws in little touches such as Disk Alert, which, if the board vendor has supplied an accurate image, will highlight the exact drive connector on the user's board linked to a problematic drive. The nForce4 also embraces the new SATA II spec.

Like the NF3, today's nForce4 also lacks integrated graphics. One industry source told me that the IGP circuitry was still within the core logic but disabled. NVIDIA flatly denies this, but it raised the interesting question: Why did NVIDIA, a company famed for its graphics technologies, drop the IGP?

"The demand for Athlon 64 and Athlon 64 FX processors was focused in the enthusiast, gaming, and hard-core PC markets," says Henry, "and in this market segment, users demanded top-of-the-line graphics, which was not something that integrated graphics solutions provided. Now that AMD has done a great job in executing a top-to-bottom product strategy, ranging from value PCs all the way up to enthusiast-class machines, now is the right time to introduce a cutting-edge core logic solution with onboard graphics capability that can support a number of required features, including DirectX 9.0, Shader Model 3.0 technology, multi-monitor environments, and more. This is what NVIDIA will be bringing to the market later this year."

As mentioned, the standard nForce4 chipset lacks ActiveArmor. It also still only supports first-generation 1.5 Gb/sec SATA. The nForce4 Ultra remedies both of these points. All nForce4 models, however, support an incredible tweaking utility NVIDIA calls nTune. As an overclocking tool, nTune gives you control over system bus speeds, disk performance, memory timings, and GPU core and memory speeds, but that's only the beginning. The automated features are what set nTune apart from anything that has come before.

"If anything, we're having the greatest challenge with promoting our new nTune technology," says Henry. "It's very different in that it allows you to set up profiles that put the system into performance modes based upon the way that you're using the system. So nTune can detect that you're playing DOOM and will clock the system into a high-performance mode. If you want to play a DVD, it'll recognize that and put the system into a low-performance, quiet mode. It's so different from what people are used to that they're still trying to figure out how to take advantage of it."

nTune includes a performance troubleshooter, BIOS updating tool, system monitor, benchmarking tests, profile manager UI, and more. Users or system builders who want to overclock their systems to a few inches shy of the stability threshold can either spend many hours to days changing BIOS settings and going through countless reboots or they can simply perform a few clicks in nTune and let NVIDIA do all of the analysis and heavy lifting. If the thought of letting your customers overclock their machines makes your stomach churn, you can use nTune to disable any and all overclocking profiles, only allowing underclocking ones.

Far and away, the biggest buzz surrounding the nForce4 concentrates on NVIDIA's Scalable Link Interface, or SLI. The idea is that the user has two identical graphics cards sharing the rendering load equally. Depending on how the driver decides to deal with the application, this is either done by having the first GPU process odd frames while the second GPU does evens or by dividing each frame evenly between both GPUs. The rendered data is compiled and sent out for display via the primary card. An SLI bridge joins the two cards via the small edge connectors at the top of each card.

The SLI approach requires two x16 PCIe slots. SLI systems don't require two adapters, however. They can function just fine in traditional single-adapter mode. When you switch to SLI, though, the x16 slots are electrically converted to x8, effectively halving their PCIe bandwidth. Fortunately, x8 is still ample for even the fastest modern GPUs.

Some people, including myself, came away from the initial SLI debut with the impression that the feature was geared primarily at super-gamers who would go to any length and expense to outperform their rivals. Certainly, when armed with two 6800 GT or 6800 Ultra cards, which were two of the first SLI-compatible models, no other graphics subsystem on the market could come even remotely close. Knowing that only a relative handful of buyers take home $399 and up cards, however, I wondered if SLI was more of a hype generator than a real-world solution. Time and the release of the $199, SLI-compatible 6600GT has corrected my assumptions.

"Having two cards does not mean two cards at $500 each," says Ed Leckliter, director, channel products for Foxconn, which manufactures nForce4 motherboards. "It means having two cards. Then you're into the scenario of getting the bang of a $500 video card with two substantially less expensive cards. That's where NVIDIA is going. It's not about dropping two $500 cards in there."

While the SLI platform emphasizes having an upgrade path, Drew Henry notes that 95% of SLI systems sold go out the door with two cards installed. Most of these are now 6600 GT cards as two 6600 GTs running in SLI will outperform a 6800 GT.

A lot of confusion continues to swirl around what is and isn't SLI. Shortly after the technology's arrival, some board manufacturers started coming out with "near-SLI" implementations. MSI, for example, released an nForce4 Ultra board with an x16 and an open-backed x4 slot. By using early versions of NVIDIA's driver, this enabled the board to deliver SLI-type functionality that came within 5% to 10% of true SLI. NVIDIA quickly altered their drivers to prevent such approaches from working.

More recently, DFI's NF4 Ultra-D motherboard went to market with two x16 slots but no SLI bridge included. Like VIA, DFI maintains that the twin slots are meant to enable three- and four-head video output, not SLI. However, one prominent Web site shows how the Ultra chip can be easily modded with a #2 pencil (yes, seriously) and effectively turned into an SLI—the chip construction is that similar­—thus turning the Ultra-D into an SLI motherboard. NVIDIA is reworking the Ultra chip to prevent this mod from working in the future, and subsequent Ultra-D boards will carry the updated chip.

This is the geeky stuff that enthusiasts gobble up. That said, it seems unlikely that such cost-cutting "cheats" will—or should—gain any traction in the channel.

"If you try to turn an Ultra into an SLI, I'm not sure it says anything except you're trying to be cheap," says Foxconn's Leckliter. "SLI is SLI to the extent that they don't support Ultra for SLI purposes. Doing these little tweaks here and there is more a matter of smoke and mirrors. I don't feel it's something we need to do. When it comes time for the customer to open his wallet, I don't see many people wanting to pay for a non-official solution. For people who want to put two cards in their system, I'd be surprised if more than 10% of them would settle for an Ultra."

With each passing week, momentum continues to build behind SLI that goes unanswered by NVIDIA's competitors. And yet, SLI may not even be the biggest story of the year for NVIDIA. At long last, for reasons that neither company is publicly discussing, NVIDIA and Intel at last shook hands in a cross-licensing deal, and now NVIDIA is on the verge of releasing its first nForce chipset for the Pentium 4.

NVIDIA maintains that the Intel parts will be in production and on store shelves during its fiscal first quarter, which ends April 30th. Not so coincidentally, this would seem to coincide right about with Intel's Lakeport and Smithfield releases. No one would be shocked if one of the terms of the Intel/NVIDIA agreement was that the two companies release their new parts near or concurrently with each other. Given NVIDIA's burgeoning reputation in performance consumer circles—a demographic Intel had some trouble impressing with its 915/925X releases—having NVIDIA onboard with a high-end SLI part when Smithfield drops makes total sense, because the Intel nForce chips will all be ready for dual-core processors. And it just might help to distract NVIDIA from promoting AMD so heavily and thus stem some of that rising AMD market share.

"We will have a variety of products for the Intel marketplace—UMA and discrete, high-end, low-end, and mainstream," says Henry. "Our first product will be aimed more at the high-end, the gaming and performance segment, the segment that is more focused on some of our technologies and values performance. So when we go there, we will focus on speed and overclockability, because that's what that market is about."

Yes, you heard right—UMA. NVIDIA's initial Intel part, as yet not named publicly, will be discrete-only and SLI-enabled, but following chipsets will offer integrated graphics. Now that NVIDIA has conquered the high-end AMD segment, count on seeing the rebirth of IGP graphics on that side of the fence, too. NVIDIA states that all UMA-based chipsets will be based on PCI Express and will arrive before the end of 2005.

ATI

We all know the only half-joking rule that it takes Microsoft three tries to get any major new product right. With luck, the same will hold true for ATI's third stab at the core logic market. Back in 2002, ATI bowed its first attempt, the RADEON 320 and 340. These were RADEON 7000-based northbridges for the Intel and Athlon platforms that looked solid on paper but had significant performance problems, particularly in hard drive throughput. It was slow even for a part aimed at the value segment. The product did, however, find success in the notebook market.

Southbridge problems returned to plague ATI again in 2003 with RS300, or what became the MOBILITY RADEON 9100, which was based on the RADEON 9000 core. The northbridge was set to be an 865G slayer, complete with DX8.1 graphics and 800 MHz P4 support. But Intel was faster on the southbridge, ATI missed including a SATA interface, there were USB compatibility problems, and the product reached streets over half a year late. When the RS350 arrived and plugged all the holes, seriously trouncing all IGP rivals in the process, it was too little too late.

Which brings us to RS480 (IGP) and RX480 (discrete), also known as the RADEON XPRESS 200 series, ATI's first core logic stab at the enthusiast market...for AMD. The chips feature 20-lane PCI Express (22 including the SB400 southbridge), a 1 GHz HyperTransport interface, advanced power management, asynchronous bus speeds, integrated TV encoding, and (in the IGP part) full DX9 compatibility.

I found it curious that ATI had completely passed over the AMD platform in its prior generation, leaving NVIDIA to all of its nForce success, so I asked Reuven Soraya, director of marketing for ATI, why his company chose that route.

"With AMD, there was no opportunity for a technology transition. What I mean by that is when you release a chipset, it's good to line it up with some kind of a technology transition. We wanted to line up with K8 PCI Express. Just to line up with K8, we would have been late. So we waited for PCI Express, and that's what we did in November of last year with the RS480 and RX480. Now, the way it turned out was that most of our orders there went to the major OEMs for some obvious reasons. For channel products, they are being made available as we speak and will probably be on the street by the end of February from companies like MSI and Gigabyte."

I was lucky enough to be on the list of reviewers who obtained a "Bullhead" RS480 reference board sample kit last November, and I was very impressed. As part of my testing, I ran the RS480's integrated X300-class IGP against an Intel 915G board as well as an older RS300. Neither of the older IGPs would run DOOM 3. While the RS480 only pulled in just over 5 fps in 800 x 600 (high detail) mode, the point is that it ran without so much as a hiccup. Moreover, 3DMark tests show ATI clearly ahead of the 915G, and ATI's southbridge showed no trouble at all. Moreover, ATI's BIOS revealed a veritable blizzard of overclocking features sure to make any AMD enthusiast drool.

Now, in all fairness, the OC options in DFI's nForce4 Ultra implementation I saw were nearly if not on par with the Bullhead board, and no opinion should be final until we see how board vendors decide to implement RS480 and RX480 in real channel products. The one place ATI will chafe is its software implementation. The chipset group uses ATI's CATALYST driver set, which is very good from a support and stability standpoint. The CATALYST group is now famous for its responsiveness and quality. But ATI isn't even close to having the sophisticated utilities, such as NVRAID and nTune, found in NVIDIA's Forceware. Time will tell if the market will hold this against the newcomer.

But the whole matter doesn't boil down to software, either. My RS480 sample had both VGA and DVI outputs onboard. ATI also has a killer feature called SURROUNDVIEW that keeps the IGP live while an ATI discrete graphics card is also in use, enabling triple-head output. (Theoretically, quad-head support should be equally possible.) No other chipset offers this. And so long as NVIDIA's IGP remains a promise rather than a reality, ATI has the de facto best IGP chipset available today.

On the Intel side, ATI will unleash the RS400, or RADEON XPRESS 200G, which is a Socket T version of the RS480. According to ATI, partner boards based on RS400 should start hitting the streets shortly after you read this.

"We're affordable, but with our features, we're giving you a good out of the box experience," says Soraya. "If you're only paying $600 or $700 on a PC, you're probably not an enthusiast gamer, but you might like to play games every now and then—something modern like Far Cry or DOOM 3. These will run on the RS400 at playable frame rates. This means that the customer will not be picking up the phone to call his system integrator and say, ‘I just bought this PC, and this thing doesn't work!' That might be the case with an SiS or an Intel chipset. We are bringing good graphics performance even to a $600 PC. You can even run Media Center on our chipset. You can't do that with SiS. Moreover, unlike competitors, when you upgrade our system to external graphics, you don't lose the integrated graphics. It's still online, able to power another monitor."

Once again, the southbridge may become a cause for concern. The SB400 supports eight USB 2.0 ports, four PATA drives, four SATA 1.0 drives, 7.1 AC'97 audio, and five PCI slots. RAID 0 and 1 are offered, but ATI has made no mention thus far of any more advanced RAID features, and by 2Q, it's going to start looking a bit late not to have SATA II. The follow-up part, the SB450, which will deliver HD Audio, is slated for the first half of 2005, but no other details were available on this by press time. Because ATI's northbridge is independent of its partner, board manufacturers also have the option to go with a third party southbridge. For example, at least one instance of the RS480 paired with ALi/ULi's M1573 already exists.

And what about a dual x16 solution for the enthusiasts ATI is courting?

"We do have a dual graphics slot strategy," says Saroya, "and as we get closer to having something to talk about, then we'll talk about it."

Staying mum on the subject may not be a bad plan for ATI at this point. The last thing ATI needs is a black eye like VIA's, having the hardware in place but sitting around and waiting for NVIDIA's cooperation to move forward with it—as if that would ever happen. If NVIDIA's claim that one-quarter of all nForce4 chips being sold are SLI, and that 95% of those are carrying a pair of NVIDIA-based graphics cards, then the pressure really is on ATI to come up with its own dual-slot solution, and no doubt VIA would welcome a more cooperative partner to make use of its own dual-slot potential. But ATI can't just copy SLI. To do that at this point in the game would be suicide. ATI needs something different, substantially better, or both, and it needs it fast. On this of all points, though, a premature move for ATI could be disasterous.

All in all, ATI's RS400 looks to be a 915G killer. But does that matter on the eve of the 945G's debut? The answer is yes if only because ATI's track record indicates that it will surpass the 915G on performance while substantially undercutting it on price, and that's good news for system builders looking to craft value-rich mainstream boxes. With no production boards visible as of this writing, RS400 looks to be a DDR1, 800 MHz FSB, single-core solution. In other words, it's got all the mainstream specs with a bunch of enthusiast performance potential under the hood.

For the next few months, this should suffice. DDR2-667, the first iteration of the newer RAM format to show a definite advantage over DDR1, still needs a lot more price erosion, and there's no sign of 1,066 MHz CPUs being even remotely affordable for quite a while. If ATI can deliver dual-core, EM64T, and XD bit support either in first-generation channel shipments or through a subsequent BIOS update, then there's no reason why the market should reject it. And besides, the rumor mill says that ATI will have a chip ready to counter Lakeport when the time is right. If this follow-up to the RS400 comes off like the RS350 did—only on a prompter time schedule—then ATI should be sitting very pretty to meet the challenge of all rivals heading into the second half of the year.

Why Chipsets Matter

Chipsets matter at the highest industry levels because they are a means of controlling the direction of technological innovation and the experience had by end-users.

"Intel is in the business of chipsets as much for the overall PC market control as for money," says IDC's Shane Rau. "Intel does the research and development for a great deal of the PC industry, and if Intel wants to drive the PC form factor forward, it can use chipsets to help the industry adopt new technologies. With its market share, it has the ability to drive things like PCI Express, DDR2 memory, and Serial ATA into the marketplace. If you add on the licensees of Intel—companies like VIA and SiS—you're approaching 90% control of the entire chipset market."

This is also why companies like Intel and NVIDIA have realized that tying software to the chipset is so important. Once a user gets hooked on NVIDIA's firewall or utility of nTune, he's not likely to ever migrate away and will, in fact, endeavor to change his other PCs over to a similar feature set. Specs are not enough. Tomorrow's bus speeds will be higher than today's, but a value-rich experience set can remain more or less unchanged year after year.

In the end, it falls to you, the system builder, to translate the potential within these chipsets into tangible end-user benefits. If you can tweak these feature opportunities into unique solutions for your clients' needs, you'll have nothing to fear from larger, deeper pocketed competitors.