|
||
|
|
|
              |
|
 |
 |
||||||||||||||
|
 |
|
||||||||||||
By William Van Winkle |
||||||||||||||
“With one year of the AMD/ATI merger behind us, perhaps now is a good time to pause and ask what the merger has accomplished for AMD, the channel, and the industry as a whole. A cynic could point to launch delays, market share numbers, and the fact that the HD 2000 family didn’t leapfrog NVIDIA at the very top of the GPU spectrum. But cynics are too busy wallowing in their own ideas of what should be and ignoring the potential upsides of reality today. If you sit around pinning your hopes on AMD’s “Fusion” effort, which will build GPU technology into multi-core CPUs, you’re going to miss the excellent products and platforms AMD has on the table now-—products from which you and your customers should be profiting. |
||||||||||||||
|
A year ago, former executive vice president of AMD Henri Richard commented, “We’re establishing clearly that this is a two-horse race. And as you would expect in a race, sometimes, when one horse is a little bit in front of the other, it reverses the situation. But what’s important is that it is a race.” If nothing else, a two-horse race would likely guarantee that both horses run even faster than they would in a larger pack. You’re either winning or losing; there is no comfort of a silver medal. And the faster and more furiously the contenders race, the more the channel and industry stands to benefit. Never mind raw performance, look at the secondary benefits in power savings that have resulted from the AMD/Intel rivalry. Look at the architectural advances paving the way for future functionality. The price war between these two titans may not benefit either’s bottom line, and whether falling ASPs benefit resellers is up for debate, but there’s little question that the end-user wins simply by getting more for his money. Resellers who do the best job of educating buyers about this increased value stand to gain the most leverage from this battle. The reason Richard commented on AMD engaging in a two-horse race is that it seems increasingly apparent that an integrated CPU/GPU/chipset platform is the key behind having control in the PC space for the coming years. Solely on the strength of its stable integrated graphics chipsets, Intel has been able to claim market share leadership in the graphics world even though its IGPs have historically trailed not only every mainstream discrete GPU at the time but also rival IGPs from ATI and NVIDIA. But when you control graphics you get to sell more CPUs, which in turn influences memory formats, virtualization technologies, and much more. Many believe that dominance in discrete graphics also weighs heavily on share in integrated graphics, which is why you now see Intel preparing to bolster its position with its own discrete graphics family, currently code-named Larrabee. The first public references to Larrabee, and apparently the formation of the Larrabee group within Intel, date from December 2006. This is a little less than five months after AMD announced the ATI merger (and presumably after Intel and NVIDIA realized their futures might not be as intertwined as some suspected). So for those cynics who think that AMD doesn’t have the power to shape events and trends throughout the industry, think again. Some things just take time to reach fruition.
SPIDER HATCHES Spider is the code name for AMD’s first concerted CPU/graphics/chipset platform play on the desktop. We have to step back and qualify the term “platform.” Historically, a “desktop platform” has implied a group of components that all interact both cooperatively and exclusively. In theory, this exclusivity exists in part to let the manufacturer do a better job with validation and optimization. AMD traditionally shunned the platform concept, saying instead that it preferred to allow choice. Admittedly, this is easy to say when you don’t have a graphics or chipset group with which to complete a platform. Now, however, AMD finds itself with all of the necessary platform ingredients and a decision to make on how it will position these pieces going forward. How can the company offer a complete desktop platform without negating its long-standing messages about industry choice? Let’s dig into the platform and find out. Spider comprises three segments: ATI Radeon HD 3800 series GPUs, AMD Phenom processors, and the AMD 7 Series chipsets. (Yes, insects have three body segments while arachnids have two, but they couldn’t call it Ant, now could they?) As you would expect, the GPUs are platform-agnostic. An HD 3800 series card will work on any PCI Express platform; the limitation is driver support, not processor or chipset compatibility. But the advantage of having all three of these platform elements under one roof, so to speak, is that there is now only the proverbial one throat to choke—-one source for all support. There is also the reassurance that these pieces were developed by the same company and therefore have that extra level of cross-validation from in-house testing. A platform should be about the whole equaling more than the sum of its parts. In Spider’s case, that something extra comes from single-source support rather than proprietary hooks. “The channel has really been hungry for an all-AMD solution,” notes Gary Bixler, director of North American marketing for AMD. “There’s always been debate and consternation about the relative merits and performance and stability of the Taiwanese infrastructure versus an all-AMD infrastructure. But now we’re excited to deliver that across the three major silicon building blocks. That’s one of our key messages to the channel—the strength of that all-AMD solution.”
7 SERIES: THE FOUNDATION Back in February of this year, AMD launched the 690G (RS690), the company’s first desktop chipset since the AMD-762 (for Athlon MP) in May 2001. The RD480/570/580 chips that went to market under ATI’s CrossFire Xpress brand were good products, but the 690G delivered a lot of sorely needed improvements. Chief among them were the Radeon X700 integrated graphics core, HDMI and HDCP support, and dual-core CPU compatibility. The 690G remains a very solid chipset, which is why it still fuels the desktop boards in AMD’s Validated Solutions stable image program (see www.amd.com/validated). In the hubbub of Spider’s launch and the 7 Series chipsets, don’t lose track of the 690G as a viable value play. That said, the 7 Series is here, marking AMD’s return to high-performance discrete chipsets and a move from the 80nm fab process of the RS690 series into 65nm designs. This fab shrink is a key element behind AMD’s messaging because it’s a large part of what enables the debut 7 Series chipset, the enthusiast-class 790FX, to operate in an 8W (nominal) to 12W power envelope. Consumption drops to 3W at idle; the forthcoming 780G sips only 1.2W at idle. Add on the SB600 southbridge, and Spider’s chipset will still average around only 15W. This is twice the power efficiency of the prior ATI-based chipset, and AMD isn’t shy about pointing to the Intel X38 rival, which operates in the 26W to 42W range. (Mind you, the Intel chip integrates a memory controller while the AMD platform puts its memory logic in the CPU.) You may recall how much AMD emphasized power efficiency as a cornerstone of its Barcelona (quad-core Opteron) launch. That message now carries down into and expands within the desktop platform. At first glance, power savings may not be a huge incentive for the types of high-end users likely to embrace the 790FX. These are performance-driven buyers accustomed to not watching their power bills. However, the opposite side of the power coin is noise. With a 15W TDP, the need for active (fan-based) cooling drops significantly, even with overclocking. This is why you see the first round of 790FX-based motherboards showing up with lots of copper and heatpiping, but not active chipset cooling. So not only are boards quieter, but all that copper looks pretty sweet behind a clear side panel. Consider offering some internal LED lighting to show off all those gorgeous grilles.
Beyond power savings, there are a few other elements of the 7 Series that deserve close scrutiny. HYPERTRANSPORT 3 HyperTransport is a packet-based I/O interconnect technology maintained by the HyperTransport Consortium but backed foremost by AMD. At its simplest, HyperTransport is a conduit for making chip-to-chip or board-to-board connections. AMD/ATI, NVIDIA, Transmeta, and others use HyperTransport in contrast to Intel’s Direct Media Interface (DMI). The flexibility and low latency aspects of Hyper-Transport are what allow the technology to compete with DMI—and presumably surpass it. (Intel will be dropping the DMI next year in trade for a new interconnect called QuickPath that mimics HyperTransport in several ways.) Whereas DMI is used exclusively as a chipset interconnect, HyperTransport not only does this but also connects CPUs, add-in cards, and more. As you start hearing more about modular co-processors in AMD’s Torrenza initiative, Hyper-Transport is the technology making those high-speed component links possible. First-generation HyperTransport featured a top clock speed of 800 MHz, enabling a maximum 12.8 GBps aggregate bandwidth with 32-bit links. HyperTransport 3.0, now debuting with the 7 Series chipsets, hits 2.6 GHz with a maximum bandwidth of 41.6 GB/s. QuickPath is expected to arrive with 24 to 32 GB/s of bandwidth per link. The third generation also introduces features such as link splitting (for example, splitting a 16-bit link into two 8-bit links), hot-plugging, and PCI Express mapping. “HD content is like four tractor trailers strapped side-by-side from a width perspective,” says Ian McNaughton, senior product manager for AMD. “On a typical data path, we’ve got system memory going to the CPU and from there down to the northbridge, and from there it’s PCI Express out to the graphics cards, right? Pretty simple. For those graphics cards to access system memory, that’s the route to travel up. But say we’re using four graphics cards—-just a massive pumping of bandwidth. HyperTransport 3 enables the doubling of those lanes to accommodate those tractors. With the things that are happening in gaming now with higher resolutions, it’s required to have broader lanes for bandwidth. As 30-inch monitors become more mainstream—-hey, don’t laugh. Who knew a couple years ago that we’d be able to buy a 22-inch widescreen for $198? So 30-inch on the desktop, 70-inch plasmas. That stuff is going to become normal in the home. As we increase the resolutions and the requirements to drive them, HyperTransport becomes much more important. You could still drive that resolution with HyperTransport 1, but it would be a slower, less satisfying experience.”
PCI EXPRESS AND CROSSFIRE X In the same vein as HyperTransport, PCI Express 2.0 doubles the bandwidth of the original version that debuted in 2004 and is only now starting to phase out. Version 1.0 featured 250 MB/s bandwidth per link in each direction. Version 2.0 doubles this to 500 MB/s, and draft 3.0, not expected to ship until 2009 or 2010, will scale to 1 GB/s per link. As a point of reference, one PCI Express 2.0 lane offers four times the bandwidth of a conventional PCI connection. Keep in mind that 16 of these lanes can connect to a single x16 graphics slot. Just before launch, AMD threw around some numbers to give a sense of what widening these pipes could accomplish. With a dual-core CPU, HyperTransport 1.0, and PCIe 1.0, you could run an ATI Radeon HD 3850 (see below) at 1920x1200 and assume a baseline in performance of 1X. Stepping up to a quad-core Phenom chip (also below) and PCI 2.0 yielded a 10% performance improvement, but this left the HyperTransport connection as the bottleneck. Increasing this to HT3.0 yielded an end gain of 20% over the baseline. While this scenario doesn’t make a night-and-day case for quad-core over dual-core in 3D gaming, it’s interesting to see the real-world benefits of expanding these two pipelines. Let’s be clear here. There are 38 PCIe 2.0 lanes in the 790FX chipset. Six lanes veer off for x1 connections to devices such as network adapters and storage controllers. Then you have two x16 connections, 32 lanes total, dedicated to four graphics slots. That’s right—-Spider supports up to four graphics cards on one motherboard. With all four slots populated, each slot would utilize eight PCI Express 2.0 lanes. You can do the math. The per-card bandwidth of four cards on a Spider PCI 2.0 platform is the same as two x16 cards on PCI 1.0. Given that we’re still a long way from tapping out 4 GB/s of bandwidth from a GPU, there’s no need to worry about bottlenecking. The bandwidth per card may stay the same as in the prior generation, but PCI Express 2.0 now lets you run and/or aggregate more cards.
This takes us to CrossFireX. In a nutshell, CrossFireX makes no changes to the basic modes that traditional CrossFire uses for 3D acceleration. Rather, there are three main feature enhancements. First, for those who don’t care much about 3D acceleration and just want maximum screen real estate, a Quad CrossFireX array (meaning four cards) enables output to up to eight monitors. This seems laughable, but some games actually don’t need acceleration as much as they need extreme resolution. And just ignoring games, the number of users adopting triple, quad, and even higher monitor output for non-professional tasks continues to climb. We almost wonder if Quad CrossFireX might take a small chunk out of NVIDIA’s Quadro NVS market—-just for those who might value budget multi-monitor deployment over software validation. Second, in a conventional CrossFire setup, when CrossFire is enabled, the 3D application runs on one screen, receiving all of the acceleration, while the other displays go dark. CrossFireX now allows those secondary screens to stay active. The applications running on those monitors won’t be accelerated, but at least they’ll be visible. So, for the first time, a user will be able to keep half an eye on his stock streamer or email inbox while enjoying a CrossFireX-boosted title. For the record, no one currently has the technology to enable single-app 3D acceleration across multiple monitors, although it’s obviously a priority for all major players. One unofficial estimate we heard recently pegged the task of developing such technology at 100 man-years. Until then, we can enjoy the fruits of CrossFireX’s third main benefit: performance scaling. Now you can throw the combined might of four GPUs against a single task to maximize acceleration. AMD’s McNaughton estimates that users will see acceleration scaling of roughly 2.7X with three graphics cards and 3.2X with four. “For system builders, CrossFireX is about the upsell,” says AMD’s McNaughton. “What’s the attach rate in new builds for three and four cards? You know the answer. It’s going to be incredibly low. But it’s not about that. It’s about the ability to scale your platform if you want to. We’re offering our consumers the ability to scale to four graphics cards and eight monitors. If they only want to run one graphics card, they can. The platform doesn’t cost them any more money. If we didn’t have it, then it would be a limitation. It’s going to help builders sell more motherboards.” For a bit of stunned disbelief and proof that end-users will buy eight monitors (or more!), check out www.wideview.it/pictures.htm.
QUAD FX The 790FX chipset also enables AMD’s Quad FX, the dual-processor desktop platform. Not one reference to this fact shows up in AMD’s pre-launch Spider literature. However, we know that the 790FX supports Quad FX, and at an analyst event in 2006, AMD’s Phil Hester showed a presentation noting “4x4+” dual-processor, quad-core desktop designs for 2007 and “4x4++” (more of the same, only with DDR3) for 2008. Quad FX earned a lot of buzz when it launched, but it yielded hardly a blip in terms of channel sales. Some of this could be blamed on the added expense, some on the high-end DIY audience, and maybe some on the fact that realizing the platform’s full potential required using the top-priced Windows Vista Ultimate Edition. Because none of those factors have changed as of Spider’s release, we can only assume that AMD is holding off on making a fuss about the feature until next year. OVERDRIVE On one hand, AMD will still tell you that it doesn’t condone overclocking of its CPUs. On the other hand, the 7 Series brings us AMD OverDrive, sometimes also called AutoXpress. NVIDIA has enjoyed some recognition for its nTune overclocking routines, enabled through the nForce chipsets. Here AMD looks to provide an even more comprehensive toolset. “Our OverDrive has an easier interface [than nTune],” says AMD’s McNaughton. “Also, ours creates a better-performing platform. There are quite a few cases in reviews where nTune will lower the performance of your system, because they have some set configurations. It’s not a smart tool. It just says to go from A to B regardless of what processor, chipset, or graphics you have. Ours actually runs tests in the background to see how stable your platform is. Then it’ll either increase the OverDrive aspect to get more performance if you’re still in the boundaries of stability or it’ll back off if you’re getting close to the stability line.” OverDrive has been a lot longer in the making than you might suspect. In early 2005, ATI went to Finland and recruited Sami Maekinen, one of the world’s most accomplished and record-holding overclocking enthusiasts. Planted in Toronto, Maekinen was ultimately given the job of taking most of what he knew and pouring it into an application that could perform its magic with one button click. This became the heart of OverDrive’s novice mode, which features only an oversized button and a slider that goes from “default” to “high performance.” Naturally, there’s an advanced mode in which experts can take granular control over their many voltages, frequencies, multipliers, and bus speeds. OverDrive also includes a dynamic tool that can automatically overclock on the fly. And don’t overlook the software’s integrated benchmark, which you can use to quickly and easily demonstrate the utility of OverDrive to potential buyers. Obviously, the OverDrive feature is not a channel differentiator; OEMs have it the same as mom and pop shops. But OEMs can’t show it to would-be buyers and say, “Just nod and I’ll give your system an expert-class performance boost for free.” You can. ...more |
||||||||||||||
 Back to top
|
||||||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
||||||||||||||
