Watching for AMD's 780 Chipset

By: David Kanter | 05-04-2008

It was nearly 2 years ago that AMD announced the intention to purchase ATI - one of the most surprising moves in the PC industry. Initially, it seemed like little had changed. After all, AMD focused on CPUs and ATI on chipsets and GPUs. The most immediate difference was that Intel and NVIDIA became somewhat friendlier than usual (for a little while) and much of ATI's chipset business disappeared. As time passed, AMD's fortunes waned as Intel offered a rejuvenated product line. At the same time, AMD was busy trying to unify two separate roadmaps for technology and development by becoming a full platform provider, rather than a platform agnostic CPU vendor. This required a substantial change in mindset, as previously, AMD had always seen the chipset as something that they must do for the first generation of a given processor. It was a chore though, and one that they would quickly abandon in favor of 3rd party partners, such as VIA, SiS, NVIDIA or ATI. Adjusting to a 'platform' strategy involved recognizing that OEMs really wanted a full solution, rather than simply a collection of components.

The first combination of CPU and chipset that AMD really emphasized as a platform was the 'Spider platform', which was targeted at high-end gaming systems, using an AMD CPU, ATI chipset and ATI GPUs. The problem with this is that high-end gamers know and want the highest performance components and are willing to pay for them - but unfortunately AMD CPUs are not the highest performance, nor are ATI GPUs. Naturally, this is why the Spider platform has not been embraced by gamers.

Two months ago, AMD announced the 780 chipset, which forms the basis of their volume notebook and desktop platforms. The 780 chipset is a brand new chipset manufactured on TSMC's 55nm process. Since this news is a little old, we won't dwell on the minute details of the chipset, except to point out that most everything has been improved over the prior generation, as covered elsehwhere. Instead we will focus on why the 780 chipset is interesting and important to AMD. Specifically, it is AMD's first credible platform product, their first DX10 integrated graphics (IGP) solution and it also contains novel technology for assymetric GPUs.

Platform and Graphics

The first thing to note is that unlike the 'Spider platform', the 780 chipset and the associated platforms (Puma and Persius) are actually well suited to the target market. While AMD does not have the highest absolute performance for CPUs or GPUs, they have good performance at very attractive prices. Top notch performance is essential for the high-end of the market, hence the lukewarm reception for 'Spider'. In the mainstream market though, a complete feature set and affordable performance at a given price level is much more important than having a system which breaks records for the latest benchmarks. So from the start, the Puma platform seems much more likely to be a success than previous platforms from AMD. It also helps that the 780 chipset excels where the competition from Intel is weakest.

AMD's primary focus for this chipset was on the graphics. The northbridge integrates what was once a full discrete DX10 GPU, the RV610 (originally fabricated on TSMC's 65nm process). This was an unusual step for AMD; traditionally ATI IGPs use half the functional units of a low-end discrete GPU to save on die area and power. The IGP for the 780 chipset uses system memory through a 3.6GT/s HT3 link (intriguingly, HT3 is not enabled for servers yet) or can be configured to use dedicated memory (GDDR or DDR) for extra performance. As a result, the IGP performance is quite impressive - roughly a 2-3X improvement compared to the 690G chipset. The most comprehensive review out there is probably from Tech Report and shows a huge performance advantage (upto 10X on Half-Life 2) relative to integrated graphics on Intel's G35 chipset.

That being said, the G35 is a year old and Intel is planning to upgrade their integrated graphics in Q2/3 of this year. It's clear they won't actually catch up to the 780 chipset, but the performance should be less embarrassing. The comparison is not totally fair without mentioning this, but it makes the point clear - AMD currently has the highest performance integrated graphics solution by a fair margin. NVIDIA's GeForce 8200 IGP should perform at roughly the same, level but it is barely available in North America and initial versions may not have functioning Hybrid SLI support.

The 780 chipset will really be the first time that AMD has aggressively begun to market and sell the platform as a whole. The reception from OEMs will be one of the first proof points of the ATI acquisition, and it will also give a hint as to how well AMD is executing on their plans. It should also indicate whether a great chipset from AMD can make up for a lackluster CPU. All of these are key questions for AMD going forward.

Hybrid Graphics

The 780 chipset is the first product to use a "hybrid" multi-GPU set up, aptly named, Hybrid Crossfire. Hybrid Crossfire operates a discrete GPU (HD 34xx) in tandem with the IGP to boost performance above what either could achieve separately. Rendering on two identical GPUs is old hat - it's been around since 1998 and typically frames are either divided in half (Split Frame Rendering), or alternate frames are rendered on different GPUs (Alternate Frame Rendering). Using two GPUs with different characteristics (memory capacity, latency, bandwidth, functional units, etc.) is vastly more complex, since it isn't clear how to parallelize work efficiently if the two aren't exactly equal. This is a general problem, not specific to GPUs - although we expect that it would be easier with GPUs, since the parallelism is abundant.

The initial implementation of Hybrid Crossfire is very difficult to judge. Performance gains relative to a discrete GPU only vary across all the different reviews on the web. There are results ranging from -5% to +100%. It seems like the improvement is extremely sensitive to the game, resolution, graphics API and other factors - it would be nice to see some real sensitivity studies which actually pin-point where Hybrid Crossfire helps the most and where it isn't effective. If the average benefit is 40% or higher, that seems like a very nice feature - at 10-20%, it's far less exciting. Hybrid Crossfire is restricted to certain GPUs, which is very understandable. Hopefully, the second generation will probably be much better as drivers receive more attention and engineering resources.

There are also a lot of technical questions remaining to be answered, such as how Hybrid Crossfire performance varies with different integrated graphics configurations (dedicated memory versus system memory, HT1 vs. HT3, etc.), but it seems like a very reasonable first implementation of a good idea.

Hybrid Crossfire is a real platform feature that relies on the interactions between two different components to provide added value (a bit of extra performance). In many cases, a 'platform' is just a combination of components with a special name and there is little or no unique benefit to buying them together. For instance, the Live! platform from AMD really was more about marketing than adding value - it didn't matter which wireless solution was present, as long as there was one there (although it did guarantee wireless). Features like Hybrid Crossfire actually make platforms a little more interesting and relevant to users. Whether Hybrid Crossfire ultimately succeeds will be measured by the attach rate for ATI cards. This should be easy enough for AMD to find out at the OEM level, but it will be very tricky for users who are purchasing add in boards separately.

The platforms based on the 780 chipset are novel both from a technical and a business perspective. The Puma platform represents a big opportunity for AMD. The platform push could increase AMD's share of the AM2+ chipset market due to two factors. First, OEM preference to have two (or three) components from a single source and second, marketing and selling the CPU and chipset (and maybe GPU) as a bundle. The Puma platform could also increase AMD's share of the processor market, if OEMs find the integrated graphics performance appealing relative to Intel's integrated offerings (at the low-end) and more attractive from a cost perspective than a low-end discrete GPU. In the coming months, the reception for these platforms should be clear, and will provide a great deal of insight into AMD's product portfolio, their approach to the market and their future evolution.


<< Prev News