   |
|||
|
|||
|
:
News
: Archives
|
|
NVIDIA GeForce 7950 GX2 Review |
|
|
More than the Sum of its Parts? |
|||
|
|
|||
|
by Josh Walrath |
|||
|
The 7950 GX2 The basic specifications of the 7950 GX2 are pretty straightforward. It is comprised of two 7950 GPU’s (G71 Mobile) running at 500 MHz a piece, 512 MB of 1200 MHz GDDR-3 for each board (a total of 1 GB of memory with a combined bandwidth of 76.8 GB/sec), and a PCI-e switch chip that features 48 lanes. The video card is divided into two PCB’s with a single slot width cooler for each GPU. This makes the card two slots wide overall, but this also makes for some interesting cooling issues. One GPU cooler is more exposed to free flowing air, while the other due to being sandwiched between the two PCB’s will be somewhat restricted, thereby making that GPU warmer. This consideration is not all that large considering NVIDIA uses a mobile binning of the G71 so it does not suck as much power nor create as much heat as the desktop binned G71 chips.
The two single slot coolers make the GX2 about the same height as the 7900 GTX. Both require a single PCI-e 6 pin power cable and about the same rated power supply. The 1 GB of video memory is a bit misleading in that the memory is not shared between the cards. This means that each chip can only address 512 MB, and due to the rendering schemes each card has to replicate data across the boards. Effectively the card has enough space for 512 MB of data. This is unlike the old Voodoo 5 which would require half the frame buffer space per VSA-100 and then only had to replicate the texture data. The board sports two dual link DVI-I outputs as well as HDCP. The outputs include the special chip which enables HDCP, so high definition content can be output by the dual link DVI. This will not be much of an issue for the next few years, but once HD-DVD and Blu-Ray require the presence of HDCP it will be much more important. The TV output also can handle high definition resolutions, as long as the necessary component cable is supplied. While using this board with both LCD’s and CRT’s, I never noticed any rendering issues, and the output was crystal clear up to 2048x1536. The GX2 is also the only “single” card that supports SLI-AA. Not only do the mixed anti-aliasing modes get a nice performance boost because of AFR/SFR, but because of the close proximity of the two GPUs and the nearly direct PCI-e connection, SLI-AA gives good enough performance to use on a wide variety of applications at higher resolutions. Both 8XS and SLI-8X are excellent performers, and many old (and some new) games can use those schemes at resolutions above 1600 x 1200. Games such as my own personal favorite, Morrowind, gain another degree of impressiveness when running at 1920 x 1200 with 8XS or SLI-8X AA. To get such a product to market though, NVIDIA had to rely on its expertise in SLI technology. The two G71 chips use a board to board connection like the “over the top” SLI solution on standalone cards, but it is routed through the PCB’s differently. NVIDIA also designed and implemented a 48 lane PCI-e switch, which provides 16 lanes to each GPU, as well as 16 lanes to the PEG connection. This means that each GPU realistically only has 8 dedicated lanes to the system, but each also has 8 low latency lanes from one GPU to the other for communication needs. This could very well have a positive effect on SLI-AA performance since it utilizes the PCI-e bus rather than the “over the top” connector. The changes made from the 110 nm G70 to the 90 nm G71 also made the architecture a lot more efficient in terms of power and transistor costs. By utilizing the faster 90 nm Low-K process from TSMC, NVIDIA could reduce the number of transistors needed to enable higher clockspeed. Typically in pipelined ASICs, by breaking up the pipeline into stages there is a lower propagation delay between the stages, so the clockspeed ceiling can be increased. What NVIDIA was able to do is utilize the faster transistor performance of TSMC’s 90 nm Low-K product, and cut out a lot of redundant transistors. This made the chips die size smaller, and perhaps more importantly there are fewer transistors to power. Again, NVIDIA apparently really optimized the SLI circuitry in the 90 nm G7x chips, making it much leaner as well as better performing. Each G71 chip features 16 ROPS, 24 pixel shader pipelines (2 ALU’s per pipe), 24 texture units, and 8 vertex shaders. The G71 is comprised of around 278 million transistors and takes up around 196 mm square of die space. When all of these hardware steps are combined with what is now 2 years of software development in SLI, we can see that the 7950 GX2 is in fact a very mature product which comes by its performance honestly. This is not some Frankenstein creation that will find a home in very few, select PC’s. The GeForce 7950 GX2 is a mass produced video card that has found a lot of acceptance in the marketplace. Now that prices are dipping under $500, more and more users are considering this product. The Empty Promise of Quad SLI NVIDIA’s dedication to SLI is truly impressive, and when we were introduced to the concept of Quad SLI late last year, many of us thought it could become gaming nirvana. Most of us also thought that there would be some major technical issues to solve before we could get there. Since NVIDIA had so much experience with SLI, we expected these issues would be solved and by now Quad SLI would be a viable product for the ultra-highend. Turns out we are not quite there yet. Many recent articles have utilized the 91.47 drivers and re-tested Quad SLI with 2 x 7950 GX2’s with a variety of motherboards. For the most part these Quad SLI machines actually performed worse than a single 7950 GX2. The issues we are seeing are probably related both to the hardware, and mostly to the software and how the OS handles multi-GPU systems. The most beneficial rendering scheme for these cards is AFR, but Direct3D only allows a maximum of 3 frames to be buffered, which makes for some interesting scheduling with 4 GPUs. OpenGL does not have that issue, but the majority of current applications are all Direct3D based. Add to this the situation of NVIDIA’s upcoming new architecture, as well as Windows Vista and DX10, we see very little interest in convincing customers to pay $1000 for two 7950 GX2s. It makes no sense for individuals wanting to spend this money and not wait for the next generation of products that are right around the corner. This does not mean though that a single 7950 GX2 is obsolete already! NVIDIA’s upcoming parts will undoubtedly be more expensive than the GX2, and Windows XP will not get the full DX10 treatment. So, if a user wants very good performance for quite some time to come, but is not terribly interested in buying Vista, then this will be the top choice. Even if a user wants to go to Vista, the GX2 prices will go down when the next generation cards are released, and these could still very well be an outstanding bargain. One other consideration is that DX10 and Vista will be much more multi-GPU friendly than Windows XP, so we may yet see another performance increase from the GX2 in the near future.
If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2006 PenStar Systems, LLC. |
|||
