: News
: Reviews
: Editorials
: About
: Contact
: Advertising
: Privacy
: Links
: Forums



NVIDIA SLI Certification Program


Why You Should Care


by Josh Walrath



            We are surrounded by certifications and marketing schemes.  Intel Inside, Authentic AMD, Cisco Certified, Designed for Microsoft Windows, Premier this, Enabled that… you get the picture.  So why write an entire article about NVIDIA’s new certification program for SLI use?  I was asking myself this very same question when I was first asked to talk to the folks at NVIDIA about this.  I was sent a PDF with the draft on SLI certification, and I thought “pure marketing”.  Once I really delved into it though, I started to change my mind.  Once I actually talked to the people behind the process, I now feel that it is very important.  Anyone who is remotely interested in SLI should actually give this certification more than a passing glance.

            NVIDIA started work on SLI (Scalable Link Interface- and not 3dfx’s Scan Line Interleave) about 3 years ago.  Ever since the release of the GeForce 4 Ti, NVIDIA’s GPU’s can be used in multi-chip configurations, which is what Quantum 3D base their visual simulation products on.  This feature is not very helpful to the average user, as buying a multi-GPU rendering board from Quantum 3D is really not an option (unlike the old days with the Obsidian brands).  So to make multi-GPU setups affordable for the consumer market, NVIDIA had to architect a new strategy.

            There are three areas that NVIDIA addressed to bring affordable multi-GPU tech to the consumer market.  The first is at the GPU level where NVIDIA has spent a good amount of time with the actual silicon of the GPU.  While older GeForce GPU’s had the ability to be placed in multi-GPU applications, it was more of a brute force affair that relied more on the software and drivers to achieve good performance.  This was not the most efficient way to do this, so I am assuming that in late 2000/early 2001, when NVIDIA acquired 3dfx, they put some of these 3dfx engineers on a next generation multi-GPU project.  While scan line interleave is no longer used, they have totally redefined the technology behind multi-GPU applications.

            The second area that NVIDIA had to address was that of the motherboard level technology.  While the AGP 3.0 specification allowed the use of multiple AGP slots on a board, no chipmaker ever implemented this feature into any chipset to date.  There are some other issues with AGP that did not lend itself nicely to multi-AGP slot products, and most of these centered on data delivery and data transfer.  While AGP 8X delivered 2 GB/sec throughput, it is not set up for tight isochronous data transfer.  If there are two AGP slots delivering that type of bandwidth, there could easily be a data skew between the info being delivered to each card.  The hurdles to get two AGP cards to work in tandem on separate AGP busses would have been huge, and there is a good reason why no one has ever attempted this.  PCI-Express offered a true next-generation approach to expansion cards, and it made multi-GPU systems a possibility.  Not only does PCI-E offer great bandwidth (4 GB/sec upstream and downstream bandwidth in a X16 layout, for a total of 8 GB/sec), but it also has the ability to transfer data in a tight isochronous fashion.  PCI-E is far less error prone than the old PCI and AGP standards, and any kind of problem that data skew might cause is offset by the embedded clock information sent with each data packet that PCI-E delivers.

            So with this in mind NVIDIA set to work on designing their nForce 4 series of PCI-E based chipsets, and the big push was to develop a solid SLI supporting platform.  NVIDIA’s experience with the nForce, nForce 2, and nForce 3 series of chips unfolded nicely when it came time to release the nForce 4.  This single chip platform features incredibly low latency, since it handles all of the different busses from one piece of silicon.  When combined with the very powerful AMD Athlon 64 memory controller, this low latency infrastructure translates into very high overall system performance.  The nForce 4 also has dedicated silicon to help balance out the SLI needs in terms of synchronization and data delivery.

            The third and final area that NVIDIA addressed was that of system and OS level technology.  Specialized drivers are enabled when SLI is detected, and these drivers help to orchestrate the entire process.  These affect not only the applications ability to render to SLI, but also control the silicon in the nForce 4 and 6x00 GPUs.  While NVIDIA didn’t comment on the overhead that these drivers consumed, they did say that they were very efficient and shouldn’t bog down the processor.  More work is being done on them, and they expect better performance in the future with later driver revisions.


Next: Why We Need Certification


If you have found this article interesting or a great help, please donate to this site.


Copyright 1999-2004 PenStar Systems, LLC.