NVIDIA's Multi-GPU Technology - SLI, it's baaack!

Is it really SLI?

Let's take a walk down memory lane, shall we? Where were you in 1999? If you were into the technology scene, especially personal computing, most likely you were scampering about like Chicken Little, wondering what you would do when the sky fell at the end of the year, as the clocks ticked over to the year 2000. The Y2K support and compliance industry was in full bloom, as we all thought our computers, mainframes and networks would somehow spontaneously combust due to the fact that our BIOS firmware, operating systems and application software couldn't handle the roll-over into the new millennium.

In addition to worrying about your soon too be possessed PC, if you were an enthusiast, you probably also were happily gaming on a pair of Voodoo 2 cards, in a Scan Line Interleaving setup, or "SLI" for short. Well, SLI is back again today, sort of. When NVIDIA bought out 3dfx in December of 2000, they also bought the intellectual property, patents and branding of the company that was once a driving force in 3D Graphics for the PC.

NVIDIA's New Breed Of SLI

Not exactly SLI but maybe better?

3dfx's SLI technology,used two graphics processors in a parallel processing setup where one chip drew even lines of an image and the other chip drew the odd lines of the image. This in many ways was a simple way of increasing available fill rate, geometry setup and load-balancing of rendering workloads on the graphics subsystem. And because it was working at the scan line level, it was fairly easy to manage timing of rendering and output to a display, so artifacts were non existent.

NVIDIA's new breed of SLI however, isn't really SLI at all. The company is just retaining the name and branding of the technology, since many folks will identify with the parallel processing concept that was made so popular by their one time competitor and now assimilated symbiotic partner, 3dfx.

CLICK ANY IMAGE FOR AN ENLARGED VIEW

Taking a page from the book of Wicked3D/Metabyte (remember them?), NVIDIA is actually employing a lesser known technology for their Multi-GPU NV45 implementations, more a kin to Metabyte's "PGC" or Parallel Graphics Configuration approach. In this implementation,both chips work on different sections of a frame. One chip renders the top half of the screen, while the other chip renders on the bottom half of the screen. This approach has the potential for rendering workloads to be unbalanced between the two GPUs, however. One half of the screen can have fewer polygons to render than the other. For example, in an outdoor environment, the top half of the screen can have much less detail to render in the sky area, versus the bottom ground half of the screen.

As you can see in the above nature scene from 3DMark03, the top half of the frame has some detailed rendering to do but the bottom half may have significantly more workload involved due to hundreds of individual blades of grass that are all set in motion. This is where NVIDIA's new software algorithms will come into play. The host CPU and chipset will manage the work load distribution evenly between each GPU. In addition, NVIDIA's NV45 architecture also incorporates an inter-GPU communications protocol that allows the two cores to signal status between them, for more efficient multi-chip rendering results.

Boards and Configurations

So, it looks like NVIDIA has been a bit more focused on "out of the box" thinking as of late, since obviously this new Multi-GPU parallel rendering technique would be nearly impossible without the availability of the new PCI Express interface for graphics. With its serial point to point communications protocol, dedicated links back to system resources and gobs of available managed bandwidth, PCI Express is a natural for these types of applications.

CLICK ANY IMAGE FOR AN ENLARGED VIEW

Obviously, there's a very different motherboard technology that is key to the architecture. Dual X16 PCI Express Graphics slots will also be required for this type of configuration. Above you'll see a motherboard from SuperMicro based on what we think is Intel's forthcoming E7515 "Tumwater" chipset for Intel's new "Nacona" Xeon workstation processor that also has 64-bit extensions ready to rumble. We did not get conformation on this from NVIDIA but the timing of their announcement, along with Intel's today, suggests this.

The question is, will we see other chipset and motherboard permutations moving forward, that have the all-important dual PCI Express X16 graphics slots? You could also conceive that an i925X solution with an add on discrete X16 switch, could be a possibility. Consumer class singe-CPU Tumwater boards could also be released. And NVIDIA's next itineration of the nForce will almost definitely support multiple PCI Express X16 slots. However again, until we get more firm details, we're simply speculating here. We'll come back and update this section as more information come forth from either NVIDIA or Intel.

Closer inspection of the a GeForce 6800 Ultra In SLI

So it is really SLI!

CLICK ANY IMAGE FOR AN ENLARGED VIEW

Above you can see NVIDIA's new board to board connector for the PCI Express/NV45 versions of their GeForce 6800. This is a simple rigid PCB with a female slot connector that mates on to the top edge connector of the graphics card. NVIDIA informed us that SLI ready PCB designs will only be made available on higher end products like the GeForce 6800 GT and 6800 Ultra models. Here's where NVIDIA is actually coining the acronym "SLI" again, which stands for Scalable Link Interface.

Historically, 3dfx cards used ribbon cables to connect master and slave cards together and pass through cables to send analog output from one card to the other. However now with SLI, all data is transferred through this connector via the MIO (multi-function IO) port, a block of circuitry on the GPU itself that provides a communications link and chip level assist for load balancing. It's a fully digital implementation, versus legacy SLI which performs a digital to analog conversion. This MIO port, in combination with NVIDIA load balancing algorithms, will supposedly allow perfect image fidelity, along with a claimed 1.87X average speed improvement. We'll be testing NVIDIA's new SLI in our labs in the coming weeks so stay tuned and we'll see if this claim rings true.

The dual card nature of and SLI configuration will also give users the ability to run a quad-monitor setup in "2D mode". NVIDIA tells us there will be a switch in the drivers that will allow the SLI functionality to be disabled, which makes each card function independently of the other.

Conclusion

So the revival of SLI and another tidbit of 3dfx nostalgia is brought back to market by NVIDIA today. It again lends credence to NVIDIA's nimble move of acquiring the IP and other assets of 3dfx. The company is obviously still utilizing the technologies of their famed acquisition, while putting a refreshed leading edge spin on it, with new available technologies like PCI Express. Obviously, NVIDIA had this capability planned for a long while now, with the inclusion of the MIO port in their new line of NV4X GPUs. It seems we are learning more and more what all those 222 million transistors are doing in the NV40 and 45, beyond just its 16 parallel processing pixel pipelines. Who knew when the company claimed their architecture was "highly scalable", that it was scalable at the chip level, as well as board to board.

For the crew here at HotHardware, we've got a twinge of that special feeling we had back in the days of 3dfx, when 3D Graphics were still being pioneered and you felt like something totally innovative was on the horizon. Perhaps this will be a nicheproduct and configuration, with only a small segment of the end user market having the bankroll to shell out over a thousand greenbacks for a pair of Geforce 6800 Ultras in SLI. Not to mention the new motherboard, CPU and most likely the new power supply needed to support all this new hardware. On the other hand consider a pair of GeForce 6800 GT cards, in single slot designs, at a sub $400 price point. If you've already made the move to Intel's new PCI Express driven platform, this could be a significantly more manageable proposition, especially as street prices begin to fall in the coming months. Most likely you'll still need to sell it a bit to the significant other, however!

The key is going to be the support hardware coming to market in the next few months. Motherboard and chipset configurations will help drive availability of this technology into the mainstream. Full systems based on NVIDIA SLI multi-GPU technology are expected to become available in the second half of 2004 from multiple PC and workstation manufacturers including:

Alienware – North America, Europe
Atelco Computer – Germany
Boxx Technologies – North America
Falcon Northwest – North America
Mouse Computer Japan (MCJ Co., LTD) – Japan
Network Technical – Sweden
Paradigit - Holland
Scan Computers - UK
TSUKUMO Co., Ltd. - Japan
UNITCOM INC - Japan
Velocity Micro – North America
ThirdWave Corporation - Japan
VoodooPC – North America

In addition, multiple add-in-card manufacturers will also be adopting the technology. These manufactures include:

Albatron Technology, Co. Ltd
AOpen
ASUS Computer International
BFG Technologies, Inc.
Chaintech
ELSA Japan
Gainward Co., Ltd.
Gigabyte Technology, Co., Ltd.
Leadtek Research, Inc.
Micro-Star International
Palit Microsystems, Inc.
Pine, XFX, a Division of Pine
PNY Technologies
Prolink Computer Inc
Sparkle

We'll be sure to come back to you with coverage in the months ahead. NVIDIA plans to have boards available around the August time frame, so stay tuned for our full review.