With dual launches today we?ve decided  to kill two birds with one stone and combine coverage.  Thus, you get a double-launch treat.  Hopefully it ends up tasting like peanut-butter and chocolate instead of, say, pickles and cherries.

I?m at a loss for my usual stunning wit and genius humor, so I?ll just get straight into the coverage.  First up: NVIDIA, Intel to follow.

NVIDIA?s SLI Technology:

We?ve been hearing tidbits and teasers from Alienware for weeks about NVIDIA?s new dual PCI-Express technology; today we finally get to take a look at what?s behind it.

Although NVIDIA is calling it SLI (originally short for Scan-Line Interleave), that display mode is only one of the various methods this new solution can use. Time for a bit of history on graphics, and why this launch is both new and old.

 

 

 For 250K to 1 million these are some pretty sweeeeet graphics.  If you squint you can actually see a texture map.  Forget trying to explain to my kids how I played games that looked like this--I can’t even remember how I played games that looked like this--and I played them!

As the above slide points out, parallel video displays have been with us for awhile.  Its not hard to see why, especially back in the days of very slow video?if you wanted to display the screen faster (or render data faster), one of the simplest ways is to break the screen up into pieces and render each piece. 

The tricky part, of course, is making sure that every piece goes to the correct video hardware, is rendered properly, and then put back together perfectly every time.  End users won?t care if quadrants 3 and 4 only switch sides ?every few hours?.

There?s a lot of different ways to potentially render a screen in parallel.  You can split it by pixels, by lines, by sections or quadrants, or by load.  The trick is to ensure that the software or drivers handling the task are up to it.  Fortunately with modern hardware, this isn?t a problem.

The parallel method 3dfx released on the desktop wasn?t new, it was simply a new phenomenon for consumer markets.  Because there were no hardware T&L engines in the days of the Voodoo2 only triangle setup was done in parallel, geometry data was, of course, still handled by the CPU.  Scan Line Interleaving (SLI) was a method of sending only every other line to a particular video card.  The cards were connected by a pass-through cable and the resulting blended image was displayed.  At $300 per card this wasn?t exactly cheap, but it did give a substantial boost to performance.

What NVIDIA is doing today isn?t *really* 3dfx SLI, though that?s one mode available to the card.  They played their cards close to their chest in terms of what rendering modes are available and what the various modes actually do, but I got the distinct impression that the card can either be forced into a particular mode or will dynamically choose what rendering most is best based on a variety of factors it detects.  If, for example, one video card is dealing with a tremendously difficult part of one scene, the majority of the screen may be allocated to the second card, in order to dynamically load-balance the rendering process.

Physically, the system is nearly identical to the old 3dfx hookup methods.  The two cards are connected via an internal cable.  No word yet on if an external connector is also used, but we?re guessing not.  There?s also no reason we?re aware of why this system couldn?t be used to power four external displays.  Fire it up with an integrated video system (if the combination is ever produced) and you?ve got a six-display-capable system using only two video cards.  Not bad.

Technical Details of NVIDIA?s SLI:

Unfortunately these are few and far between, but we?ll tell you what we know.

• Texture Data duplicated across both video cards:  Ideally, an SLI configuration would be able to send each video card only the data needed to render its portion of the screen, thus giving the entire video array double the RAM on any given card (in this case, 512 meg total). 

One constraint that?s existed since the 3dfx days, however, is that its extremely difficult to properly route texture data intelligently and the cost of failing to do so (ie, a ?miss?) is severe.  The video card must then fetch the proper data, which could result in a performance hit, or worse, a screen mis-match. 

Those of you who?ve been around awhile, however, will remember that one of the dis-advantages to 3dfx?s VSA-100 design was that while each chip required its own discrete RAM, the total did not equal the sum of the parts.  The Voodoo5 5500 had 64 meg of RAM, but performed like a 32 meg card, as did the Voodoo5 6000.  NVIDIA?s dual solution will have a similar limitation.

In all fairness, however, it shouldn?t matter.  We?ve yet to see games that strain the RAM capacity of a 128 meg video card, much less a 256 meg version.  Just keep in mind that the duplication of texture data effectively means you have 2x the cores and 2x the bandwidth, but still only 256 meg of ?effective? video RAM.  Trust me?you won?t be complaining.

• You?ll Need Power?But not TOO Much:  Initially NVIDIA specced their new Ultra?s as requiring a 480W PSU.  NVIDIA has since stated that their initial estimate was made using extremely conservative measurements, and now recommends a 350W PSU for using a 6800-class card.  Obviously if you?re packing six hard drives, 4 CD-ROMs, and two CPU?s already that claim may not apply, but for the rest of us its welcome news.

We?re guessing that this dual solution will require 500-600W of power; an expensive-but-do-able amount far below the 800-1 Kw demands that were initially discussed.  Big numbers may be fun, but I?m in no hurry for my power supply to require its own heatsinks, fans, and separate casing.

• Will be Offered in Future Athlon 64 Chipsets:  Although NVIDIA wasn?t exactly jumping up and down to offer details, we know that future Athlon 64 chipsets will offer dual PCI-E support.  When and which isn?t clear; presently support is only available on Intel?s Tumwater.  
   
• Will be Present on All 6800 PCI-Express cards:  We don?t know yet about the 6800 standard, but SLI will be possible using both the 6800 GT and the 6800 Ultra (PCI-Express versions only, obviously).
   
• Available for both Quadro and GF FX Cards:  NVIDIA isn?t going to confine SLI to just the desktop market; Quadro PCI-Express cards will be SLI-capable as well. 

 

next >

---

• NVIDIA Page 2. (1 of 2)
• Intel's Nocona: 800 MHz FSB's and x86-64 Support. (2 of 2)