Slashdot Mirror
Voodoo5 6000 Preview
Robert writes: "VoodooExtreme posted a hands-on preview of the mysterious 3dfx Voodoo5 6000 video card from the I/ITSEC (http://iitsec.org) Interservice/Industry Training, Simulation and Education Conference which comes equipped with the first-ever published benchmarks of Quake III. Formerly the Voodoo5 6000 was planned to be a retail product, but its high price tag and need for an external power supply led them to sell the rights to it to Quantum 3D, where it will be used in Visual Simulation industry..." I always thought a video card was a selling your video short ;)
Let's be honest here, when was the last time 3dfx had a leading edge product?
Was it the Voodoo5? Bzzt, no, sorry.
The Geforce2 GTS was released first and blew the 5500 out of the water in terms of frame rate. But that was OK, the Voodoo5 had FSAA and we'd pay for that. Then NVIDIA "leaked" drivers for the GF2 that had FSAA. No problem for 3dfx though, they'd produce a 4-chip Voodoo5 6000 and retake the frame rate crown. After showing the product at e3, they proceed to figure out that nobody will buy a card that's too damned big to fit in most cases.
How about the Voodoo3? That was pretty neat right? Unless you like 32 bit color, then you're screwed since 3dfx didn't think that competing with the features in the Rage128 and GF256 was worthwhile. The shareholders must have loved that.
How about the Banshee? What a deal! Inadequate 2d combined with the aging Voodoo2 chip but without SLI!! What were they thinking?
So yeah, 3dfx hasn't had a decent product since the Voodoo2. That's one fucked up company.
--Shoeboy
I always thought a video card was a selling your video short
Understand jibberish not. Dotslash someone would please editor hire for.
"It proves what most people suspected, the card was being sold to the public because even after nearly 6+ months of work they can't make it work."
Although I don't have confirmation of this, I believe the V5-6000 ran into a problem with certain motherboard BIOSes.
In the V5-5500, the two VSA-100 chips on board are both directly on the AGP bus. However, with 4 VSA-100s, the V5-6000 had to have a bridge chip in between the bus and the graphic chips. Bridge chips are actually pretty common with PCI devices.
However, certain motherboard BIOSes refused to recognize the bridge chip as an AGP device. The problem could have been fixed with a new BIOS flash, but unfortunately there were enough motherboards with the problem that they couldn't release a retail product that way.
Though at first, I thought the sparc-like parallel CPU archetecture was kind of cool. But it has some serious flaws. First, there are some serious paracitic forces that impeed parallel operations.. Next it's like the NASA way of doing things.. Cost is irrelevant, we want 2, 3, or 5 way redundancy. Even if volume brings the chip cost way down, you still have to duplicate memory and controller connects, etc.
Next as far as I've been able to tell, taking a pseudo-multi-threaded application and throwing more [CG]PUs at it very quickly dies off. 64 processor SPARC machines work well mainly because they multi-task, not multi-thread. Rendering a single graphic scene is not multi-taskable, nor even multi-threadable. At best what you get is a heavy and independant pipelines coupled with SIMD operations.
As it turns out, The VSA requires a single-tasked/ single threaded stage in the pipeline that is a bottleneck for all chips. Namely the dividing and sorting of vertexes before distribution to the individual plane processors. Though a novel and "scalable" approach, this provides such over-head on low-end cards, that it can't compete with traditional archetectures. In fact, the more polygons a scene has, the slower the VSA architecture will go. This series is best suited for average numbers of polygons with incredible texture dependancies - which utilizes the BW and the scalability of the chips. Unfortunately, even for CAD/Graphics designs which Quantum seems to think they can sell to, you're going to need massive polygon counts. And unless their distribution / sorting stage scales well with the number of processors, this isn't going to be good for them.
This information I've deduced from the various pages on sharkyextreme, anandtech, and tomshardware. Specificly in relation to the ATI Radeon / nVida Detonator 3 drivers verses the infinite plains approach VSA takes.
-Michael
-Michael