MaximumPC 2006 12

T

he industry has known for months that

Nvidia would be the first GPU manu-

facturer to introduce a Direct3D 10 part, but

absolutely no one anticipated the 681-mil-

lion-transistor behemoth now lumbering out

of the TSMC fabs in Taiwan.

Here’s another fact about the 8800 that

no one anticipated: It’s based on a uni-

fied shader architecture. Looking back

on everything that Nvidia executives

have said—for attribution, off the record,

or in guidance they reportedly gave

to third-party, supporting-component

manufacturers—we’ve come to the con-

clusion that either the company made a

brilliant course correction in the middle

of its product-development process, or it

waged an even more brilliant and years-

long campaign sowing FUD.

Nvidia chief scientist David Kirk, for

instance, has been quoted as saying

“there’s plenty of mileage left in [discrete

shader-unit] architecture,” that “change

will probably happen progressively,”

and that Nvidia would release a uni-

fied-shader part “when it makes sense.”

ATI, meanwhile, has openly informed the

press that its next part would definitely

feature a unified shader architecture, one

adopted from the Xenos chip they devel-

oped for the Xbox.

Just weeks before Nvidia’s announce-

ment, power-supply vendors had been

whispering about upcoming videocards

that would

require 300 watts

of power, Nvidia

reps quizzed

journalists about

their thoughts

on water-cool-

ing, and “leaked”

photos of a G80

prototype card

outfitted with

what looked like

hose connec-

tions popped up

on the web.

Here’s the

reality: The

ultra high-end

GeForce 8800

GTX ($600) has

128 shader units (Nvidia calls them

“stream processors”) running in paral-

lel at 1.35GHz (the core itself runs at a

more laid-back 575MHz), and each one is

capable of executing geometry, vertex, or

shader code. In other words, it’s a unified

architecture. The GPU has a 384-bit inter-

face to 768MB of GDDR3 memory running

at 900MHz.

A reference-design GeForce 8800

GTX card consumes just 185 watts (the

GPU has twice as many transistors

as the 7950 GT, so it must be incred-

ibly power efficient), and the GPU and

memory are air-cooled (by a Zalman-

designed cooler very similar to the one

on the 7900 GTX). Nvidia claims that a

single 8800 GTX is faster than two 7900

GTX cards running in SLI. Nvidia also

announced a slightly less-expensive

GeForce 8800 GTS that’s outfitted with

96 shader units running at 1.2GHz

(the core is clocked at 500MHz). Cards

using this chip will be equipped with a

320-bit interface to 640MB of memory

running at 800MHz; Nvidia expects them

sell for to $450.

Nvidia’s announcement came just as

we were going to press, but you know

we’ll have a more in-depth report—with

coverage of SLI, Nvidia’s new antialias-

ing technologies, hands-on benchmark

results, and a whole lot more—in our next

issue. Stay tuned.

Preview

Nvidia GeForce 8800

Nvidia has unveiled the first Direct3D 10 accelerator. Here’s everything we know about it, so far.

r&dBreaking down tech­­­­—present and future

66 MAXIMUMPC december 2006

With a unified shader architecture and relatively meager

power requirements, Nvidia’s GeForce 8800 GTX surprised and

delighted us!

by michael broWN

Nvidia NForce 680i SLi

S

o you finally got a mobo using Nvidia’s

590 SLI Intel Edition chipset? Well

prepare to get pissed, because Nvidia is

releasing a new chipset that supercedes

the late-to-the-party 590 SLI Intel part.

Dubbed nForce 680i SLI, the top-end

chipset uses essentially the same south

bridge as the 590 SLI Intel Edition chipset

but adds a redesigned north-bridge chip.

Nvidia has yet to provide a satisfactory

explanation for why the 590 Intel Edition

arrived so late, but sources outside the

company have hinted that poor over-

clocking performance was the problem.

Apparently the 590 SLI IE didn’t have the

overclocking chops of other Intel chipsets.

The 680i should address that in spades.

The chipset supports dual x16 PCI-E con-

nections plus a third x16 physical PCI-E

slot with x8 electrical plumbing for GPU

physics support, packet prioritization,

gigabit teaming, massive RAID structures,

and a ton of overclocking features.

In a quick test of an early 680i-based

motherboard from eVGA, we ran a

1.83GHz Core 2 Duo E6300 up to 2.7GHz

with the FSB clocked up to 390MHz,

or 1.5GHz, without it breaking a sweat.

That’s 500MHz out of spec on the FSB.

With further tinkering and more voltage,

you should be able to take it even further;

and, hell, you get SLI to boot. —GMU