nv4

[Arkady Ostrovski]

nVidia Announces RivaTNT

by Jonathan Greenberg
March 23, 1998

Earlier this month, OGR reported that nVidia is going to be releasing a new 3D
accelerator chip, the Riva128ZX, which should improve 3D performance over the
current Riva128 by about 1.5 times. But, this chip is mainly going to be a
Spring product for nVidia, something fresh to have on the market until Fall,
when they'll be ready to release the Riva TNT. The Riva TNT, also known as the
Riva 4 internally at nVidia, should truly be one powerhouse of a 3D chipset.
Why?
First off, the TNT will be fast. Really fast. This card should easily outrun a
Voodoo2, and may even have a serious shot at killing a Voodoo2-SLI equipped
system. Of course, that probably comes across as a wild claim, and I can just
imagine a lot of the Voodoo2 equipped readers out there immediately cursing my
name and shouting expletives. But no, the TNT should be able to meet this claim.
How can the RivaTNT outperform the Voodoo2? Well, let's get into a bit of the
under-the-hood stuff, shall we?

For nVidia, the TNT is a generational leap. The TNT isn't simply a reworking of
the Riva128. Instead, it's a whole new architectural design from the ground up,
investing in some proven ideas that carried over from the '128, but also
introducing a bunch of new tech into the fray. The TNT has quite a few snazzy
internal features, the first of which is their Perfect-Prefetch technology. What
Perfect-Prefetch does is when the card is sent a polygon with texture
coordinates, instead of just reading the texture from on-chip memory, it looks
at the portion of memory that the polygon specifically requires. First, a
stencil of the polygon and its mipmap is brought into memory. It looks at the
size of the polygon in the framebuffer, and determines the scale that the
texture will require. Once this is accomplished, only the required sub-block of
memory for the texture is actually brought in. Reducing the amount of texture
memory that needs to be transferred reduces bandwidth, and allows the card to
just fly along.

In addition to the Perfect-Prefetch, which is important enough that nVidia is
patenting it, the chip has two texel engines, allowing it to crunch through two
pixels per cycle. Not only does this allow for fast rendering, but the pixel
engines operate on 32 bit RGBA colors, so the image quality coming out of cards
equipped with the chip should be quite high as well. All effects, such as fog,
mipmapping, alpha, etc., are all per-pixel, so the appearance of these effects
won't be compromised in anyway either. The chip will also feature a 24 bit
ZBuffer, to reduce those ugly popping effects that sometimes arise from the
lower precision of a 16 bit ZBuffer most current cards have, and an 8 bit
Stencil buffer, to speed up and simplify shadow and mirroring effects.

Of course, the huge benefit of having TwiN Texel engines (which is where the
TNT acronym comes from, by the way) is the ability to support multitexturing and
all of the new bonuses it provides, including the ability to do single pass
lightmaps, bumpmapping and whatever else coders can think of. All computations
on the textures for blending are done at full 32b color as well, so feel safe in
knowing that texture quality won't degrade when multitexturing is enabled.

The chip also features a full triangle setup engine, and one that is
tremendously improved over the one found in the Riva128, allowing a total
maximum throughput of something in the range of 8 megatriangles per second. Of
course, that limit is entirely defined by how fast your CPU will be able to
deliver polygons to the card, something that a lot of Voodoo2 owners are
starting to realize. To get full performance out of the Riva TNT you'll need one
heck of a CPU behind it. Much like with the Voodoo2, don't expect to come even
close to seeing the full potential of this chip unless you've got something in
the class of a high-end Pentium II, or a chip with AMD-3D or MMX2 on board. The
faster you can get polygons to the chip, the higher the TNT's rendering will be.

As to full performance... Well, I've implied it's fast, but I haven't given you
any numbers yet, and I'm sure you're fairly hungry for them. Well, since the
chip is still in the development stage, numbers are iffy, and nVidia is trying
to stay conservative with their numbers so as not to incite unrealistic
expectations. But here are some basic numbers, nonetheless. First, the absolute
maximum pixel throughput of the card is 250 megapixels per second, although most
likely you won't see that kind of performance anytime soon. By the way, that
number is arrived at simply by multiplying the clock speed of the chip (125 MHz)
by the 2 texels per cycle number. A more conservative estimate for average game
performance (say on something of the Quake2 variety) would be around 150Mps.
Some games should be able to reach sustained rates as high as 200Mps (more than
2 Voodoo2's in SLI), although quite frankly if you turn on every feature, expect
the lower number to be more realistic. The TNT is a total 2D/3D solution as well,
so nVidia doesn't exactly see themselves as directly competing with the
Voodoo2's market anyway.

Additionally, the chip will support the full DirectX 6 feature set, and will
include a full OpenGL ICD, making it ideal for both gaming, and development
systems. Additional features include support for both PCI and AGP 2x, up to 16MB
of framebuffer memory, and a 250 MHz RAMDAC with support for resolutions up to
1600x1200x32b @ 85 Hz, enough to satisfy almost any high-end developer.

Expect the Riva TNT to be supported by a wide variety of OEMs, much like the
Riva128, and to available in time for the back-to-school season (or as I like to
call it
"The-beginning-of-falling-behind-in-your-studies-as-you-play-every-game-that-co
mes-out season"). This chip is certainly one to watch out for. Watch the OGR's
Technology Talk section for more information on the Riva TNT as it nears
release.

Arky
mailto: ark...@corbina.ru
http://www.corbina.ru/~arkady