Why so few geometry chips?
Phil Hindle
Hi,
There are tonnes of 3d cards coming onto the PC market, which generally
tend to accelerate various rasterisation functions, and apply effects to
polys etc.
But why do none of them perform geometry? PlayStation has the GTE, and
sticking a few gates into a chip, to do quick matrix operations must be
really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
has, are more difficult and costly to implement than a quick matrix
multiply function?
Or am I missing something obvious?
--
Cheers,
- PHIL
******************************
Codemasters web site :
www.codemasters.com
******************************
Stefan Boberg
"Phil Hindle" <ph...@codemasters.com> wrote:
>There are tonnes of 3d cards coming onto the PC market, which generally
>tend to accelerate various rasterisation functions, and apply effects to
>polys etc.
>But why do none of them perform geometry? PlayStation has the GTE, and
>sticking a few gates into a chip, to do quick matrix operations must be
>really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
>has, are more difficult and costly to implement than a quick matrix
>multiply function?
Well I'm no hardware designer, but as far as I know, it takes TONS
of gates to make a fast floating-point unit with the kind of
throughput that would be necessary to get worthwhile acceleration.
--
=====================================================================
Stefan Boberg calm / fitter, healthier and more productive
bob...@team17.com a pig / in a cage / on antibiotics
Frants P
bob...@team17.com (Stefan Boberg) writes:
>"Phil Hindle" <ph...@codemasters.com> wrote:
>>But why do none of them perform geometry? PlayStation has the GTE, and
>>sticking a few gates into a chip, to do quick matrix operations must be
>>really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
>>has, are more difficult and costly to implement than a quick matrix
>>multiply function?
> Well I'm no hardware designer, but as far as I know, it takes TONS
>of gates to make a fast floating-point unit with the kind of
>throughput that would be necessary to get worthwhile acceleration.
Well, Tritech must have done it with the Pyramid3D:
As far as I know there are no boards out yet, but it's supposed to be
totally awesome!
Patrick Frants
--
/***************************************************************\
Name: Patrick Frants E-Mail: pfr...@cs.vu.nl
Phil Carlisle - Team 17 Internet Programmer
On 7 Jul 1997 12:37:56 GMT, "Phil Hindle" <ph...@codemasters.com>
wrote:
>Hi,
>
>There are tonnes of 3d cards coming onto the PC market, which generally
>tend to accelerate various rasterisation functions, and apply effects to
>polys etc.
>But why do none of them perform geometry? PlayStation has the GTE, and
>sticking a few gates into a chip, to do quick matrix operations must be
>really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
>has, are more difficult and costly to implement than a quick matrix
>multiply function?
>
>Or am I missing something obvious?
Well, a lot of the better 3D chip manufacturers also have geometry
chip units that can be integrated with the original chips, I guess its
mainly down to cost, look at the 3dfx based cards, essentially the
3dfx is a chip that can be combined with geometry chips (ala obsidian
boards), but to keep the costs down, they release the rendering chip
first. Presumably once the main chips come down in costs (as volumes
increase), they will start including the geometry chips and call it
version 2!!
If you look at the pyramid site, is also shows the same path,
initially you get a render only setup, then later they release a
render+geometry setup.
If you think about it, it makes sense, given that PC's are capable of
doing geometry transforms quite nicely... what else would you use the
main PC chip for??? admittedly, it would be nice to be able to do
geometry setup in hardware, 3d sound in hardware, then all the main
processor does is hold everything together, network stuff, and update
1million objects :))))
Hey, you never know, maybe we'll see some accelerated network cards
coming out with 8mb of network buffer :)))
Phil.
rebel
>
>
> bob...@team17.com (Stefan Boberg) writes:
> >"Phil Hindle" <ph...@codemasters.com> wrote:
> >>But why do none of them perform geometry? PlayStation has the GTE, and
> >>sticking a few gates into a chip, to do quick matrix operations must be
> >>really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
> >>has, are more difficult and costly to implement than a quick matrix
> >>multiply function?
>
> > Well I'm no hardware designer, but as far as I know, it takes TONS
> >of gates to make a fast floating-point unit with the kind of
> >throughput that would be necessary to get worthwhile acceleration.
>
> Well, Tritech must have done it with the Pyramid3D:
>
> http://www.tritechmicro.com/
>
> As far as I know there are no boards out yet, but it's supposed to be
> totally awesome!
>
"Supposed to be" meaning? From what I saw of the demonstration it wasnt too
impressive.
Mike Young
Phil Hindle wrote:
>
be
> really easy & cheap, right? Surely all the fancy effects that a 3dfx say,
> has, are more difficult and costly to implement than a quick matrix
> multiply function?
>
=====
The obvious something is (probably) the lack of a DirectMultiply HAL.
Even the knuckleheads in the room voted against this one, I imagine.
Mike.
Paul Shirley
In article <33c1459b....@news.team17.com>, Phil Carlisle - Team 17
Internet Programmer <p...@team17.com> writes
>Well, a lot of the better 3D chip manufacturers also have geometry
>chip units that can be integrated with the original chips, I guess its
>mainly down to cost, look at the 3dfx based cards, essentially the
>3dfx is a chip that can be combined with geometry chips (ala obsidian
>boards), but to keep the costs down, they release the rendering chip
>first. Presumably once the main chips come down in costs (as volumes
>increase), they will start including the geometry chips and call it
>version 2!!
The primary difference between Obsidian and consumer 3dfx boards is an
extra TexelFX. This allows true trilinear mipmapping (you have to settle
for dithered trilinear - nearly as good but still shows artifacts), 1
pass modulated texturing (which would help Quake run up to 2* faster)
and gives higher quality lighting, and you have twice as much texture
memory!
I've always wondered how D3D handles chained texture sources. I supsect
it simply cannot, so no consumer card will ever have multiple TexelFX
processors.
---
Visit www.dukepsx.com: see what I do all day.
Paul Shirley: my email address is 'obvious'ly anti-spammed
Michael I. Gold
"Phil Hindle" <ph...@codemasters.com> writes:
| There are tonnes of 3d cards coming onto the PC market, which generally
| tend to accelerate various rasterisation functions, and apply effects to
| polys etc.
| But why do none of them perform geometry?
Most PC cards are targetting the games market. For most games, the
bottleneck is going to be fill rate, rather then polygon throughput.
Pentium has pretty decent fp performance, so geometry is often not the
bottleneck. Adding a faster geometry processor will price you out of
the market, until geometry becomes a significant bottleneck and you
can command a higher price for your added value.
Geometry processors will be more interesting if/when CAD on PC's takes
off in a big way. This is still a stronghold for workstations, and
will remain so for the forseeable future. (My opinion, okay? Lets
not start a religious war on that too. I'm not saying that PC's
aren't used for CAD, just that serious CAD users are still buying
workstations.)
--
Michael I. Gold Silicon Graphics Inc. http://reality.sgi.com/gold
And my mama cried, "Nanook a no no! Don't be a naughty eskimo! Save your
money, don't go to the show!" Well I turned around and I said, "Ho! Ho!"
Phil Hindle
Frants P <pfr...@cs.vu.nl> wrote in article <5pqu3n$o...@star.cs.vu.nl>...
> bob...@team17.com (Stefan Boberg) writes:
> >"Phil Hindle" <ph...@codemasters.com> wrote:
> >>But why do none of them perform geometry? PlayStation has the GTE, and
> >>sticking a few gates into a chip, to do quick matrix operations must be
> >>really easy & cheap, right? Surely all the fancy effects that a 3dfx
say,
> >>has, are more difficult and costly to implement than a quick matrix
> >>multiply function?
>
> > Well I'm no hardware designer, but as far as I know, it takes TONS
> >of gates to make a fast floating-point unit with the kind of
> >throughput that would be necessary to get worthwhile acceleration.
>
> Well, Tritech must have done it with the Pyramid3D:
>
> http://www.tritechmicro.com/
>
We were given a demo of the TriTech chip at work recently, by one of their
sales people, and the pixels it knocks out are fantastic! The bump mapping,
and radiosity shading are excellent. My only doubt was the pixel fill rate
seemed similar to 3dfx, but this was a pre-production chip (or so the
tritech guy said)
And with regard to the matrix chip, a consumer level IC (i.e. games
dedicated, a la 3dfx), would only have to handle 32-fixed point values
(using 64-bit intermediate values). Much easier than all that nasty
floating point stuff :-)