S3 Savage 4 - number of pipelines

N

no-X

Veteran
I have a question about pipelines/texture units configuration in S3 Savage 4 core. Many websites tells, that Savage 4 is 1x2 (like Voodoo3) and very small number of sources rate S4 as 2x1 (like TNT2). I tried some fillrate benchmarks:

Savage 4 Plus (110*125/64bit):

3D Mark 1999 default:
single-texturing: 104,2
multi-texturing: 107,5

3D Mark 2000 default:
single-texturing: 101,0
multi-texturing: 109,1

3DMark 2001SE default:
single-texturing: 45,6
multi-texturing: 76,0


TNT2 M64 2x1 (125*150/64bit):

3D Mark 1999 default:
single-texturing: 150,1
multi-texturing: 192,3

3D Mark 2000 default:
single-texturing: 146,7
multi-texturing: 201,7

3DMark 2001SE default:
single-texturing: 114,8
multi-texturing: 136,4

Voodoo 3 - 1000 1x2 (125*125/128bit)

3D Mark 1999 default:
single-texturing: 118,4
multi-texturing: 230,5

3D Mark 2000 default:
single-texturing: 121,8
multi-texturing: 239,5

3DMark 2001SE default:
single-texturing: 150,3
multi-texturing: 267,5
------

Results from 3DMark 01 are a bit strange, so numbers from 99 and 00 seems to be more relevant. Savage 4 appears to be 2x1 core, but can anybody confirm this?

Thanks
 
IIRC, Savage4 was actually 1x1. It could do single-cycle Trilinear, which meant that the performance difference to TNT2 etc in actual games was usually much smaller than the raw fillrate number indicated.
 
Don't forget that 3DMark fillrate tests use alpha blending, so they're not very good at seeing raw fillrate due to bandwidth restrictions. You should use one of the other various tools out there, assuming they work with Savage4 and whatever version of directx they have drivers for.
 
Mintmaster said:
Don't forget that 3DMark fillrate tests use alpha blending, so they're not very good at seeing raw fillrate due to bandwidth restrictions. You should use one of the other various tools out there, assuming they work with Savage4 and whatever version of directx they have drivers for.
Click to expand...
I don't think this applies to Savage4; searching for benchmark results on it indicates that it was one of the cards where the difference between 16 and 32 bit performance was smallest (~5-10%). Also, the fact that Savage4 supports multitexturing (note: "single-pass", not "single-cycle") while at the same time having the same fillrate in single-texturing and multi-texturing modes suggests that memory bandwidth was almost never what held the Savage4 back.
 
arjan de lumens said:
IIRC, Savage4 was actually 1x1. It could do single-cycle Trilinear, which meant that the performance difference to TNT2 etc in actual games was usually much smaller than the raw fillrate number indicated.
Click to expand...

Interesting. So, both (Savage 3D and Savage 4) are 1x1, both are clocked similarly, both use 64bit memory bus, but Savage 4 is much faster. Whats the cause of this performance difference?
 
By my knowledge only LT and GT used 64 bit memory interface. Savage 4 packed lots of extras in its large core, one of those more evident was about 40% faster triangle setup over first Savage.
 
Last edited by a moderator:
Could be many things; for e.g. Quake3, which exhibits very large differences, I would suspect that the 8MB onboard memory of Savage3D simply isn't enough to avoid AGP texturing, while the 32MB of Savage4 is.
 
arjan de lumens: I'm confused, because somebody told me, that according to OpenGL extensions for Savage 4, it can't be 1x1 configuration...
 
AFAICT, the Savage4 appears to be capable of "single-pass" but not "single-cycle" multitexturing, so that if you feed it a polygon with 2 textures it will spend 2 cycles per pixel, using the same physical texture unit twice to apply the two textures. This is different from most other 3d accelerators of that era where the number of textures you could apply per pass was tied directly to the number of texture units that were physically present in the hardware.
 
The savage 4 was a 1*2

It's fill rate for Single and dual texture was almost identical. a TNT was a 1*2 or 2*1 it could put out 2 single textured pixels per clock or 1 dual textured pixel per clock. a TNTs dual textured fill rate was 1/2 of it's single textured fill rate.

A savage 2K was able to put out 1 quad textured pixel per clock or 2 single or dual textured pixels per clock. It was speculated that this was the result of using 2 savage 4 3d cores (see above) and adding a T&L unit (unfortunately the clipping engine was borked and it killed them)
 
Last edited by a moderator:
arjan de lumens said:
You have any fillrate benchmarks to support this? The OP's benchmarks appear to contradict you ..?
Click to expand...

Look at the first post in the thread

single 104.X

multi 107.x

All you have to do is google it and you will find that the sites say 2 pipes 1 tezxture unit per pipe or 1 pipe with 2 texture units (3dmark bears out the second)

I had a savage 3d, savage 4 and savage 2000 and I can tell you that the original poster's results were the same as mine for the savage 4
 
Last edited by a moderator:
YeuEmMaiMai said:
Look at the first post in the thread

single 104.X

multi 107.x

All you have to do is google it
Click to expand...
I think you too should look at the first post in the thread; he lists Voodoo3 results too:

3D Mark 1999 default:
single-texturing: 118,4
multi-texturing: 230,5

It appears to be well established that Voodoo3 is 1x2; Savage4 very cleary does NOT behave like a Voodoo3.
 
arjan de lumens said:
I think you too should look at the first post in the thread; he lists Voodoo3 results too:

3D Mark 1999 default:
single-texturing: 118,4
multi-texturing: 230,5

It appears to be well established that Voodoo3 is 1x2; Savage4 very cleary does NOT behave like a Voodoo3.
Click to expand...

A savage 4's fill rate is the SAME for single and multi texturing indicating that it is a 1*2 design. Savage 4 was able to put out 1 single textured pixel per clock or 1 dual textured pixel per clock that is why the fill rate is the SAME for both texts. How else can you explain the fill rate being the same for both tests?
 
YeuEmMaiMai said:
A savage 4's fill rate is the SAME for single and multi texturing indicating that it is a 1*2 design. Savage 4 was able to put out 1 single textured pixel per clock or 1 dual textured pixel per clock that is why the fill rate is the SAME for both texts. How else can you explain the fill rate being the same for both tests?
Click to expand...
These tests meaure TEXEL fillrate, not PIXEL fillrate. I thought the Voodoo3 results would make that OBVIOUS to you.
 
YeuEmMaiMai: arjan de lumens is right, these results indicates, that Savage 4 has one texture unit per pixel pipeline. I thought, that S4 consists of 2 pixel pipelines (2x1) and its fillrate is a bit lower due to memory bandwidth limitation. But 1x1 configuration seems to be more logical. For comparision - 3Dfx Banshee (1x1):

3D Mark 1999 default:
single-texturing: 105,2
multi-texturing: 105,4

3D Mark 2000 default:
single-texturing: 103,2
multi-texturing: 103,2

3DMark 2001SE default:
single-texturing: 137,9
multi-texturing: 137,2

and Matrox Millenium G200 (1x1):

3D Mark 1999 default:
single-texturing: 73,0
multi-texturing: 73,1

3D Mark 2000 default:
single-texturing: 76,7
multi-texturing: 76,7

3DMark 2001SE default (not enough VRAM)
 
It (Savage 4) fits too nicely as 1x1 to not be one according to your benchmarks, but what is going on in 2001SE? Anyone care to remind me why that should be happening or is that just a case of the drivers acting screwy?
 
Ok:

Savage 3D: 1x1, trilinear texture unit
Savage 4: 1x1, trilinear texture unit
Savage 2x: 2x2? but which texture units?
 
You must log in or register to reply here.

Similar threads

J
My R9700 experince and Aniso still "flawd"?
10 11 12
Replies
232
Views
24K
KimB
K
Back
Top