MDolenc's Fillrate Tester: GF FX FP32/FP16 Performance

Dave Baumann

Dave Baumann

Gamerscore Wh...
Moderator
Legend
We've seen a few posts speculating that the FP16 performance of 5900 has increased over 5800. MDolenc's 'Fillrate Tester' app has tests for both PS2.0 (FP32) and PS2.0 PP (FP16):

Fillrate Tester
--------------------------
Display adapter: NVIDIA GeForce FX 5800 Ultra
Driver version: 6.14.10.4403
Display mode: 1024x768x32bpp
--------------------------
FFP - Pure fillrate - 1877.099976M pixels/sec
FFP - Single texture - 1516.172729M pixels/sec
FFP - Dual texture - 1274.996704M pixels/sec
FFP - Triple texture - 734.205444M pixels/sec
FFP - Quad texture - 698.807068M pixels/sec
PS_2_0 - Per pixel lighting - 121.043259M pixels/sec
PS_2_0 PP - Per pixel lighting - 163.160095M pixels/sec

Fillrate Tester
--------------------------
Display adapter: NVIDIA GeForce FX 5900 Ultra
Driver version: 6.14.10.4403
Display mode: 1024x768x32bpp
--------------------------
FFP - Pure fillrate - 1735.464355M pixels/sec
FFP - Single texture - 1637.989868M pixels/sec
FFP - Dual texture - 1416.287720M pixels/sec
FFP - Triple texture - 759.043274M pixels/sec
FFP - Quad texture - 738.639954M pixels/sec
PS_2_0 - Per pixel lighting - 149.738754M pixels/sec
PS_2_0 PP - Per pixel lighting - 181.771698M pixels/sec

As we can see the pure PS2.0 performance of 5900 has increased a little (not exactly 2x though!), however, while the PS2.0 PP performance also increases with 5900 the actual relative difference from PS2.0 and PS2.0 PP between the board actually decreases - would this be the case if 5900 had more FP16 units?
 
Hmm, interesting.
But I'd take great care with those results. I've seen sites claiming that the FX5800 sometimes used FX12 with the 44.03 without asking anyone's opinion.
Some sites say it's false though. All very confusing.

BTW, could you try the Dawn patches I did on your FXs?
www.notforidiots.com/DawnQuality.zip
www.notforidiots.com/DawnFull12.zip

( The FX12 in DawnQuality.zip is bugged, use the one in DawnFull12.zip instead )

Of particuliar interest would be any difference between FP16 in FP32 and pure FP32. The FP16 score would also be very nice, of course :)


Uttar
 
Uttar said:
Hmm, interesting.
But I'd take great care with those results. I've seen sites claiming that the FX5800 sometimes used FX12 with the 44.03 without asking anyone's opinion.
Some sites say it's false though. All very confusing.
Click to expand...

Could anyone with a FX5800 try my pixel shader benchmark program? (Download)

Type:

fillrate9 precision.txt

and check the output.txt file, and it will give the precision test results. It is not designed for fixed point numbers so there could be some problems if FX5800 really uses FX12 with 44.03 driver.
 
FX 5600 w/ 44.03 said:
Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 10 bits
PP Precision: 10 bits
Fillrate: 10.827 Mpix/s
Click to expand...
 
tested

Radeon 9500 PRO
typed fillrate9 precision.txt

got this:

Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 4 bits
PP Precision: 8 bits
Fillrate: 64.504 Mpix/s


interesting...
 
Err, what?
pcchen, are you sure that program actually works? :) If so, what does those results mean?


Uttar
 
Radeon9700Pro:

Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 0 bits
PP Precision: 8 bits
Fillrate: 58.843 Mpix/s
 
9800Pro

Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 16 bits
PP Precision: 16 bits
Fillrate: 96.667 Mpix/s
 
Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 16 bits
PP Precision: 16 bits
Fillrate: 82.368 Mpix/s

r9700pro
 
got the problem: i've switched off AA & AF on the D3D panel in the driver, and the new results:

Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 16 bits
PP Precision: 16 bits
Fillrate: 70.897 Mpix/s
 
and for the clear picture, here are some results from MDolenc's 'Fillrate Tester' with the 9500 Pro (AA , AF on application preference)

Fillrate Tester
--------------------------
Display adapter: RADEON 9500 PRO / 9700
Driver version: 6.14.10.6343
Display mode: 1024x768x32bpp
--------------------------

Color writes enabled, z-writes disabled:
FFP - Pure fillrate - 1660.003906M pixels/sec
FFP - Single texture - 1044.150757M pixels/sec
FFP - Dual texture - 659.143250M pixels/sec
FFP - Triple texture - 397.180237M pixels/sec
FFP - Quad texture - 292.569489M pixels/sec
PS_2_0 - Per pixel lighting - 160.018143M pixels/sec
PS_2_0 PP - Per pixel lighting - 160.006973M pixels/sec



there is no difference between the ps 2.0 and the pp results...and my 9500 pro beats the 5900u in the ps 2.0 test...wow :oops:
 
Radeon 9700 Pro

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 16 bits
PP Precision: 16 bits
Fillrate: 77.810 Mpix/s


FX5200 Ultra
Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 10 bits
PP Precision: 10 bits
Fillrate: 10.756 Mpix/s
 
I think somethings wrong here this is the results I got.

Direct3D 9 Fillrate tester V0.3 [beta]
Copyright(c) 2003 Ping-Che Chen

Initialize: 1024x768x24
No shader file specified. Perform accuracy test.
Precision: 245 bits
PP Precision: 245 bits
Fillrate: 72.189 Mpix/s

This is on my 9700pro cat 3.4's
 
Wavey, can you include all of these results (with some of the proposed analysis), from MDolnec's and pcchen's applications in your 5900 ultra review?
 
Luminescent said:
I guess what the pcchen's application means by "precision" is the mantissa precision of the applications results.
Click to expand...

I also think this is the mantissa. _PP is skipped on Radeon 9500/9600/9700/9800 and I'm sure that the detonators drivers decrease precision. So they could be correct results.


The question is : what's the range ? I think Radeons may use FX16 or FP24. But it's pretty sure that they do FP24 here.

What about GeForce FX ? FX12 and FP16 could have a precision/mantissa of 10 bits (I could be wrong with FX12). By default maybe they do FP16 and with _PP they do FX12 ?

If by default they do FP16 or FX12, they're not DX9 compliant :(
 
You must log in or register to reply here.

Similar threads

CarstenS
AMD vs. Nvidia: Per Unit/clock performance of recent architectures
Replies
4
Views
5K
fellix
F
B
Why Barts is really VLIW4, not VLIW5 (and more on HD 5830/6790 being mainly 128-bit)
5 6 7
Replies
132
Views
37K
Alexko
A
Z
3D Labs Wildcat Realizm 100
Replies
11
Views
4K
Nite_Hawk
N
A
  • Locked
List of Rendering Resolutions
Replies
4
Views
1M
AlNom
A
Z
Matrox Millennium P650
Replies
7
Views
2K
Zvekan
Z
Back
Top