More info about RSX from NVIDIA
- Thread starter nAo
- Start date
does this make sense?
""I crunched the numbers for anyone who's interested. I don't have the time to evaluate wht they all mean right now..... go for it
VS = Vertex Shader
PS = Pixel Shader
Instructions/ALU
RSX = 2 per VS x 8 VS = 16
RSX = 5 per PS X 24 PS = 120
Xenos = 2 per ALU X 48 ALU = 96
Operations / ALU
RSX = 5 per VS X 8 VS = 40
RSX = 10 per PS X 24 PS = 240
Xenos = 5 per ALU X 48 ALU = 240
FLOPS / ALU
RSX = 10 per VS X 8 VS = 80
RSX = 27 per PS X 24 PS = 648
Xenos = 10 per ALU X 48 ALU = 480
Instructions / clock
RSX = 16 per clock per VS X 8 VS = 128
RSX = 120 per clock per PS X 24 PS = 2880
Xenos = 96 / ALU X 48 ALU = 4608
Operations / clock
RSX = 40 per clock per VS X 8 VS = 320
RSX = 240 per clock per PS X 24 PS = 5760
Xenos = 240 per clock per ALU X 48 ALU = 11,520
FLOPS / clock
RSX = 80 per clock per VS X 8 VS = 640
RSX = 648 per clock per PS X 24 PS = 15,552
Xenos = 480 per clock per ALU X 48 ALU = 23,040
Instructions / second
RSX = 8.8B per VS per second X 8 VS = 70.4
RSX - 66B per PS per second X 24 PS = 1584
Xenos = 48B per second per ALU X 48 ALU = 2304
Operations / second
RSX = 22B per second per VS X 8 VS = 176
RSX = 132B per second per PS X 24 PS = 3168
Xenos = 120B per second per ALU X 48 ALU = 5760
Floating Point Operations / Second
RSX = 44B per second per VS X 8 VS = 352
RSX = 356B per second per PS X 24 PS = 8544
Xenos = 240B per second per ALU X 48 ALU = 11,520
But just remember. You can't compare RAW numbers of a standard shader stright across to a Unified shader as Unified shaders are FAR more efficient. The difference is that in a real world gaming scenario, a standard shader doesn't have a chance at actually attaining those figures...... a Unifed shader can."
Is it not poignant at this point to raise that Unified Shaders are theorectically 'far' more efficient? I don't doubt they are, but it's an unproven tech AFAIK. It needs trials on the field to determine how well they perform relative to fixed-function shaders.dukmahsik said:
ShootMyMonkey
Veteran
In theory, you can point to the fact that while RSX would have some 44 GFLOPS dedicated to vs and 356 GFLOPS dedicated to ps, you can say that any fraction of R500's 240 GFLOPS can be for either vs or ps. And it is generally true that between vertex and pixel shaders, vertex shaders tend to be far larger.
For the brave, here's a machine translation of the entire thing:
Translation
I'm starting on it now...
Translation
I'm starting on it now...
V3 said:
It might be an issue of it being proprietary to NVidia, hence not widely used. But in a closed box, that kind of issue evaporates..there is no standard.
Doom3 uses it, IIRC, hence one of the reasons it performs so much better on NVidia cards.
ShootMyMonkey said:
Do you mean the total vertex load typically outweighs the total pixel load? I'm sure that's not the case. There's a reason why pixel shaders have always outnumbered vertex shaders..the load is "typically" weighted toward pixel shading.
They don't say, but the chart implies 9ops - which would make it the standard DOT4 + RSQRT.ERP said:
True, it's just that this is a feature for accelerating a technique that's kind of a deadend IMO. I'd rather have few more free samples of PCF and no stencil accelerator.Titanio said:
Fafalada said:They don't say, but the chart implies 9ops - which would make it the standard DOT4 + RSQRT.ERP said:
Did I ever mention PSP can do this sequence in just 3 clocks?
Yes that's what I don't get
7 ops fot the DP, 1 for the RSQ, and 4 for the scale that makes 12 in my book.
Also is it now a free 32 bit Normalise now or still just 16 bit?