RV730 - where are the 32 TMUs?

CarstenS

Moderator
Moderator
Legend
Supporter
I've been testing the HD 4670 for today's launch but have not been able to extract more than (very close to) 16 textures per clock with any fillrate-meter I know and have at my disposal. That includes the 3DMarks' obviously, but also Rightmark 3D, Archmark and various in-house tools.

On page two, I've described this.
http://www.pcgameshardware.com/aid,659355/Reviews/PCGH_Review_Radeon_HD_4670_and_HD_4650/

Did any of you have more luck wrt to this? I've already talked to AMD and they have some in-house tools (which they would not disclose for obvious reasons) which were able to do just that.

Now, I wonder, if it's that difficult to extract more than 16 tex/clk - how would the chip be able to profit from it and thus justify their tribute in die-size?


edit:
FWIW: I can rule out pure bandwidth limitation - even with the core down to 500 and the memory (which suffices for about 15.9 GTex/sec in normal operation) up to 1.050 MHz, I'm only getting (very close to) the theoretical peak of 16 TMUs @500 MHz.


edit 2:
I've done some rightmarking, especially those test, which were enormously improved on with RV770 vs. RV670. Here's two of them: RV730 scores are almost (too exactly) 40% of RV770s which would also be the case if it really had only 16 TUs instead of [strike]32[/strike] 40.
Code:
	       HD 4870  HD 4670 HD3870
D3D Rightmark 1.0.5.0 - Proc. Fur (PS3.0)		
- 1600x1200	55,41	23,96	15,21
- 1920x1200	45,42	20,07	12,47
- 2560x1600	24,72	10,87	6,74
			
D3D Rightmark 1.0.5.0 - Light. St. Par. Mapp. (PS3.0)			
- 1600x1200	150,48	59,81	38,73
- 1920x1200	127,91	50,54	32,41
- 2560x1600	76,19	29,96	19,45
 
Last edited by a moderator:
Is this another case of fewer interpolators than texture units?

I don't think so - AMD said, it wasn't the case with RV730.

Why not test with trilinear filtering or 2x bi-AF? So you could see if there are 32 TFs. ;)
I thought of that, too. But as I said to Pete already, AMD told me this is not a matter of missing Interpolators.
 
Well, at least the reviews that are now officially out all say 16 TMUs, was that 32 TMUs actually confirmed by AMD or just a rumor?
 
Out of the 3 reviews I saw (hothardware, anand, techpowerup) Anand said 16, the rest nothing regarding it and none had that slide
 
Out of the 3 reviews I saw (hothardware, anand, techpowerup) Anand said 16, the rest nothing regarding it and none had that slide
Anands table says 16, but look at the AMD slide under it:
Texture-Fillrate: ...
http://www.anandtech.com/video/showdoc.aspx?i=3405
;)
And Hot HW says 32, too: http://www.hothardware.com/Articles/ATI-Radeon-HD-4670-Redefining-The-Mainstream/?page=2

But the question is still, why you could not measure this 32 TMUs in synthetic benchmarks, that there here show the CoD4 results, which is tex-demanding.
 
Since I am not a hundred percent sure as to how to interpret the results from Aruns programme, I'd rather leave the (possible) solution to this riddle to him.
 
I think you are right.
No matter what AMD said ... :LOL:
Sounds unlikely imho. Note that all diagrams say there are 8 (8-wide) simd clusters, with each having a texture unit. Somehow I don't think those tus will have 2 TA / 4 TF capability - seems far more likely they are the same (much more simple design than rv670) design as rv770.
 
Back
Top