NV43 on the road

[Evildeus]

Seems like it's not that far from launch... At least first samples/pics are available.

Some more informations over here

 

[PowerK]

what's this nv43 ? Isn't nv48 (nv40 refresh) due Q4 this year ?

 

[Dave Baumann]

NVIDIA usually fill in the mainstream entrants at the same time, or before, the high end refresh - in this case this is either the mainstream or low end version of the NV4x line. Judging by the die size I would take a stab at an 8 pipe variant.

 

[anaqer]

Could be, considering that red bold "8" among all those question marks.
Also it looks like they're making it on 110nm ...?

 

[Evildeus]

How did you find out 110nm?

 

[Unknown Soldier]

in between all that scribble it says 0.11 in red

·|¾É¤J·sªº»sµ{¡A¦]¦¹±À´úNV43À³¸Ó±Ä¥Îªº¬O0.11·L¦Ì»s

US

 

[Evildeus]

Oh thanks, i was looking but idin't find anything, must take a closer look next time

 

[Unknown Soldier]

yw

 

[991060]

I think you guys need a little translation, I'll just pick the point:
1. NV43/44 will be native PCIE chip, due to Q3 this year.
2. The chip in the picture is a sample delivered from nVIDIA to an unkown AIB located in China.
3. NV43 is just a shrinked NV40(with 8 pipelines left), all the features of NV40 are retained, including SM3.0 of course.
4. The official name of NV43 is probably PCX 6700, PCX 6500 or even PCX 6200( I think they don't have an idea also). Performance wise, it will be about the same as FX5900. NV44 is a even weaker version, maybe only 4 pipelines.
5. They speculated that NV43 is manufactured using 0.11 technology. Since the heat generation and power consumption are reduced a lot, looks like laptop manufacture can use NV43 as well.

And don't forget to take a look at the comparsion on die size between MX4000, FX5700, NV43 and NV40:
http://www.d-cross.com/cgi-bin/TB_4/upload/28/nvidia-nv43-all-big.jpg

Note that I'm just translating, nothing more, nothing less.

edit: spell

 

[PowerK]

991060,

thanks for the translation.

 

[KimB]

And don't forget to take a look at the comparsion on die size between MX4000, FX5700, NV43 and NV40:
http://www.d-cross.com/cgi-bin/TB_4/upload/28/nvidia-nv43-all-big.jpg

Hrm. The size comparison between the 5700 and NV43 really seems impressive to me. For a chip that has twice the texel rate, and four times the z/pixel rate, the NV43 really isn't much bigger.

 

[RussSchultz]

And don't forget to take a look at the comparsion on die size between MX4000, FX5700, NV43 and NV40:
http://www.d-cross.com/cgi-bin/TB_4/upload/28/nvidia-nv43-all-big.jpg

Hrm. The size comparison between the 5700 and NV43 really seems impressive to me. For a chip that has twice the texel rate, and four times the z/pixel rate, the NV43 really isn't much bigger.

If the pictures are cut/paste next to each other, we can't be sure of the scale between them.

On the 5700, for example, the passives (resistors and caps) seem smaller than the others.

 

[anaqer]

Hard to tell because of the watermarks, but it looks like actual chips lined up to me.
( At least I couldn't find anything that's pointing to photochopping. )

Chalnoth... how did you get the 4x multiplier for the Z perf.?

 

[KimB]

Well, I had thought that the 5700 had half the z-rate of the 5900 per clock, putting it at 4 z ops per clock. I could be wrong, I suppose.

 

[KimB]

If the pictures are cut/paste next to each other, we can't be sure of the scale between them.

On the 5700, for example, the passives (resistors and caps) seem smaller than the others.

That is true, but your observation would seem to make things look even better for the NV43.

 

[anaqer]

Well, I had thought that the 5700 had half the z-rate of the 5900 per clock, putting it at 4 z ops per clock.

You're right, it was just a brain fart on my part.

 

[PeterAce]

And don't forget to take a look at the comparsion on die size between MX4000, FX5700, NV43 and NV40:
http://www.d-cross.com/cgi-bin/TB_4/upload/28/nvidia-nv43-all-big.jpg

Hrm. The size comparison between the 5700 and NV43 really seems impressive to me. For a chip that has twice the texel rate, and four times the z/pixel rate, the NV43 really isn't much bigger.

Yes, I was impressed with increase of number of quads and feature ratio to die size of NV4X over NV3X.

It's been posted here before that one likely reason for this is that Nvidia outsourced the 'design and transistor layout' of the TEX/ALU units in NV4X. Looks like they did a good job.

 

[Dave Baumann]

I'm not sure they outsourced the entire design, I think they brought in cell designs for some mathematical units, which were supposed to be more optimal in terms of layout.

Bear in mind, though, that they did remove some things (SIN/COS) and the two ALU's in each are not really duplicates of each other, and instructions are distributed between them.

 

[Tridam]

I'm not sure they outsourced the entire design, I think they brought in cell designs for some mathematical units, which were supposed to be more optimal in terms of layout.

Bear in mind, though, that they did remove some things (SIN/COS) and the two ALU's in each are not really duplicates of each other, and instructions are distributed between them.

I also think that we can't compare NV35 pipelines to NV40 pipelines. One NV40 pipeline doesn't require as much transistors and die space than an NV35 pipeline does. NVIDIA removed one TMU per pipeline and removed the combiners.

Maybe we could say that they've also added a second big ALU. But, as you say, the 2 shading units are basically a splitted NV30 big ALU.

However, on the other side, they've added a lot of flexibility to the new pipelines (co-issuing...), a modifiers support, an independent fp16 nrm, more registers...

Regarding sincos, it isn't as efficient as it was with NV3x. However it is still very fast (way faster than with R3x0/R420) : 2 cycles for sin or cos and 3 cycles for sin and cos. and some resources are available during these cycles.