NVIDIA GT200 Rumours & Speculation Thread

[ninelven]

2 times the G92 transistors and 1.5 times G92 performance is an efficiency increment ?

The 1.5x is referring to architectual improvements (a performance per unit per clock increase).

 

[no-X]

Improvements to what? G80 or G94?

 

[ninelven]

I asked first .

nVidias magicians work in marketing department and their misson is simple - to conceal weak points of the new GPU (which seems to be i bit older than G92/G94)

Like what?

 

[no-X]

Like what?

http://forum.beyond3d.com/showpost.php?p=1173034&postcount=2273

 

[ninelven]

So what specifically is wrong with texture adressing in these cards that puts them behind G9x cards... I need you to spell it out for me.

 

[Jawed]

So what specifically is wrong with texture adressing in these cards that puts them behind G9x cards... I need you to spell it out for me.

40 TAs instead of 80?

Jawed

 

[Rys]

It's not 40 though.

 

[Jawed]

It's not 40 though.

Aha, so it is 1:1 TA:TF.

Jawed

 

[Arty]

It's not 40 though.

60.

 

[trinibwoy]

http://forum.beyond3d.com/showpost.php?p=1173034&postcount=2273

Heh how does that picture say anything about GT200 vs G92?

 

[Lukfi]

It's not 40 though.

You mean nVidia NDA'd people and then told them false info?

 

[Rys]

How am I supposed to know?

 

[Lukfi]

How do you know it's not 40?

 

[Jawed]

Double Precision

Something I've been told: each SM has a dedicated double-precision MAD unit, so there's 30 in total. That's a surprise, 1/12th of single-precision, way less than I was expecting, 78 GFLOPs

A HD3650 has about 44GFLOPs DP.

It seems this was being rather over-optimistic:

http://forum.beyond3d.com/showpost.php?p=1129054&postcount=51

suggesting 1/8th double-precision.

I wonder if double-precision transcendentals are directly supported?... ATI doesn't.

Jawed

 

[3dilettante]

I would have thought it would have saved hardware if the DP unit shared hardware with the SP units.

At 1/12, it's almost like only one SIMD of 3 can run quarter-rate DP math.

 

[Jawed]

I would have thought it would have saved hardware if the DP unit shared hardware with the SP units.

At 1/12, it's almost like only one SIMD of 3 can run quarter-rate DP math.

No, it's a dedicated unit. That's what I've been told.

It's really quite a stark difference, it looks like HD4870 will do 311GFLOPs, 4x the performance. Or, HD4870X2 with 8x the double-precision performance for the same money?

Jawed

 

[AnarchX]

How do you know it's not 40?

Read the right, latest documents.

 

[Lukfi]

=>Jawed: How does ATi hardware handle double-precision computing? Through the same SPs that are used for single-precision, or are there some of them that support FP64 while others don't?

=>AnarchX: I don't have those. I just know what nVidia tols us under NDA weeks ago. SP count, TU count, memory bus and size, new features... since you're now saying that it was wrong, I see no problem in telling you, the TU count was supposed to be 40 (I thought 40 TAs + 80 TFs like on G80). So, nVidia did NDA people and told them false info. Feh!

 

[Jawed]

=>Jawed: How does ATi hardware handle double-precision computing? Through the same SPs that are used for single-precision, or are there some of them that support FP64 while others don't?

ATI uses four out of the five ALUs to perform a single-cycle double-precision MAD. It can do 2 independent ADDs per cycle, x,y lanes for one ADD, z,w lanes for the other.

Start reading here for some gory speculation:

http://forum.beyond3d.com/showthread.php?p=1141518#post1141518

based upon Mike's information.

I've just realised that the GFLOPs for HD4870 would be much less than I indicated it should be 249GFLOPs double-precision, 3.2x faster than GTX 280.

Jawed

 

[Love_In_Rio]

=>Jawed: How does ATi hardware handle double-precision computing? Through the same SPs that are used for single-precision, or are there some of them that support FP64 while others don't?

=>AnarchX: I don't have those. I just know what nVidia tols us under NDA weeks ago. SP count, TU count, memory bus and size, new features... since you're now saying that it was wrong, I see no problem in telling you, the TU count was supposed to be 40 (I thought 40 TAs + 80 TFs like on G80). So, nVidia did NDA people and told them false info. Feh!

one dummy question, is double precision used for gaming graphics or only for gpgpu tasks ?