NVIDIA GF100 & Friends speculation

[Jawed]

http://www.szgalaxy.com/Fermi/

GTX480 512 shaders
GTX470 448 shaders

Memory speed not specified for some reason.

Lists a bunch of events in China starting the 26 March, giving away GT240's and Fermi t-shirts, and "priority Fermi purchasing opportunities" if i have the translation correct.

Wow, GTS250s are in short supply? Is this the end of an era?

Jawed

 

[ferro]

Haven't they done this before? Have a launch party at the end of the week, with NDA's lifted the following Monday?

Without any performance numbers this announcement seems a bit over-the-top:

Special "Must See" Presentation
Main Theater (Friday 3/26 from 6 PM to 7 PM only)

Come see NVIDIA unveil the next generation of PC gaming. Want to see what’s hot and what’s next? If you’re even vaguely a fan of PC games and miss this special event, you’ll likely be spending the next few months kicking yourself. Line up early as seating is limited. ‘Nuff said.

Maybe the Nordic Hardware article is wrong though.

 

[AlexV]

I think were dealing with integers and not FP, and INT multiplying I believe is faster on Evergreen than RV770.

That would depend on the INT. For INT32 it should be the same, IIRC.For INT24 it's mucho faster,yes.

 

[Jawed]

That would depend on the INT. For INT32 it should be the same, IIRC.For INT24 it's mucho faster,yes.

These password generation kernels use lots of bitwise operators: left-shift, right-shift, OR, AND and EXCLUSIVE OR, e.g.:

03 LOOP_NO_AL i1 FAIL_JUMP_ADDR(58) VALID_PIX 
    04 ALU: ADDR(187) CNT(109) KCACHE0(CB0:0-15) 
         39  x: XOR_INT     ____,  R30.y,  R2.y      
             y: XOR_INT     ____,  R30.x,  R2.x      
             z: XOR_INT     T0.z,  R2.y,  R33.y      VEC_120 
             w: XOR_INT     ____,  R2.x,  R33.x      VEC_120 
         40  x: LSHL        T0.x,  PV39.z,  1      
             y: LSHL        T0.y,  PV39.w,  1      
             z: XOR_INT     ____,  PV39.x,  (0x000002A0, 9.416725680e-43f).x      
             w: XOR_INT     ____,  PV39.y,  (0x000002A0, 9.416725680e-43f).x      
             t: LSHR        T0.w,  PV39.w,  (0x0000001F, 4.344025239e-44f).y      
         41  x: LSHL        ____,  PV40.z,  1      
             y: LSHL        ____,  PV40.w,  1      
             z: LSHR        ____,  PV40.z,  (0x0000001F, 4.344025239e-44f).x      
             w: LSHR        ____,  PV40.w,  (0x0000001F, 4.344025239e-44f).x      
             t: LSHR        T0.z,  T0.z,  (0x0000001F, 4.344025239e-44f).x      
         42  x: OR_INT      R34.x,  PV41.y,  PV41.w      
             y: OR_INT      R34.y,  PV41.x,  PV41.z      
             z: XOR_INT     ____,  KC0[1].x,  R31.x      
             w: XOR_INT     T0.w,  KC0[1].y,  R31.y      
             t: OR_INT      R4.x,  T0.y,  T0.w      
         43  x: XOR_INT     T0.x,  PS42,  PV42.z      
             y: OR_INT      R4.y,  T0.x,  T0.z      
             z: XOR_INT     ____,  KC0[1].y,  PV42.y      
             w: XOR_INT     ____,  KC0[1].x,  PV42.x      
             t: LSHL        T1.w,  R20.x,  (0x00000005, 7.006492322e-45f).x      
         44  x: LSHL        ____,  PV43.z,  1      
             y: LSHL        ____,  PV43.w,  1      
             z: LSHR        ____,  PV43.z,  (0x0000001F, 4.344025239e-44f).x      
             w: LSHR        ____,  PV43.w,  (0x0000001F, 4.344025239e-44f).x      
             t: XOR_INT     T0.z,  PV43.y,  T0.w      
         45  x: OR_INT      R14.x,  PV44.y,  PV44.w      
             y: OR_INT      R13.y,  PV44.x,  PV44.z      
             z: LSHL        ____,  T0.x,  1      
             w: LSHL        T0.w,  PS44,  1      
             t: LSHR        ____,  T0.x,  (0x0000001F, 4.344025239e-44f).x

Jawed

 

[rpg.314]

I am gonna stir the DP-heavy pot once again. So far, I had thought that DP was a minuscule overhead at best and minor at worst. However, prima facie, the int alu is useless for graphics. And a 32b multiplier is ~40% larger than a 27b multiplier. I am wondering whether it is of any use (apart from address generation) in graphics/DXCS? If not, then it seems to be a pretty big investment for no apparent return.

So, do you have a use case for INT32 mul in graphics/DXCS?

 

[KimB]

I am gonna stir the DP-heavy pot once again. So far, I had thought that DP was a minuscule overhead at best and minor at worst. However, prima facie, the int alu is useless for graphics. And a 32b multiplier is ~40% larger than a 27b multiplier. I am wondering whether it is of any use (apart from address generation) in graphics/DXCS? If not, then it seems to be a pretty big investment for no apparent return.

So, do you have a use case for INT32 mul in graphics/DXCS?

It's conceivable that DP and 32-bit integer are useful for non-graphics processing, such as PhysX.

 

[Jawed]

Byte-wise addressing, a feature of UAVs, doesn't go very far with 24-bit addresses. Bit-shift doesn't work when the row pitch isn't a power of 2, without writing your own multiplier based upon bit-shifts etc.

Jawed

 

[FenderBender]

Ithe int alu is useless for graphics. And a 32b multiplier is ~40% larger than a 27b multiplier....If not, then it seems to be a pretty big investment for no apparent return.

Is it really a big investment? What percentage of the die is used for integer ALUs? Like DP, it may be pretty minor transistor expense to add the wider int multipliers. (This doesn't address your actual question about the use of int32 in graphics, though...)

 

[Jawed]

Is it really a big investment? What percentage of the die is used for integer ALUs? Like DP, it may be pretty minor transistor expense to add the wider int multipliers. (This doesn't address your actual question about the use of int32 in graphics, though...)

NVidia could have lowered the investment by putting MUL32 in the special function ALU, which has considerably lower throughput.

Then NVidia would be stuck without an option for high-performance DP architecture.

Jawed

 

[ZerazaX]

Board Power: < 300W

 

[r9800pro]

Board Power: < 300W

Probably fake

It should be Confidential not Confdential

 

[Ailuros]

I'm afraid TDP values are as exaggerated as early price estimates for the 480.

 

[Silus]

I'm afraid TDP values are as exaggerated as early price estimates for the 480.

Done by the usual suspects

Let's see if that applies to the never-ending speculation on die-size too! And performance as well of course.

Hopefully we get some real game benchmarks in the next few days.

 

[Picao84]

I'm afraid TDP values are as exaggerated as early price estimates for the 480.

What do you mean by that? Exaggerated in what sence? The real ones (both TDP and price) are lower or higher?

 

[Psycho]

edit: too late
I guess they wouldn't write < 300 if it was <275.

 

[neliz]

So, it's 512 CC now again. despite Fudo's estimates earlier this week

 

[Picao84]

So, it's 512 CC now again. despite Fudo's estimates earlier this week

ROFL. nVIDIA must have been doing a disinfo campaign then xD

 

[Lonbjerg]

ROFL. nVIDIA must have been doing a disinfo campaign then xD

People are just getting really frantic this time...I "blame" NVIDIA...for keeping the cards/information so tight and firm this time around.

The rumoursites must be going out of their minds...and making up a lot of stuff in deperation.

This is really "silly season" for full effect

 

[Picao84]

People are just getting really frantic this time...I "blame" NVIDIA...for keeping the cards/information so tight and firm this time around.

The rumoursites must be going out of their minds...and making up a lot of stuff in deperation.

This is really "silly season" for full effect

I dont know really if its rumour sites. nVIDIA might be baiting people with info, to see where the leaks come from

 

[Lonbjerg]

I dont know really if its rumour sites. nVIDIA might be baiting people with info, to see where the leaks come from

Agredd, it seems like NVIDIA really is doing everything (and then some) this launch...I SO look forward to the 27th when we can talk about facts, not FUD/misinformation...