NVIDIA confirms Next-Gen close to 1TFlop in 4Q07

humm , it's like a compilation so ?


and why G80 ALU's MUL function is not utilized ?

 

Am I the Only One?

Am I the only one who read the thread title as "1 TFlop by 4007"?

I was ready to reply that I hoped we could reach a two or three times improvement over what we have now, given another two-thousand years to work at it.

 

Question: Was the 1Tflop number reached with a single GPU solution or two e.g GX2 style?

The thing is that nVIDIA could release a GX2 type card is where the move to 65nm makes this possible and they can easily reach that 1tflop number using 2 GPUs instead of one. Another evidence that backs this up is that GX2 style waterblock at cebit.

http://xtreview.com/addcomment-id-1995-view-GeForce-8950-GX2-alpha-cool-water-block.html

Could they move to 256bit if this was their approach?

Just food for thought.

 

It is utilized...kindof...maybe...sometimes.

 

Should be G92-Hummer

 

The whole history, yes. Also the fact that they never focused on performance, it's not like they can suddenly make a leap over 2-3 generations to catch up. also, nV and ATI are developing their stuff further, so Intel is shooting for a (fast) moving target. From today's POV their Larabee etc. might look like it had lots of potential, but don't forget that the competition is not sleeping either.

 

TSMC is simply cheaper, I think.

 

bur not for shading , right ?

 

If I remember correctly, Nvidia also experience bad yields when they used IBM's fabs. I could be thinking of something else however.

Regards,
SB

 

Well the bad yields were probably the main reason for higher prices on the IBM side, I ass-ume

 

Great Scott!

 

Well, I'd presume it would be quite practical as a room heater. And at that level of power, the ecologists might come complaining about global warming - but it turns out not all of us are against that!

 

I still ask my self about the ~1 TFLOPs.
Is this really meant for one G92 or possible solutions based on this chip ->GX2(2xG92)?

 

My speculation on the subject certainly would imply so: http://forum.beyond3d.com/showpost.php?p=983681&postcount=795

That kinda assumes I'm right on the die size though (200-240mm²). If the chip is even bigger, then 8 clusters (ala G80, with 32 TMUs) with 8x3 ALUs and 8x2 Interpolators might be an attractive design choice. Interestingly, if the core clock is 800MHz and the shader clock is 2.5GHz (both of which are a bit high but heh), then they have effectively doubled the ALU ratio anyway (((192*2.5/0.8)/(128*1.35/0.575)) = ~2.0)!

That still doesn't tell me what they did with the ROPs though. heh.

 

1 TFLOP is easily attainable on a G80 based architecture if Nvidia continues to count the MUL.

G80 has been pushed to over 1600Mhz shader clock at 90nm. Given the probable inclusion of PV2 in the refresh and a smaller die compared to G80 I'd say a 25% bump in shader count to 160 and a 2Ghz clock isn't too much to expect.

 

Yeah. It's also far from impossible to get to that number without the MUL (192*2.6GHz+), but it *is* much more ambitious.

We'll see. I certainly wouldn't say no to a Quad-SLI setup at 4TFlops (MADD-only), but that would be a tad much to ask for on 65nm, perhaps.

 

I still can't believe there is no 64 shader variant of G80. I guess ATi might be onto something with cutting back on the number of different chips for the R700 era.

 

Spring 2008

 

681 million transistors, 612MHz in the main core and 1500MHz in the shader core also seemed impossible at 90nm and yet, here we are.
It all depends on how good their 65nm and 55nm products turn out in terms of yields, clocks, etc, but so far not a single one has surfaced.

 

Yeah, don't get me wrong - I actually think 192SPs at 2.4GHz+ is very likely. Even if I think it's likely, that doesn't mean I wouldn't be positively surprised anyway though!