[3dcenter] GK104 specs

[Silverforce11]

You guys who think a die smaller than Tahiti while being faster than it is dreaming.

In the past several generations, NV has always had to have much bigger die sizes to offer better performance.

AMD has been the one offering superior perf/mm2.

At 342mm2, its going to be around gtx580 AT BEST.

 

[Keysplayr]

Based on these specs, yes. The GK104 would have at least GTX560Ti SLI performance without the SLI overhead. (scaling). If this is available at a 299.00 price point, that would be truly amazing. Hard to see that happening though.

 

[boxleitnerb]

You guys who think a die smaller than Tahiti while being faster than it is dreaming.

In the past several generations, NV has always had to have much bigger die sizes to offer better performance.

AMD has been the one offering superior perf/mm2.

At 342mm2, its going to be around gtx580 AT BEST.

Yeah, things never change...
You honestly think that Nvidia would make its GF114 successor only 30-40% faster? That is ridiculous.

 

[Silverforce11]

Based on these specs, yes. The GK104 would have at least GTX560Ti SLI performance without the SLI overhead. (scaling). If this is available at a 299.00 price point, that would be truly amazing. Hard to see that happening though.

Assuming each CUDA core is equivalent throughput on all processing.

That's a big IF.

 

[sontin]

And why it's dreaming?

GF114 is 360mm^2 huge and 30% faster than the GT200b - with only 81% of the bandwidth.

GK104 with these specs and a die size of around 350 mm^2 will also much faster than the GTX580 - with only 8x% of the bandwidth.

 

[BallaTheFeared]

It's not, this leak has no hot clocks, thus the increase in core count.

It has a smaller bus and many other changes there is no way for anyone to make the comment you did.

You have no idea what Nvidia may or may not have done to reduce die size, everything about these reports is totally different than what we've seen over the past three generations at least.

 

[Keysplayr]

You guys who think a die smaller than Tahiti while being faster than it is dreaming.

In the past several generations, NV has always had to have much bigger die sizes to offer better performance.

AMD has been the one offering superior perf/mm2.

At 342mm2, its going to be around gtx580 AT BEST.

Silverforce, this is a huge falsehood. Nvidia has bigger dies due to the GPGPU driven architecture, not because it needs a bigger die to compete in gaming with AMD, which has NOT been geared toward GPGPU, which is now changing slowly for them. GCN. And what's more, I have no idea how anyone can guess at gaming performance from die size alone. Can't be done.

 

[Silverforce11]

You have no idea what Nvidia may or may not have done to reduce die size, everything about these reports is totally different than what we've seen over the past three generations at least.

Exactly, nobody knows yet thus all speculative options are open.

My bet is gk104 with these die size and specs will struggle to match a gtx580. Sure, the specs could be all crap and we won't know til next to release.

 

[Silverforce11]

Silverforce, this is a huge falsehood. Nvidia has bigger dies due to the GPGPU driven architecture, not because it needs a bigger die to compete in gaming with AMD, which has NOT been geared toward GPGPU, which is now changing slowly for them. GCN. And what's more, I have no idea how anyone can guess at gaming performance from die size alone. Can't be done.

So you're saying Kepler will not focus on GPGPU as usual? Dream on, NV has a hard-on for that market segment.

It's a HPC product that can run games, not going to change in that aspect.

The only big variable and unknown is hotclocks and the performance of the new cores. If it was a straight Fermi shrink, you could have a better chance at guessing performance based on die size but its not. My point is purely based on the past few generations.. for NV to suddenly release a new HPC part with a small die offering much faster performance, unlikely and thus, dreaming.

ps. My original statement is fact, not falsehood. Whatever NV design their GPU for (HPC), its perf/mm2 has been subpar compared to AMD. For them to turn this around on a new node (!) with a new architecture (!)... yeah...

 

[boxleitnerb]

All things point to a part engineered to be better at gaming (but lacking in the HPC department) and the usual big chip. So no "dream on", think outside of the box.

 

[Silverforce11]

Btw, speculate this; If NV really will drop their hotclock design and move the architecture to more but less complex cores.. why do you think they went down this road?

Is it for perf/w improvements?

Better throughput on sp while retaining decent dp? This was a big weakness with CUDA cores.. not appealing to some HPC segment that depended on sp performance (and bitcoin mining servers!)

On specs its not lacking in HPC at all, the Dp is 1/4th the sp output and very high for a mid-range card (nearly matching gtx580, impressive feat for a "non HPC, gaming focused part").

 

[sontin]

Huh? Fermi's Perf/mm^2 improvement was huge. nVidia struggled to compete with their 480mm^2 die against rv770 with 260mm^2. But with GF1x0 and GF1x4 they were very competitive. They will not to do a step back to the old GT200(b) days.

 

[Keysplayr]

So you're saying Kepler will not focus on GPGPU as usual? Dream on, NV has a hard-on for that market segment.

I'm saying:

"Silverforce, this is a huge falsehood. Nvidia has bigger dies due to the GPGPU driven architecture, not because it needs a bigger die to compete in gaming with AMD, which has NOT been geared toward GPGPU, which is now changing slowly for them. GCN. And what's more, I have no idea how anyone can guess at gaming performance from die size alone. Can't be done."

 

[Crap Daddy]

Exactly, nobody knows yet thus all speculative options are open.

My bet is gk104 with these die size and specs will struggle to match a gtx580. Sure, the specs could be all crap and we won't know til next to release.

Based on these specs this chip was designed to be twice the GF104-114.
The only thing we can't tell now is HOW much faster will it be than the 580 and not IF.

 

[Silverforce11]

I'm saying:

"Silverforce, this is a huge falsehood. Nvidia has bigger dies due to the GPGPU driven architecture, not because it needs a bigger die to compete in gaming with AMD, which has NOT been geared toward GPGPU, which is now changing slowly for them. GCN. And what's more, I have no idea how anyone can guess at gaming performance from die size alone. Can't be done."

Note i never compared gk104's potential vs 7970.. i compared it to gtx580.

So whats your point? Kepler is HPC focused as was Fermi.

 

[BallaTheFeared]

Nvidia could never match, let alone exceed AMD's per/nm guys, stop living in your dream world Nvidia sucks and needs more die space to compete with the master crafter that is AMD.

 

[Keysplayr]

Note i never compared gk104's potential vs 7970.. i compared it to gtx580.

So whats your point? Kepler is HPC focused as was Fermi.

My point was made. It's just a matter of you acknowledging it. It cant get simpler.

 

[Silverforce11]

Based on these specs this chip was designed to be twice the GF104-114.
The only thing we can't tell now is HOW much faster will it be than the 580 and not IF.

What makes you say that? You are assuming the new cores are just effective vs the old ones at the same speed? If so, it's magic, because somehow NV managed to squeeze in ~4 times as many cores and make the chip smaller than the gtx560, all on a 40nm -> 28nm shrink.

Doubling specs also requires bandwidth. It's not magic.

 

[Silverforce11]

My point was made. It's just a matter of you acknowledging it. It cant get simpler.

Your point is wrong and accusing me of falsehood which you cannot prove. My statement that NV has required much bigger die sizes to offer better performance stands.

My gk104 perf estimate is vs gtx580. Again your point is irrelevant because Kepler will still focus on HPC.

 

[Keysplayr]

Your point is wrong and accusing me of falsehood which you cannot prove. My statement that NV has required much bigger die sizes to offer better performance stands.

My gk104 perf estimate is vs gtx580. Again your point is irrelevant because Kepler will still focus on HPC.

People here are smart enough to see what's real and what's BS. So, suit yourself.

 

[Silverforce11]

People here are smart enough to see what's real and what's BS. So, suit yourself.

Hey, if you're going to make a habit of accusing people stating facts as falsehood, i'll make note of it to ignore those comments for future reference.

 

[sontin]

Hey, if you're going to make a habit of accusing people stating facts as falsehood, i'll make note of it to ignore those comments for future reference.

Wait, he is right:

ps. My original statement is fact, not falsehood. Whatever NV design their GPU for (HPC), its perf/mm2 has been subpar compared to AMD. For them to turn this around on a new node (!) with a new architecture (!)... yeah...

Maybe you should check your facts. Fermi has a much higer perf/mm^2 improvement over the last nvidia gen than Cypress or Tahiti.

 

[Crap Daddy]

What makes you say that? You are assuming the new cores are just effective vs the old ones at the same speed? If so, it's magic, because somehow NV managed to squeeze in ~4 times as many cores and make the chip smaller than the gtx560, all on a 40nm -> 28nm shrink.

Doubling specs also requires bandwidth. It's not magic.

It's not four times it's 2x. With Fermi it would have been 768 SP or CUDA cores, as they name them. The chip is not much smaller than Tahiti XP and 560Ti has 1.9 bil Transistors so to squeeze double that amount on 28nm node on the aforementioned die-size is a piece of cake, not that higher transistor count equals better performance.

 

[boxleitnerb]

Also Kepler != Kepler. GK110 might be much more HPC oriented than GK104. Everything from GK104 down seems to have been engineered for maximum efficiency in perf/mm and perf/W. Again, why make a successor only 30-40% faster? It is impossible that GK104 is only on par or below GTX580. That would be to little improvement for a tweaked architecture and a node jump.

 

[Silverforce11]

Wait, he is right:

Maybe you should check your facts. Fermi has a much higer perf/mm^2 improvement over the last nvidia gen than Cypress or Tahiti.

Did i say improvement?

I said plainly, NV has much bigger die sizes to offer faster performance. That statement is truth. You can look it up from several generations past, 4870 vs gtx280, 5870 vs gtx480, 6970 vs gtx580...

The die size/perf gap is big, now you suddenly expect NV to turn it all around and have their 342mm2 gk104 faster than 7970? If that isn't dreaming..