[Xbit] Nvidia: GeForce "Kepler" Graphics Processors Will Be Unbeatable.

[Silverforce11]

As enthusiastic as I am about Nvidia's upcoming GPU's, I don't think GK104 will be, on average, 20% faster than an hd7970. If it is, then color me super impressed.

There's two scenario of my expectation of gk104 performance.

1. ~ gtx580 perf at much reduced power use (~175W), but with a big OC headroom.
2. ~20% faster than gtx580 at ~225W, little OC headroom.

If its better than these two scenarios, I will be very impressed.

The launch end of this month should coincide with 7870 release... something to look forward to.

 

[Keysplayr]

Well, we know it weill be slower than GK104, odd as that sounds. Because GK104 *is* unbeatable.

Nvidia would never claim that a card is unbeatable if they were planning to beat it themselves, would they?

Omg you're right. Nvidia should close up shop after gk104
released since they can never outdo it. Or could "unbeatable"
Just be directed at Nvidias competition?

 

[railven]

Omg you're right. Nvidia should close up shop after gk104
released since they can never outdo it. Or could "unbeatable"
Just be directed at Nvidias competition?

They did say it was their best chip ever, so you might be onto something

 

[BallaTheFeared]

Pretty sure they said Kepler, not GF104 though.

 

[tviceman]

This. I'd be pretty much be dumbfounded if a ~340mm² GPU beat a 352mm² GPU by 20%.

I don't think it's completely unrealistic, especially given that Tahiti is AMD's first full-on dual purpose compute GPU AND they were pretty conservative with Tahiti's clocks.

Keep in mind that Tahiti has more die spaced devoted to it's memory controller, which means GK104 has more die space devoted to cores, cache, etc. A lot of people are saying there is no way GK104 will be able to keep up with an hd7970 at 2560x1600 or higher resolutions because of the lower memory bandwidth on GK104, and while I think that hd7970's performance will degrade less than GK104 at super high resolutions, I don't think bandwidth is going to be a big detriment to GK104. The gtx560ti has 20% less memory bandwidth (from the perspective of a gtx285, a gtx285 has 25% more bandwidth) yet is 40-50% faster across the board than a gtx285. The most current rumors say GK104 will have 160gb of memory bandwidth, 20% more than a gtx560ti. Optimizations in architecture and cache efficiency can do quite a bit to improve performance before memory bandwidth hits a wall.

Just food for thought.

 

[AnandThenMan]

Keep in mind that Tahiti has more die spaced devoted to it's memory controller...

How much more?

 

[railven]

Pretty sure they said Kepler, not GF104 though.

Haha touche! Hopefully we get some leaks to rain on 78xx launch. Finally time to see what's under the hood!

 

[tviceman]

How much more?

I don't know exactly, I'm just stating possible explanations as to why a smaller chip with less memory bandwidth could be just as fast as a bigger chip with more memory bandwidth. With Cayman, it has less die space devoted to it's memory controller than GF100/GF110 and beats GF110 on performance per mm^2. So I'm just saying we could be looking at a situation similar to that, where GK104 might clearly be faster per mm^2, and since the dies are close in size, GK104 would end up faster than an hd7970.

But keep in mind I am just speculating.

 

[AnandThenMan]

I don't think the memory controller takes up much of a transistor budget at all, but it does contribute to heat output in a more significant way vs. other parts of the die from what I understand. And a wider bus probably makes the die physically bigger out of connectivity constraints, although I'm also just speculation here.

The way I look at, AMD has more room to grow with Tahiti than Nvidia with Kepler because of the memory controller. There is only so much Nvidia can do to bump up performance on a 256bit bus. For the target market the first iteration of Kepler is aiming for, it probably wont' matter, but at the same time it may just prevent Nvidia from taking the performance crown until "Big K" comes out.

 

[BallaTheFeared]

Kepler probably doesn't only feature a 256 bus card and that's it.

GF104 is the GTX 460/560ti replacement, not the 470/570 replacement or even the 480/580 replacement. All three of those generational cards had a different size bus, so I wouldn't assume Kepler is stuck on 256.

The really interesting thing, which above all else I'm looking forward to seeing is if we'll see the same trend AMD has set, with the same or nearly the same price/performance ratio, or if the GK104 card will take a slot outside where the 460/560ti released at, which was the $250~ range.

 

[tviceman]

I don't think the memory controller takes up much of a transistor budget at all, but it does contribute to heat output in a more significant way vs. other parts of the die from what I understand. And a wider bus probably makes the die physically bigger out of connectivity constraints, although I'm also just speculation here.

The way I look at, AMD has more room to grow with Tahiti than Nvidia with Kepler because of the memory controller. There is only so much Nvidia can do to bump up performance on a 256bit bus. For the target market the first iteration of Kepler is aiming for, it probably wont' matter, but at the same time it may just prevent Nvidia from taking the performance crown until "Big K" comes out.

I wasn't refering to transistors or to transistors per mm^2, only to actual amount of die space used by memory controllers. But insofar as performance I again point to the comparisons of gtx285's memory bandwidth and gtx560ti's memory bandwidth vs. performance. The gtx285 has 25% more memory bandwidth, yet the gtx560ti is 40-50% faster across the board.

The really interesting thing, which above all else I'm looking forward to seeing is if we'll see the same trend AMD has set, with the same or nearly the same price/performance ratio, or if the GK104 card will take a slot outside where the 460/560ti released at, which was the $250~ range.

No I don't see it coming in at $250 like the gtx560ti. It will have more memory (which will increase costs) and 28nm wafers are more expensive. Beyond those two variables, there will likely be an additional premium. But I believe, and at the very least strongly hope, relative to hd7970 and hd7950's prices, GK104 will completely obliterate those cards in value. Putting the card at $350 alone should make Nvidia froth at the mouth. At similar yields, adding $40 per card for 1gb of memory and 28nm increased costs, that's an additional $60 per card vs. the gtx560ti and $90-110 more than the gtx460 1 gig/768 gig intro'd at. I think Nvidia has better business sense than AMD. I think they are going to look at their sales and see that the gtx560 (ti and non ti) outsold gtx570 and gtx580 combined and won't forget where most of their revenue from desktop GPU's comes from.

Again, all speculation but it won't be long now until we find out.

 

[AnandThenMan]

I wasn't refering to transistors or to transistors per mm^2, only to actual amount of die space used by memory controllers. But insofar as performance I again point to the comparisons of gtx285's memory bandwidth and gtx560ti's memory bandwidth vs. performance. The gtx285 has 25% more memory bandwidth, yet the gtx560ti is 40-50% faster across the board.

You are comparing a completely different architecture and a vastly more powerful chip in general in Fermi, it's simply not a valid comparison. A GTX 285 could have 10 times the bandwidth and it simply would not able to utilize it. GCN and Kepler are by all accounts very similar, so the bandwidth comparison is much more valid IMO.

 

[tviceman]

You are comparing a completely different architecture and a vastly more powerful chip in general in Fermi, it's simply not a valid comparison. A GTX 285 could have 10 times the bandwidth and it simply would not able to utilize it. GCN and Kepler are by all accounts very similar, so the bandwidth comparison is much more valid IMO.

I was comparing a gtx285 with a gtx560ti.
texture fill rate: gtx285 - 51.8, gtx560ti - 52.5
pixel fill rate: gtx285 - 21.4 gtx560ti - 26.3
memory size: both 1 gig
memory bandwidth: gtx285 - 159 gb/s, gtx560ti - 128 gb/s

I think the different architectures makes it an even more valid comparison, because the point I was trying to make originally was that Nvidia has been able to improve their current mid-range offering's performance over their previous flagship, despite having lower memory bandwidth.

EDIT: And we really don't know how similar Kepler and GCN are. There have not been any white papers or technical analysis done on Kepler. GCN has quite a bit in common with Fermi, but Kepler is obviously not Fermi. Even still, as similar as Kepler and GCN might be, the fundamental designs of both architectures are very different, and thus are doing very different things internally to come to the same result. Different cores, different cache sizes, I really could go on and on so I personally don't see how comparing a GT200 to Fermi and Fermi to Kepler is any less valid than comparing Kepler to GCN. I think comparing the same company's own architectures against one another is the best way to predict future performance. Just like one wouldn't predict Ivy Bridge's performance based on bulldozer, or Piledriver's performance based on Ivy Bridge.

 

[omeds]

I think the point was though, mem bandwidth isnt a fixed level of performance, and with other improvements and tweaks it could still perform very well despite is memory bandwidth, and faster than say the GTX580 which has more.

In saying that though, high end 7 series will still murder it in eyefinity I'm sure, but this GK104 could still do very well on single displays upto 1600p.

 

[wahdangun]

I was comparing a gtx285 with a gtx560ti.
texture fill rate: gtx285 - 51.8, gtx560ti - 52.5
pixel fill rate: gtx285 - 21.4 gtx560ti - 26.3
memory size: both 1 gig
memory bandwidth: gtx285 - 159 gb/s, gtx560ti - 128 gb/s

I think the different architectures makes it an even more valid comparison, because the point I was trying to make originally was that Nvidia has been able to improve their current mid-range offering's performance over their previous flagship, despite having lower memory bandwidth.

EDIT: And we really don't know how similar Kepler and GCN are. There have not been any white papers or technical analysis done on Kepler. GCN has quite a bit in common with Fermi, but Kepler is obviously not Fermi. Even still, as similar as Kepler and GCN might be, the fundamental designs of both architectures are very different, and thus are doing very different things internally to come to the same result. Different cores, different cache sizes, I really could go on and on so I personally don't see how comparing a GT200 to Fermi and Fermi to Kepler is any less valid than comparing Kepler to GCN. I think comparing the same company's own architectures against one another is the best way to predict future performance. Just like one wouldn't predict Ivy Bridge's performance based on bulldozer, or Piledriver's performance based on Ivy Bridge.

then how about HD 5770 vs HD 4870 ?? or HD 7770 vs HD 6870 ?? and its clearly that HD 4870 will drop it performance if it have lower bandwidth like HD 4850, but how come HD 5770 have equal and sometimes have higher performance than HD 4870 ???

 

[IlllI]

Yeah I remember that launch and Rollo getting banned from Anandtech lol

starting to wonder if bella isnt the reincarnation of rollo or something

 

[Arzachel]

then how about HD 5770 vs HD 4870 ?? or HD 7770 vs HD 6870 ?? and its clearly that HD 4870 will drop it performance if it have lower bandwidth like HD 4850, but how come HD 5770 have equal and sometimes have higher performance than HD 4870 ???

You proved his point: less memory bandwidth doesn't always mean lower performance.

Anyways, I'd say the top-end GK104 will either perform between the 7950 and 7970 or be clocked to slightly beat/tie the 7970 resulting in a high TDP and little overclocking headroom.

 

[Keysplayr]

You are comparing a completely different architecture and a vastly more powerful chip in general in Fermi, it's simply not a valid comparison. A GTX 285 could have 10 times the bandwidth and it simply would not able to utilize it. GCN and Kepler are by all accounts very similar, so the bandwidth comparison is much more valid IMO.

We'll, fortunately for you, in your arguments against, it seems you're answering your own questions.

And what's with these questions? "How much more" as in how much space the memory controller in Tahiti takes up.

And then you follow with a comment like:

"I don't think the memory controller takes up much of a transistor budget at all"

So you know? Or don't you? I mean, why ask if you know?

 

[taltamir]

They did say it was their best chip ever, so you might be onto something

did nvidia ever release a GPU inferior to its predecessor? Because if not then every single one is their best ever (best =! fastest, they of course have low cost chips which are weaker then high end ones; or lower power chips, or near same performance for half price, etc)

 

[Silverforce11]

did nvidia ever release a GPU inferior to its predecessor? Because if not then every single one is their best ever (best =! fastest, they of course have low cost chips which are weaker then high end ones; or lower power chips, or near same performance for half price, etc)

This is too logical, it does not belong on the VCG forum.

 

[railven]

did nvidia ever release a GPU inferior to its predecessor? Because if not then every single one is their best ever (best =! fastest, they of course have low cost chips which are weaker then high end ones; or lower power chips, or near same performance for half price, etc)

Smiley face, figured that would imply that I was just kidding.

Next year they'll release "their best GPU ever." And so forth and so forth.

 

[tviceman]

then how about HD 5770 vs HD 4870 ?? or HD 7770 vs HD 6870 ?? and its clearly that HD 4870 will drop it performance if it have lower bandwidth like HD 4850, but how come HD 5770 have equal and sometimes have higher performance than HD 4870 ???

Good job you just backed up my point and made my argument even more valid.

 

[KCfromNC]

On March 23, NVIDIA Will Launch GK104 Kepler GTX 600 Cards



Saw this date mentioned at a few forums today, it's a Friday.

Didn't nV also push the Fermi release late on a Friday afternoon? Interesting.

 

[railven]

Didn't nV also push the Fermi release late on a Friday afternoon? Interesting.

Date's marked on me calendar. Since my Microcenter is gaga for nVidia, I hope they have cards on shelves day of launch. If the bench marks are favorable I'd pick one up (or just get a GTX 580.)

WTB leaked benchmarks!

 

[Leadbox]

Quote:
Originally Posted by notty22
On March 23, NVIDIA Will Launch GK104 Kepler GTX 600 Cards



Saw this date mentioned at a few forums today, it's a Friday.
Didn't nV also push the Fermi release late on a Friday afternoon? Interesting.

A little under 3 weeks from now. so I'm guessing the shiny new hardware is in the hands of reviewers or is arriving this week. Would mentioning that they have the hardware be in breach of NDA?