NVIDIA Kepler speculation thread

[Jawed]

And there's absolutely no reason why NVidia couldn't have better performance per mm² or per Watt.

Also if this chip has "slow" double-precision (i.e. the hardware rate is so slow that it doesn't have to be artificially slowed for marketing purposes, because it's not bound for Tesla cards) then that's an easy adjunct to performance per mm² as Tahiti is "full performance" double precision in hardware. Same prolly goes for that ECC nonsense.

 

[xDxD]

http://www.geforce.com/News/article...-running-on-next-generation-nvidia-kepler-gpu

interesting information about using fxaa

 

[trinibwoy]

Its a 256-Bit 2GB performance card with only ~300mm² chip.

300? What happened to 320?

So a low power, small, fast chip with surround capability? It'll be quite a miracle if nVidia addressed all Fermi's major shortcomings in one shot.

 

[Jawed]

~200W is low power? 300mm² is small? Yes I suppose if you think those things then it's a miracle.

 

[trinibwoy]

Of course it's low power and small compared to Fermi and apparently Tahiti too if rumors are to be believed. How about we wait for reviews before starting inane arguments?

 

[Jawed]

Which Fermi? It's clearly not a high end chip.

 

[leoneazzurro]

I don't understand well, AMD lineup is not only Tahiti. Even if GK104 will end as fast as or slightly faster than the 7970 (which I personally don't believe to be true in all situations, especially in high resolution + AA where bandwidth is still quite important) Pitcairn is much better positioned for a price war (smaller die, probably better power consumption, lower performance than competition but still quite adequate), especially with "custom" versions. AMD could also release higher clocked/higher power versions of the 7970. It could be a situation like the GF110/Cayman, but with reversed roles. And in that case the die size difference was much bigger.
So I think that competition will be quite high anyway.

 

[Picao84]

Which Fermi? It's clearly not a high end chip.

It still is smaller than GF104/114, while alegely performing better than a GTX580 and not consuming that much more power than GF104/114.

So, I suppose Thaiti is also not a High End chip for you, although it does cost as much as one?

 

[Jawed]

It still is smaller than GF104/114, while alegely performing better than a GTX580 and not consuming that much more power than GF104/114.

Hmm, why wouldn't the replacement for GF114 do this?

So, I suppose Thaiti is also not a High End chip for you, although it does cost as much as one?

I'm talking about NVidia's range, i.e. the expectation there's a faster chip of the same family due in the next few months. Tahiti is the high end chip in AMD's line-up.

Tahiti is clearly over-priced compared to its predecessor (Cayman) but blame GTX580's high price for that. GTX580 was always wildly over-priced, but over-pricing is only evil when AMD does it.

 

[trinibwoy]

Which Fermi? It's clearly not a high end chip.

Based on what? If it comes with high-end performance (as defined by Tahiti) and high-end pricing (also as defined by Tahiti) then it's a high-end chip.

 

[trinibwoy]

Tahiti is clearly over-priced compared to its predecessor (Cayman) but blame GTX580's high price for that. GTX580 was always wildly over-priced, but over-pricing is only evil when AMD does it.

Ok, fair enough. Let's say Tahiti is overpriced so GK104 has to earn its stripes against GF114. Current rumors put it at 70-90% faster than the 560 Ti in BF3 with a smaller die size....would that be impressive enough?

 

[AlexV]

NVIDIA needs to move to an obfuscated codename scheme, this numeral using stuff is too easy to abuse in making inferences. If they start naming their chips as differing sorts of cakes (lots of cake around the planet), their actual competitive positioning will be determined by their characteristics, nut by their codename. On a separate note, Kepler seems nice.

 

[Jawed]

Ok, fair enough. Let's say Tahiti is overpriced so GK104 has to earn its stripes against GF114. Current rumors put it at 70-90% faster than the 560 Ti in BF3 with a smaller die size....would that be impressive enough?

Yes that's definitely more impressive than the 45-50% gain we see with Pitcairn over Barts. I'll wait until the numbers arrive, to be sure.

 

[Picao84]

Hmm, why wouldn't the replacement for GF114 do this?

Why should it? We are talking about market segments after all, so it would have been normal if GF104/114 die size was kept, instead of shrinking.

I'm talking about NVidia's range, i.e. the expectation there's a faster chip of the same family due in the next few months. Tahiti is the high end chip in AMD's line-up.

Tahiti is clearly over-priced compared to its predecessor (Cayman) but blame GTX580's high price for that. GTX580 was always wildly over-priced, but over-pricing is only evil when AMD does it.

Since when does you or me define NVIDIA's range? Thats something dictated by performance and price. Also, time between die shrinks seem to be increasing, so its only normal for this to happen, since one node shrink must be used for at least two generations of products. When HD5870 was launched did you also say it was not a high end chip, because you were expecting Cayman?

I didnt say Tahiti is overpriced, I just talked about it because it has a die size which is not that much bigger than GK104, and is sold as a high end part, and I didnt see you accusing it of being just a mid-range chip with a high end price. No, instead you choose to justify it with GTX580 overpricing. See, I can do that too.

EDIT - Dont forget also that AMD looks to be supply constrained (as probably most of TSMC 28nm clients), so the GTX580 price might not be the only and definitive reason for 7970 price... neither for GK104 price for that matter.

 

[xDxD]

I don't understand well, AMD lineup is not only Tahiti. Even if GK104 will end as fast as or slightly faster than the 7970 (which I personally don't believe to be true in all situations, especially in high resolution + AA where bandwidth is still quite important) Pitcairn is much better positioned for a price war (smaller die, probably better power consumption, lower performance than competition but still quite adequate), especially with "custom" versions. AMD could also release higher clocked/higher power versions of the 7970. It could be a situation like the GF110/Cayman, but with reversed roles. And in that case the die size difference was much bigger.
So I think that competition will be quite high anyway.

mmmm...yes, Pitcairn and Cape Verde, and what about gk106/gk107? I think this time with the same die-size the nvidia gpus will be better the other party, at least that is derived from rumors.

 

[A1xLLcqAgt0qc2RyMz0y]

Based on what? If it comes with high-end performance (as defined by Tahiti)

There is the problem. You are basing your point around Tahiti. Nvidia will also be releasing a single massive sized GPU for the Tesla/HPC Market. That GPU will also be Nvidia's high-end and will crush Tahiti in performance. So GK104 is not Nvidia's high-end whereas AMD's Tahiti is.

and high-end pricing (also as defined by Tahiti) then it's a high-end chip.

To quote a famous US President, "There you go again". Nvidia will not price the GK104 at Tahiti's high price even if the performance is nearly the same. The high-end price for Nvidia will be for the GK110/112.

 

[Blazkowicz]

I agree with Picao84, hard to complain that cards are expensive when they're on the cutting edge process, are the first significant consumer product on that process, supply constrained and making all 40 nm GPUs obsolete. wait if you want them cheaper.

 

[Kaotik]

There is the problem. You are basing your point around Tahiti. Nvidia will also be releasing a single massive sized GPU for the Tesla/HPC Market. That GPU will also be Nvidia's high-end and will crush Tahiti in performance. So GK104 is not Nvidia's high-end whereas AMD's Tahiti is.


To quote a famous US President, "There you go again". Nvidia will not price the GK104 at Tahiti's high price even if the performance is nearly the same. The high-end price for Nvidia will be for the GK110/112.

You really think nVidia will price GK104 low just because GK110 is coming in something like 4-6 months from now (being positive)?
GK112 or whatever the dual GK104 is, will be priced higher than Tahitis of course, just like 7990 will be, and GK104 will be priced according to performance and competition pricing at similar performance, not some lala-land charity midrange price

 

[Ailuros]

AFAIK, the original source of information was Theo Valich's piece over at VR-Zone, all others are referencing him. And he was writing power planes. With an "e". If that makes it more believable is another question as the earlier news on heise.de he linked indeed mentioned the dynamic clock adjustment, but first on a far smaller scale (+7% max) and secondly heise.de didn't talk about power plan(e)s at all.

I won't insist on something that I don't know first hand, but it still sounds like bullshit. I'll combine your theory with mine and we'll see what happens in the end:

There's no hotclock anymore, but a geometry "cold clock" which can scale upwards according to demand to a higher degree and a core clock which represents the majority of units that can scale upwards to a more conservative degree. I'd even go as far and suspect that the geometry cold clock on desktop Keplers will be held deliberatly low to give Quadros an advantage.

 

[Silent_Buddha]

And there's absolutely no reason why NVidia couldn't have better performance per mm² or per Watt.

Also if this chip has "slow" double-precision (i.e. the hardware rate is so slow that it doesn't have to be artificially slowed for marketing purposes, because it's not bound for Tesla cards) then that's an easy adjunct to performance per mm² as Tahiti is "full performance" double precision in hardware. Same prolly goes for that ECC nonsense.

That's what I was thinking as well. If they removed virtually anything related to HPC that doesn't directly impact consumer graphics performance then in theory they could make a small and fast GPU.

The flip side of that is, where does that leave their traditional enthusiast class consumer oriented chip?

It could be...

[1] As Carsten postulated before that GK100/GK110 could be just GK104/GK114 with HPC bits added. Which leads to my thinking that...

[2] GK104/GK114 may be positioned as the top end enthusiast level consumer chip.

In other words, GK100/110 may end up being mostly used for the HPC/professional/pro-sumer markets with perhaps a limited edition enthusiast consumer model at a ridiculous price for ~10% more perf. than a GK104/110.

If this is the case (and I'm not saying that it is), that would put AMD into a bit of a bind. As now they'll be the ones with the oversized die (Tahiti [big and full featured] compared to GK104 [small and less full featured]) in the consumer space while Nvidia takes up AMD's former position in the consumer space with smaller but less capable dies (GF100 [big and full featured] compared to Cypress [small and less full featured]). And by full featured, I mean HPC/professional market oriented features and specifications.

Anyway, that's just my mind running through the possibilities. Especially when you hear people on forums complaining about why a consumer GPU needs all these HPC/professional features. Perhaps, Nvidia listened to that and is making a bit of a break with anything sub 100/110 being for the consumer market with many things cut out or greatly reduced while anything 100/110 will be mainly meant for the HPC/professional/pro-sumer market.

Regards,
SB