NVIDIA Kepler speculation thread

[Mianca]

Then, I guess, we old-worlders are just lucky.

Supply is rather limited in Europe, too.

That overview you posted is kind of misleading, as "in stock" basically just means that at least one of the hundreds of stores within the database has at least one card in stock. In most cases, you'll be lucky to find a card actually available in one of the more reputable stores.

That's not to say you don't have a point about availability being better in Europe - but it's far from "plenty of stock" at the moment.

 

[UniversalTruth]

Galaxy 68NPH6DV5ZGX GeForce GTX 680 available.

http://www.newegg.com/Product/Produ...H&N=-1&isNodeId=1&Description=gtx+680&x=0&y=0

Or not so much.

http://www.newegg.com/Product/Product.aspx?Item=N82E16814162095

 

[NathansFortune]

OcUK have got stock in for any UK people who want in.

 

[Florin]

OcUK have got stock in for any UK people who want in.

Heh, the curious thing about discussions like these is that none of those who display concern about a lack of available cards is actually interested in buying one.

 

[UniversalTruth]

Heh, the curious thing about discussions like these is that none of those who display concern about a lack of available cards is actually interested in buying one.

Yup, because we are waiting for NV to flood the market, no one actually to buy AMD products, and those to decrease finally that ridiculous pricing.

BTW: GK104 is 100% faster than GF104. Ridiculous scaling.

Expect GK110 to be around 100% faster than GF100/110 too.

 

[Ailuros]

Compute performance yes.

Compute performance comes from ALUs and unless you'd have an insane for the time being 1:1 DP/SP ratio in order to have N GFLOPs DP, you'd need automaticall (assuming a 2:1 ratio) N*2 GFLOPs SP. TMUs are also relatively bound to clusters since Fermi.

Under that reasoning all you need to define is what N stands for the example above.

Given that GF114 was less deficient at compute than GK104 the 3D performance gap should narrow while the compute gap gets much wider. What that means in absolutes is anybody's guess.

A crapload of additional transistors and probably way more than the transistor difference between GF110 and GF114 (roughly say 35%). The problem now is that despite the differences for Fermi GF110 was still roughly 40% ahead in terms of 3D performance compared to 114. In other words that 25% whatever 3D performance difference between GK104 and GK110 is most likely the usual senseless bullshit.

 

[Kaotik]

Yup, because we are waiting for NV to flood the market, no one actually to buy AMD products, and those to decrease finally that ridiculous pricing.

BTW: GK104 is 100% faster than GF104. Ridiculous scaling.

If you're referring to GTX560 Ti by "GF104", GTX680 is only ~87-88% faster at 2560x1600 despite 1GB vs 2GB, at 1920x1200 the difference is only 67%

Expect GK110 to be around 100% faster than GF100/110 too.

With GTX680 scaling being 67% on resolution where memory rarely has an effect, why on earth would one expect 100% faster GK100 compared to GF110, especially when big part of GK104's speed comes from sacrificing GPGPU speed, while GK110 isn't doing the same sacrifices.

edit:
If you were being sarcastic, I blame english not being my native language making sarcasm so much harder to detect from written text

 

[UniversalTruth]

It's logical to expect I had in mind GTX 460. We should compare apples to apples, first generation 40 nm process vs first generation 28 nm. 560 Ti is the second iteration.

not to be understood.

 

[trinibwoy]

Compute performance comes from ALUs

Well that's obviously not true as proven by Fermi. You can have a ton of ALUs but if they're idling most of the time it does you no good.

A crapload of additional transistors and probably way more than the transistor difference between GF110 and GF114 (roughly say 35%). The problem now is that despite the differences for Fermi GF110 was still roughly 40% ahead in terms of 3D performance compared to 114. In other words that 25% whatever 3D performance difference between GK104 and GK110 is most likely the usual senseless bullshit.

The absolute numbers will be known eventually. All I'm saying is that the gap between compute and graphics performance will probably grow significantly from GF104vGF110 to GK104vGK110.

 

[Kaotik]

It's logical to expect I had in mind GTX 460. We should compare apples to apples, first generation 40 nm process vs first generation 28 nm. 560 Ti is the second iteration.

not to be understood.

We never had fully enabled GTX460, fully enabled one shouldn't be far from GTX560Ti

 

[Albuquerque]

We never had fully enabled GTX460, fully enabled one shouldn't be far from GTX560Ti

Indeed. It appears exactly like one whole 'partition' was enabled between the 460 and the 560Ti; ROPs stayed identical as did memory bus, but texture units and cores both moved by the exact same ratio (1/8th)

Card    ROPs    TexAddr   TexFilt   Cores     MemBus  CoreClock
460     32      56        56        336       256b    675Mhz
560Ti   32      64        64        384       256b    822Mhz

So, yeah, the 560Ti is exactly a 'full' GF104 plus a fat clock bump. Maybe the confusion is the 560Ti-448? That one isn't related to the 460/560 line...

 

[DuckThor Evil]

There is fairly "large" die size difference between gf104 and 114 so they've done some tweaking to it, other than just enabled the additional 48 cuda cores etc.

 

[Ailuros]

Well that's obviously not true as proven by Fermi. You can have a ton of ALUs but if they're idling most of the time it does you no good.

Fermi also lacks severely in bandwidth for HPC.

The absolute numbers will be known eventually. All I'm saying is that the gap between compute and graphics performance will probably grow significantly from GF104vGF110 to GK104vGK110.

I'm only reacting to the nonsense that GK110 will be about 25% faster than GK104. If you take that under account compute performance cannot obviously grow signficantly compared to Fermi unless as I said they go for 1:1 DP/SP.

 

[trinibwoy]

I think you're underestimating the significance of GK104's static scheduling and dual-issue Ail. Both conspire to reduce throughput and if the big boy sheds them it could mean better compute performance without a big bump in theoreticals. The DP:SP ratio only matters for DP performance which as we know is pretty much non-existent on GK104 anyway.

 

[Albuquerque]

There is fairly "large" die size difference between gf104 and 114 so they've done some tweaking to it, other than just enabled the additional 48 cuda cores etc.

Dunno if 30mm^2 is "large"; in a relative sense, it's ~10%. From what I'm able to tell, a lot of the die size change between GF104 to GF114 was related to optimizing for higher speed without a radical change in power consumption (160w TDP vs 170w TDP, respectively) rather than truly any additional functionality. Further proof of the 'not much change in functionality' is reflected in the total reported transistor count being the same between both GF104 and GF114 - 1.95B transistors.

Anand has a page dedicated to talking about the physical transistor changes that NV put into the GF114 'remake' of GF104; specifically targeting certain key areas of the GPU to use equally specific transistor types that are more appropriate for the workload. See here: http://www.anandtech.com/show/4135/nvidias-geforce-gtx-560-ti-upsetting-the-250-market/2

I'm pretty sure that's where nearly all of the die size change came from.

 

[Mianca]

Expect GK110 to be around 100% faster than GF100/110 too.

~90-100% faster than GF100 sounds plausible.

~100% faster than GF110 seems a little too optimistic.

Should still be a great leap in performance. No question about that.

 

[iMacmatician]

There is fairly "large" die size difference between gf104 and 114 so they've done some tweaking to it, other than just enabled the additional 48 cuda cores etc.

What are the two die sizes?

 

[AlphaWolf]

332mm2 and 360mm2, I'd say that's a ways off of a large difference. It's under 9%

 

[LordEC911]

Which means that you actually count.
Pitcairn and Tahiti- 20-30% performance difference. I said higher for GK104 and GK110, you agreed- 35-40%.
I expect the usual optimisations here and there on the chip which will allow it to improve compared to what launched 6 months earlier.

20-30%? For Pitcairn - Tahiti? Wow, that is lowballing.

 

[DuckThor Evil]

I specifically used "" around large and "fairly" to imply that it was meant relatively. GF114 wasn't just a case of enabling something that was already there, but something else also. Nearly 10% difference is fairly large compared to 0%. 0% would only basically be a bios change like opening the shaders on the 6950.