NVIDIA Kepler speculation thread

[whitetiger]

For Desktop one part, actually. GTX 460 SE.


edit
Just for the record: GK104-400 and -335 are product numbers only, not unit counts.

In which case I believe it!
There will be 2 versions of the GK104!
A high-end part, and a lower-performing part at a lower cost!


I also predict nVidia will invest a horrendously obscure product naming scheme that will confuse and annoy the hell out of everyone!

 

[Man from Atlantis]

I think if the shader alus in a SM really are beefed up by a factor of two it would make sense if the SMs also had twice the shader export capabilities.

Agreed but according to Fermi numbers 32ROPs @5gbps should be heavly bottlenecked..

It could be one of Charlie's infamous ruses to winkle out copy cats
- 335 being an unlikely number for anything....
- even 400 is a strange number, of say, CCs.....
- but if 400 was correct, then 336 would be more logical
- but, doesn't call into the 1/8 scheme, ....

So probably both numbers are made up, somewhere along the line....

i think you're a bit confused.. for example GTX 460 336CC 256bit is named GF104-325 and 192 bit part GF104-325-KA/KB/KC/KD(depends on which memory controller is disabled) and GTX 460 SE(288CC 256 Bit) GF104-225.. GTX560Ti(384CC) GF114-400 etc

 

[mczak]

For Desktop one part, actually. GTX 460 SE.

Oh I forgot about that indeed. Still it was launched nearly a half year later (and close to GF114 based products) and I think it had quite low visibility (doubt it sold in large numbers). So yes maybe there were some chips which didn't have 7 working SMs, but it can't have been that bad.

 

[whitetiger]

Agreed but according to Fermi numbers 32ROPs @5gbps should be heavly bottlenecked..



i think you're a bit confused.. for example GTX 460 336CC 256bit is named GF104-325 and 192 bit part GF104-325-KA/KB/KC/KD(depends on which memory controller is disabled) and GTX 460 SE(288CC 256 Bit) GF104-225.. GTX560Ti(384CC) GF114-400 etc

Yes, it's just the part numbering scheme... sorry for taking your time!

 

[iMacmatician]

Just for the record: GK104-400 and -335 are product numbers only, not unit counts. GF100 had -275 and -375 suffixes for example.

Yeah, I was making a joke (if it didn't come across as such).

EDIT:

Does "group" have to be equivalent to a SM? Could the "7 group" just be referring to 7/8 of the total CCs active, even though the disabled CCs could be in various places not entirely dependent on the SMs (for example, 6 full SMs active plus half the CCs of each of the other two active).

(Charlie claims something along these lines later in his report.)

 

[itaru]

maybe
8 GPC or 7 GPC
I think that it resembles echelon
8 lane
4 SM
8 cluster

Yeah, I was making a joke (if it didn't come across as such).

EDIT:

Does "group" have to be equivalent to a SM? Could the "7 group" just be referring to 7/8 of the total CCs active, even though the disabled CCs could be in various places not entirely dependent on the SMs (for example, 6 full SMs active plus half the CCs of each of the other two active).

(Charlie claims something along these lines later in his report.)

 

[xDxD]

http://www.3dcenter.org/news/die-aktuellen-spezifikationen-zum-gk104-kepler-performance-chip

 

[CarstenS]

Yeah, I was making a joke (if it didn't come across as such).

No, I really didn't get it - sorry.

EDIT:

Does "group" have to be equivalent to a SM? Could the "7 group" just be referring to 7/8 of the total CCs active, even though the disabled CCs could be in various places not entirely dependent on the SMs (for example, 6 full SMs active plus half the CCs of each of the other two active).

(Charlie claims something along these lines later in his report.)

I am not sure if his wording is chosen in such an interpretable way deliberately or by accident. In any case, distributing Warps between different SMs seems like it would require quite a bit of overhead on the housekeeping side of things.

 

[Man from Atlantis]

S|A - GK104 pops up in the wild
Kepler is in the hands of the AIBs

The short story is that some places have been getting early variants, others later, and in various states of functionality. Since they are meant for early hardware design and testing, they are more than adequate for the task. The lack of polish seems to indicate that Nvidia is pretty hell-bent on getting cards out the door ASAP.

If things go as normal, it takes 4-6 weeks from AIB sampling to cards on the shelves. This would mean late March or early April,

 

[CarstenS]

WTF??? Charlie's back on NV-hatin'-track again?

S|A - GK104 pops up in the wild

[…]The lack of polish seems to indicate that Nvidia is pretty hell-bent on getting cards out the door ASAP.

Kepler is in the hands of the AIBs

[my bold]
vs.

GK104 cards seen by SemiAccurate all look very polished and complete, far from rough prototypes or “Puppies“.

[my bold]
http://semiaccurate.com/2012/01/19/nvidia-kepler-vs-amd-gcn-has-a-clear-winner/

 

[boxleitnerb]

Maybe his cheques from AMD cleared...

 

[Ailuros]

WTF??? Charlie's back on NV-hatin'-track again?

[my bold]
vs.

[my bold]
http://semiaccurate.com/2012/01/19/nvidia-kepler-vs-amd-gcn-has-a-clear-winner/

Or you could read into the text in a different manner:

The short story is that some places have been getting early variants, others later, and in various states of functionality. Since they are meant for early hardware design and testing, they are more than adequate for the task. The lack of polish seems to indicate that Nvidia is pretty hell-bent on getting cards out the door ASAP.

 

[CarstenS]

Still doesn't explain the discrepancy in polish. Maybe it wears off in time...

 

[AlexV]

Gentlemen, please, enough with the Charlie this Charlie that dance...can we not find anything else of interest do discuss beyond what one of the less interesting aspects of this industry has done/said/imagined? Whilst I do fully understand why "The young and the restless" had a trillion episodes and quite a following, I would very much like to think that we're in a company that's somewhat more favourable to a quality book...

 

[Ailuros]

Still doesn't explain the discrepancy in polish. Maybe it wears off in time...

I'd say when you're in a rush, you'd use anything you have available at the time to get the job done. And that's exactly how I understood the specific passage. I disagree probably more than others with Charlie's writeups, but in the given case I don't see anything wrong.

***edit: AlexV I apologize since we probably typed at the same time. CarstenS: via PM if you want.

 

[ninelven]

I'll change my guess to:

GK-104 (high)
4 GPC
12 SM/GPC
32 SP/SM
2 TU/SM

GK-104 (low)
2 GPC
16 SM/GPC
32 SP/SM
2 TU/SM

That should cover a whole range of possibilities, while still being very wrong.

 

[Ailuros]

Thank God I'm no professional and I wouldn't have to create any schedulers for those scenarios.

 

[trinibwoy]

Boards in the wild? Sweet, that means the bullshit rumors are nearing an end. Next stop - bullshit facts!

 

[Ailuros]

Boards in the wild? Sweet, that means the bullshit rumors are nearing an end. Next stop - bullshit facts!

Now where's the shotgun when you need it? You cook it if I hit one

 

[whitetiger]

S|A - GK104 pops up in the wild
Kepler is in the hands of the AIBs

Hopefully soon we'll get some board shots!

I'm intrigued to see if they've dropped the hot-clock
- if makes sense to that they would - in these days where everything is power limited, the hot-clock seems like a mistake architecturally