NVIDIA Kepler speculation thread

[fellix]

I guess the current FP+INT co-issue design is simply a cheaper way to get faster DP, without bloating the ALU size too much, and I think NV will stick to it. Also, integer and bit-wise throughput is becoming rather important side in todays and future GPGPU workloads -- making the INT ALU full-rate in Kepler will not only boost INT performance, but will likely be used for a speedier DP.

 

[Gipsel]

I guess that depiction of separate INT and FP ALUs is simply bogus. Just as an example, it doesn't make sense not to reuse the multiplier circuits for both.
They can't co-issue INT and FP instructions anyway.

 

[psurge]

Presumably that's why things like 24 bit integer multiplies are full speed (fits in the mantissa of a single precision float) and 32/64 bit integer multiplies are slower (at least, that's was the case on AMD HW). I would think that some things really are separate - bit ops like popcount, ro[rl] and so forth. But I would guess that compared to FMA, the transistor/power cost of most (all other?) ALU ops is small.

Actually - I do wonder how big the HW divider block is on modern x86s, and whether GPUs contain any dedicated HW for full precision correctly rounded divide/square root, or if those things are just software.

 

[3dilettante]

For GCN, divide is done via a macro.

There is hardware for approximating the reciprocal, square root, and combinations thereof for DP. I suppose it would apply to SP as well.
Fermi also has SFU hardware for those.

 

[psurge]

I was curious and looked into this a bit. It turns out that including HW for a divide is probably stupid if you've got FMA, because then you can use Newton-Raphson to produce a correctly rounded result. It's a quadratically convergent method, and so should be able to at least compete with HW implementations that apparently compute some fixed number of quotient bits per clock. [In case anyone cares, details are at http://perso.ens-lyon.fr/jean-michel.muller/ASAP2010-a.pdf

 

[trinibwoy]

http://vr-zone.com/articles/nvidia-kepler-to-arrive-in-march-or-april--hard-launch-only/14538.html

March now, no mention of which SKU.

 

[neliz]

http://vr-zone.com/articles/nvidia-kepler-to-arrive-in-march-or-april--hard-launch-only/14538.html

March now, no mention of which SKU.

Meh, optimistic.

Maybe not even the SKU everyone is talking about either...

 

[Ailuros]

http://vr-zone.com/articles/nvidia-kepler-to-arrive-in-march-or-april--hard-launch-only/14538.html

March now, no mention of which SKU.

From that link:

Will it be in the dying weeks or March or steal Ivy Bridge’s thunder with the launch around April 8... only time will tell.

There's a question mark missing at the end. Just because way too many are clueless it doesn't mean anything. Here another one to make it even more colourful:

http://www.fudzilla.com/graphics/item/25563-kepler-to-come-out-in-early-q2-2012

You can bet anything you want that it at least one desktop Kepler GPU will launch from February until December 2012. If needed you can also give mGPUs a new core codename.

​

 

[Sinistar]

Fudzilla also said Fermi would launch in December, but that article seems to be missing from his site now.

http://www.fudzilla.com/content/view/16185/34/

quote saved from that deleted article on Overclock.net

http://www.overclock.net/t/600033/fud-fermi-gf100-to-launch-until-early-december

 

[BRiT]

*ahem* All the completely off-topic noise has been deleted. Do not revive that topic in this thread.

 

[Man from Atlantis]

GK104

RE: GTX680 (not sure if this is the final model number)

May be seen as early as in February (not paper launch but retail), as Mr Huang doesn't like the HD7970 to shine most in the 28nm era.

The performance would be around the same as the HD7970, depending on the drivers.

So far the core clock is at 780MHz, with 2GB vram.

Source is from an AIC manufacturer, saying that even the retail packages are being printed now.

http://www.chiphell.com/thread-346223-1-1.html

 

[neliz]

Let's just say that that rumor isn't double-confirmed.

 

[DegustatoR]

Rumour: NVIDIA GeForce GTX 680 to release in February

Well this is fun.

 

[trinibwoy]

It sure is

 

[BRiT]

Rumour: NVIDIA GeForce GTX 680 to release in February

That's just repeating the ChipHell rumor. Is there any other source?

 

[DegustatoR]

Yeah, B3D is slow for me today for some reason so I haven't seen the 35th page until I've posted that VRZone link.

 

[trinibwoy]

Does chiphell have a good track record with these things? It's a nightmare trying to read the translated english on that forum! Any volunteers to summarize the gist of the conversation?

 

[Man from Atlantis]

and a mysterious 3dmark11 run..
translated
http://www.xtremesystems.org/forums/showthread.php?277763-%E2%80%9CKepler%E2%80%9D-Nvidia-GeForce-GTX-780&p=5036246&viewfull=1#post5036246

original source
http://www.chiphell.com/forum.php?mod=redirect&goto=findpost&ptid=346223&pid=10613945

when everyone is speaking you wont expect obr shut his mouth

You've all probably read (especially on Fudzilla) VGAs with graphics architecture Kepler is postponed to April, May etc. It's all just BS and Camouflage maneuver. Launch is literally in a few weeks/days, all this is just to appease the competitor AMD ... before crushing blow!

 

[mczak]

and a mysterious 3dmark11 run..
translated
http://www.xtremesystems.org/forums/showthread.php?277763-%E2%80%9CKepler%E2%80%9D-Nvidia-GeForce-GTX-780&p=5036246&viewfull=1#post5036246

That's a pretty strange score. GT2 and GT4 look good, but the overall score is so low that GT1 and GT3 would have to be no faster (or possibly even slower) than on GTX580.

And, with that quite low clock, would that mean hot clock is still there?
I think there's too much wild random speculation out there, so far nothing really trustworthy has materialized imho.

 

[gkar1]

*Fake exposed* http://forum.beyond3d.com/showpost.php?p=1612568&postcount=864