vacancies advertise contact news tip The Vault
facebook rss twitter

NVIDIA Kepler CUDA make-over

by Alistair Lowe on 10 February 2012, 11:35

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qabcjb

Add to My Vault: x

Details of NVIDIA's Kepler architecture still remain scarce, with many rumours considered to be propaganda and heavily unsubstantiated. Amongst these rumours are suggested release specifications, dates and prices for the Kepler line-up, released by LenzFire.

Rumoured Kepler Details

Rumoured Kepler Prices

With no sources cited, we suspect these figures to be complete conjecture but is it possible to glean elements of truth from these tables? Certainly, press feedback from CES suggested a smug feeling of superiority from NVIDIA when asked about its concerns over AMD's first-to-goal Radeon HD 7xxx series, leaving us to expect much from Kepler, especially given its delayed release.

TechPowerUp has claimed that its information does in fact stem from "reliable sources" and suggests that Kepler's revamped CUDA architecture will focus on improved parallel performance, with an extra, and presumably cached layer at the top of the hierarchy, dubbed the Graphics Processing Cluster (GPC), with each cluster enclosing four Streaming Multiprocessor (SM) cores, with each SM containing 96 Stream Processors (SP).

This is in contrast to Fermi where SMs are not clustered, with the GTX 580 containing one third of the SPs per SM. This massive increase in Stream Processors follows somewhat with trend, as later dies in the 5xx series such as the GF114, GF116 and GF118 featured 48 SPs per SM. With TechPowerUp suggesting that there will be four clusters in the GK104-based 6xx cards, this would result in 1,536 CUDA cores, a three fold increase over the current generation. What's more is the site claims that the new design will feature twice the number of TMUs, with eight per SM.

NVIDIA Kepler Shader Design

All this points to an incredibly large die, with the added heat lowering possible clocks, with the site suggesting that CUDA cores will no longer function inside their own clock domain and will tick over at the same rate as the rest of the card, which looks to be 950MHz. If TechPowerUp's claims are legit, we'll soon be seeing cards with 2.9 TFLOP/s of single-precision floating point compute power and 486 GFLOP/s at double-precision.

The site's claims fail miserably to match up with the above figures, though TechPowerUp's claims of a 340 mm² die, 32 ROPS, 16 SMs per cluster and ~950MHz clock do support figures for the GTX 660, with the mistake, perhaps, that it would feature four clusters, where perhaps this configuration could be how NVIDIA intends to accomplish its high-end GK110 implementation.

Only time will tell but we wish that time would come sooner, as the suspense is too much to bear.



HEXUS Forums :: 7 Comments

Login with Forum Account

Don't have an account? Register today!
sounds interesting,time to upgrade
Enormous gaming cubes ?
</pedant>
Phage
Enormous gaming cubes ?
</pedant>

Don't ask me how that happened… I really don't know :S Think I'm losing my marbles.
Auto spell check thingy ?
Brilliant! With the amount of wtfpwnage rumours im beginning to think that Nvidia have a flop ready :P.

If those specs are right then the 690 is simply two 670's, seems a bit silly as this time they shouldnt be held back by the TDP like last time :D. I like the performance comparisons in the second chart, 45% better than the 7970, it will interesting because the fact is the 7970 has so much untapped speed available, at 1100mhz core it increases performance insanely, i think thats why AMD havent released with high clocks, they are using this time to test the waters and see how to improve the process while Nvidia is taking their time to release kepler because by then if the 680 is better than the 7970 then AMD will simply launch 1100mhz+ core clocked 7970's and that would be the end of it :D.

One thing that seems baffling is the memory, sure the 690 is using the 670 but surely Nvidia should know that the raw power from two would benefit from a wider bus? I have a feeling its going to be a mars type price card!.