Radeon HD 7000:  HD 7790 |  HD 7850 |  HD 7870 |  HD 7870 XT |  HD 7950 |  HD 7970 |  HD 7990 |  Radeon HD 8000:  HD 8750 |  HD 8770 |  HD 8850 |  HD 8870 |  HD 8950 |  HD 8970 |  HD 8990 
GeForce 600:  GTX 650 Ti Boost  |  GTX 660  |  GTX 660 Ti  |  GTX 670  |  GTX 680 |  GTX 690 |  Titan: GTX TITAN |  GeForce 700:  GTX 760  |  GTX 760 Ti  |  GTX 770  |  GTX 780 |  GTX 790 

April 20th, 2012

NVIDIA To Present 7-Billion Transistor Kepler GPU during GTC 2012

NVIDIA To Present 7 Billion Transistor Kepler GPU during GTC 2012 geforce gtx 690 geforce gtx 685 geforce 600
NVIDIA will present a new chip from Kepler series during GTC 2012 Event in San Jose. GPU Technology Conference which will take place between 14-17th May, will focus on high performance computing, revealing latest breakthroughs and new technologies in GPU-enabled computing (sounds great doesn’t it?).

NVIDIA is reveling in brief information that Stephen Jones, CUDA Developer, and Lars Nyland, Senior Architect from NVIDIA, will present probably the most powerful GPU ever made. For know, the GPU which has the highest number of transistors is Radeon HD 7970′s Tahiti XT. Tahiti XT which holds 4.3-billion transistors has a die size of 356mm2. The latest card from NVIDIA – GeForce GTX 680 – based on GK104 holds 3.54-billion transistors, with a die size of 294mm2.

7-billion transistors would be twice the amount of transistors from GK104, so it’s basically an upcoming GTX 690 card, but in one chip. So what is it then? Kepler GK110 is not planned for this quarter and it wouldn’t feature so many transistors. We think this could be some engineering sample, which wouldn’t even be released as a GeForce graphics card. Is it Tesla? What do you think?

S0642 – Inside Kepler
Stephen Jones ( CUDA Developer, NVIDIA )
Lars Nyland ( Senior Architect, NVIDIA )

In this talk, individuals from the GPU architecture and CUDA software groups will dive into the features of the compute architecture for “Kepler” – NVIDIA’s new 7-billion transistor GPU. From the reorganized processing cores with new instructions and processing capabilities, to an improved memory system with faster atomic processing and low-overhead ECC, we will explore how the Kepler GPU achieves world leading performance and efficiency, and how it enables wholly new types of parallel problems to be solved.

Topic Areas: Parallel Programming Languages & Compilers
Session Level: Beginner

  • BestJinjo

    Since they are discussing compute functionality of Kepler, I am thinking it may be a preview of the GK110 for professional markets. Eventually we’ll probably see a derivative of this chip in the consumer space later in the year.

    The thing about GTX690 is it’s still based on 2x GK104. It’s not possible to magically fix the compute deficiency of GK104 since the way that chip was designed is an evolution of the GF104. The 1/24th SP performance in dual precision is not gimped by accident, but a design decision in how the SMX clusters are actually formed. Further, GK104 uses static not dynamic scheduling which to begin with means it will not be suitable for high computation workloads.

    This is why I would rule out GTX690 as the “7-billion GPU” for compute. Finally, it sounds like the memory subsystem is improved and processing cores are reorganized. That doesn’t smell like GTX690 either. It could mean a 384 or 512-bit monster chip we’ve been hearing rumors about for months. That would make sense. But if they don’t launch it for sale, then what’s the point?

    “From the reorganized processing cores with new instructions and processing capabilities, to an improved memory system”

  • http://videocardz.com VideoCardz

     I meant this would be a GPU with the same total amout of transistors like GTX 690. GK110 with 7-billion transitors? That would cost more than GTX 690.

  • Mufa_gwef

    They obviously need a better performing compute card for their workstation line (GK110) but 7-billion sounds impractical – what kind of cooling would it require, and the yields, which aren’t great on 28nm as it stands, would be abysmal. Financially and practically it doesn’t seem to make sense but I wouldn’t put it past Nvidia to do it anyway. I suppose it could be a one-off for demonstration purposes or its some kind of play on the 7 million dollar man!

  • BestJinjo

    It’s possible this card is meant for professional markets (i.e., Tesla). Right now with TSMC capacity constrained, NV could sell large die advanced compute chip for thousands of dollars for those markets. I think Telsa cards cost $4-5k! They could do that and wait until 28nm capacity picks up by Q3/Q4 and launch GTX685 in the consumer space. Who knows though. 7 Billion transistors sounds massive. That surely mean something like 520-530 mm^2 die. Right now I can’t see how its feasible to sell such a chip in the consumer space for less than $1000 given the 28nm capacity issues. 

  • BestJinjo

    GTX680 is a 195W TDP but it really operates at 170W. If they curb clocks to 900mhz and install a better cooler, they can make a much larger chip with 250-275W TDP. GTX280/285/480/580 were massive 520-576mm^2 die monsters! 

    They have room to play in terms of cooling and power. Cooling can be easily addressed by putting a better cooler. I agree with you it’s probably a demonstration of what they plan to launch sometime later, but not soon. 

  • http://videocardz.com VideoCardz

    Fair point, but those cards are usually released later than GeForce cards, so it’s quite confusing, as we still have only one Kepler ‘available’.

  • John Pombrio

    Whoa, I did NOT see this coming. I would imagine that this in NOT meant for any consumer market but as a uber model for massive parallel processing mojo for supercomputing applications. What else it would be for?

  • BestJinjo

    I know. If NV launches GK110 with 2304 SPs, 512-bit and so forth today though, that card will be 30-40% faster than HD7970. That would be almost unthinkable performance lead for a flagship. I suppose if they do limited run of these cards at $800 like 8800 GTX Ultra Extreme was, then it could make sense. 

  • The_empire007

    die size??

  • http://videocardz.com VideoCardz

     Around 500-550mm2