Kepler Architecture to Support 7-Billion Transistor Chip?
We just saw the Kepler GPU architecture revealed in the recent NVIDIA GeForce GTX 680 graphics card launch that positioned it as the new flagship GPU. True enough, our findings also point to the new graphics card coming out tops in most scenarios with its clever GPU Boost algorithms in place to maximize the GPU's efficiency and utilization.
However, we weren't quite right when we thought the GPU Kepler architecture was fully revealed. Apart from the internet rumors of a more powerful NVIDIA GPU, here's something direct from a session synopsis of NVIDIA's upcoming Graphics Technology Conference 2012 in the USA:-
While NVIDIA's GPU managers tried their best to conceal and not comment of any future products, the internet was abuzz with speculations on the true heir of NVIDIA's Kepler architecture based line-up. These are the giveaway pointers we noted from our perspective:-
- GeForce GTX 680's GPU core code is GK104. The numerals normally indicate that the GPU is just a 'mid-range' contender
- GK110 / GK100 has long been the rumored flagship core code for the Kepler series
- 195W TDP of the GeForce GTX 680 is yet another telling sign even though it could match and surpass 250W TDP cards from AMD (Radeon HD 7970)
- Only 256-bit wide graphics memory interface - again a trait of mid-range GPUs (which luckily for the GTX 680 is somewhat sufficient)
- Spare space for extra graphics Molex power connector
Further to this, we've spoken to some of the industry folks and they too have indicated something more powerful underway. Now with the above screen shot direct from NVIDIA, it is certain that there's a more powerful Kepler GPU later this year. While NVIDIA's Graphics Technology Conference is a high-level conference session where scientists, professors, GPU and software engineers come together yearly to share their progress and ways to optimize processes for GPU compute and propel computing to the next level, the GPUs that would be featured are those of the Quadro and Tesla series. Does this also signal that the GK110 GPU is more suited for GPU compute rather than gaming?
With a GPU boasting 7 billion transistors on the current 28nm silicon fabrication process technology, you can guess that it's going to be a very expensive GPU with its massive die size and current yield rate. Since gaming needs are taken care by the GeForce GTX 680 fairly well, NVIDIA could focus on the more lucrative workstation/server market segment and offer the GK110 GPUs to contribute in more serious engineering and research tasks - until at least it becomes more cost effective for the consumer market.
So far estimates for the GK110 has been reported to have 6 billion transistors, up to 2500 CUDA cores, 512-bit memory interface and a TDP that can range between 250W and 300W. Given the renewed confirmation of the transistor count at 7 billion, these stats sound totally plausible. This however brings up an interesting question if the die shot of Kepler that has been shared to-date belongs to the the GK104 or the final GK110. Stay tuned as we find out more in the days ahead.