On March 22, 2012 Nvidia launched its latest graphic chips, codenamed GK104. This chip, powering the high-end GeForce GTX 680, proved to be a formidable contender for AMD cards.
However, while the newly launched GeForce GTX 680 is indisputably an high-end card, the GK104 chip itself seems a little “small” for Nvidia's standard: it weight at about 294 mm2 only. This, combined to the removal (eg: ECC memory protection), or attenuation (eg: very low FP64 rate) of certain GP-GPU specific features, leave the door opened to the future presentation of a new, bigger, GP-GPU-centric chip.
This article try to elaborate on GK104 die size and features to project some possible solutions for the bigger GK100 / GK110 processor. Remember that these are speculations only! Final GK100 / GK110 can be very different from what we expect.
UPDATE: it seems that Big Kepler chip's codename will be GK110 rather than GK100. I updated the article to reflect this.
Before entering the “grey area” of speculations and hypothesis, let's recap why the GK104 is not too well suited for GP-computing:
- no ECC memory protection
- very low (1/24) FP64 (dual precision) rate
- few load/store units in comparison to ALUs
- low L1 bandwidth per ALUs (0.33 Byte/FLOP for 32 bit operations, 0.66 Byte/FLOP for 64 bit operations).