Kepler big core broke the news: 512bit 4GB memory

The slightly quieter Czech OBR recently brought us new news: The Kepler big-core GeForce cards that some users have been looking forward to will be listed in a month that ends with “ber”.

As for a more specific date, the OBR has already been clarified, not in November/December but in September/October, but this point seems to me not very reliable. Prior to this, the most optimistic Kepler core Tesla will be in August/September. Available, the first batch of GK110 will be used for professional cards.

As for the specific specifications of the GK110 core, OBR said that although the core area is indeed large, the GK110 is still smaller than the GF110. The integrated memory controller will also rise to 512bit with 4GB GDDR5 standard. OBR sources said that the performance of the large core Kepler is about 20-25% better than the GeForce GTX 680.

But Sumit Gupta, the head of NVIDIA Tesla’s GPU computing business group, once said that the Kepler architecture's second core (hereafter referred to as Kepler II) will add a lot of new computing features and much higher double-precision floating-point calculations than Kepler I, the GK104. ability.

HPCwire, a professional supercomputer news report website, speculated that Kepler II will have 1.3-1.5 TFlops of double-precision floating-point computing capability, which is roughly double that of the previous Ferris architecture Tesla M2090. The Kepler architecture Tesla card will still maintain a TDP of around 225W. The product may be released on paper on the GTC 2012 in May. As for actual production, according to Gupta's previous statement that it would wait until the end of the fourth quarter before the end of the year. From this point of view, the time-to-market portion of OBR’s sources is even higher than the most optimistic speculation.