We decided to start Monday on a different scale - your Ed-in-Chief woke up at 2AM European time, just in time for Asia to wake up... and we decided to hear all that we can from interested parties. Information we received was quite interesting - companies know that nVidia has the samples of NV70/G300/GT300 silicon, but they were not delivered through usual sources - which would explain the lack of pictures coming from Chinese websites.
As usual, nVidia is keeping their partners in the dark, but there is a certain power struggle between the companies, with questions flying left and right - in fact, we heard partners repeating the information that originated from our site. At recent annual shin-ding for partners, nVidia was cocky and claimed that the company call help the partners out of the crisis. The GT300 just might be the reason for over-confidence typical when nVidia has a strong part coming out.
First internal cards are based on A1-stepping and feature 1024MB of GDDR5 memory, with all the memory chips located on front of the card. There are also cards with 2048MB, but we were unable to confirm the codename for the 2GB board. Oddly enough, the 1GB card has a codename fitting nicely in the current line-up, even behind one upcoming dual-GPU single-PC part, which would imply that nVidia has the cards for several weeks now.
According to the German site Hardware-Infos, the sample boards now come with a higher GPU clock. The alleged GT300 A1 silicon ticks at 0.7 GHz, with 512 cores working at 1.6 GHz. GDDR5 memory was upped to 1.1 GHz QDR or 4.4 GT/s [billion transfers per second], resulting in memory bandwidth of massive 281.6 GB/s. According to information we had, nVidia firstly clocked the at 1:1 ratio with GeForce GTX280 for a baseline comparison. If the leaked informationis true and those clocks remain final, that would mean GTX380 missed it targeted clock by a double-digit percentage. Still, a beast nevertheless.
If the figure of 1.6 GHz for 512 shaders is final, then the chip can achieve 2457 GFLOPS, e.g. 2.46 TFLOPS. This is a very good jump from 933 GFLOPS on the original GTX280. But the dark horse here is not the single-precision score, but rather the efficiency of dual-precision format. If our sources were true, GT300 chip has radically improved way of calculating dual precision operations and should bring anywhere between 7-10x improvement in performance. Given that ATI's Radeon 4800 series destroyed GT200 series in dual-precision, we are not surprised to hear about large improvements in DP performance.
