AMD is finally splitting up the 4+1 shader cluster. Though a very useful feature when utilized, the final complex shader was rarely used in practical applications. Hence, out of the 1600 SP of the HD 5870, only about 1280 SP were in full utilization. HD 6000 cuts each shader cluster down to 4 shaders of medium complexity, rather than 4 simple + 1 complex. The end result is greater throughput as well as practical performance. In addition to this fundamental improvement, the NI uncore is also said to be improved. NI was designed for 32nm, however, and the cost of moving to 40nm means a large die size. SemiAccurate suggests 380-400 mm2 for Cayman / HD 6800 series. This still makes it much smaller than GF100's 528 mm2. The performance is expected to increase much more than the 10%-20% increase in die size over the HD 5800 series, however.
