A real monster seems to be taking shape at NVIDIA. The company's next big graphics processor looks like a leap ahead of anything current-generation, the way G80 was when it released. It is already learned that the GPU will use a new MIMD (multiple instructions multiple data) mechanism for its highly parallel computing, which will be physically handled by not 384, but 512 shader processors. The count is a 112.5% increase over that of the existing GT200, which has 240.
NVIDIA has reportedly upped the SP count per cluster to 32, against 24 for the current architecture, and a cluster count of 16 (16 x 32 = 512). Also in place, will be 8 texture memory units (TMUs) per cluster, so 128 in all. What exactly makes the GT300 a leap is not only the fact that there is a serious increase in parallelism, but also an elementary change in the way a shader processor handles data and instructions, in theory, a more efficient way of doing it with MIMD. The new GPU will be DirectX 11 compliant, and be built on the 40 nm manufacturing process. We are yet to learn more about its memory subsystem. The GPU is expected to be released in Q4 2009.
NVIDIA has reportedly upped the SP count per cluster to 32, against 24 for the current architecture, and a cluster count of 16 (16 x 32 = 512). Also in place, will be 8 texture memory units (TMUs) per cluster, so 128 in all. What exactly makes the GT300 a leap is not only the fact that there is a serious increase in parallelism, but also an elementary change in the way a shader processor handles data and instructions, in theory, a more efficient way of doing it with MIMD. The new GPU will be DirectX 11 compliant, and be built on the 40 nm manufacturing process. We are yet to learn more about its memory subsystem. The GPU is expected to be released in Q4 2009.
