Method of and system for time-division based parallelization of graphics processing units (GPUs) employing a hardware hub with router interfaced between the CPU and the GPUs for the transfer of geometric data and graphics commands and rendered pixel data within the system
A method of dynamic load-balancing in a PC-based computing system employing a multiple GPU-based graphics pipeline architecture supporting multiple modes of GPU parallelization. During the execution of the graphics application, the stream of geometrical data and said graphics commands is analyzed, and the mode of parallelization of the GPUs during each frame, is determined using results of the analysis of the stream of geometrical data and graphics commands, and one or more policies for determining the mode of parallelization. The stream of geometrical data and graphic commands is distributed to the GPUs according to the determined mode of parallelization. During the generation of each frame, one or more of GPUs are used to process the stream of geometrical data and graphic commands, or a portion thereof, while operating in the parallelization mode, so as to generate pixel data corresponding to at least a portion of an image of 3D object. The pixel data output is transferred from one or more of the GPUs and composing a frame of pixel data, representative of the image of the 3D object. The frame of pixel data is displayed on a display surface of the PC-based computing system.
1. A multi-GPU graphics processing subsystem to work with a computing system having a CPU, said multi-GPU graphics processing subsystem comprising:
a plurality of GPUs divided into N sets of GPUs, each GPU taking N frame time units to render a full frame of graphics data, one of said GPUs is a display-designated GPU that is connectable to a display and the remaining GPUs are non-display-designated GPUs; and
a hardware hub interfaced between said CPU and said plurality of GPUs, wherein said hub has a router to alternately direct an incoming frame of a sequence of frames of geometric data and graphics commands to a different one of said N sets for rendering of said frames into image pixel data and to alternately redirect rendered frames to said display-designated GPU for display.
2. The multi-GPU graphics subsystem of claim 1 , wherein the hardware hub comprises a separate chip.
3. The multi-GPU graphics subsystem of claim 2 , wherein said hardware hub is interfaced with the computer system by a PCI express bus.
4. The multi-GPU graphics subsystem of claim 1 , wherein the hardware hub comprises a part of an I/O chipset.
5. The multi-GPU graphics subsystem of claim 4 , wherein said hardware hub is interfaced with a CPU chip by a proprietary bus.
6. The multi-GPU graphics subsystem of claim 4 , wherein said hardware hub is interfaced with said plurality of GPUs by a PCI express bus.
7. The multi-GPU graphics subsystem of claim 4 , wherein said hardware hub is interfaced with said plurality of GPUs by a PCI express bus.
8. The multi-GPU graphics subsystem of claim 1 , wherein said hardware hub is part of a CPU chip.
9. The multi-GPU graphics subsystem of claim 8 , wherein said hardware hub is interfaced with said plurality of GPUs by a PCI express bus.
10. A method of processing data within a multi-GPU graphics subsystem working with a computing system having a CPU, said method comprising the steps of:
(a) providing a plurality of GPUs divided into N sets of GPUs, each GPU taking N frame time units to render a full frame of graphics data, wherein one of said GPUs is a display-designated GPU that is connectable to a display and the remaining GPUs are non-display-designated GPUs; and
(b) alternately directing an incoming frame of a sequence of frames of geometric data and graphics commands to a different one of said N sets for rendering of said frames into image pixel data and alternately redirecting rendered frames
to said display-designated GPU for displaying.