AnarchX
2010-01-11, 21:03:05
http://www.freepatentsonline.com/7616206.pdf
Patent vom November 2009, was beschreibt wie ein Teil des Speicherinterfaces für einen High-Speed-Interconnect zwischen zwei GPU genutzt werden kann. Dadurch sollte wohl dann besser skalierendes SFR möglich sein.
Das erinnert doch etwas an den Crossfire-Sideport des RV770 (http://www.forum-3dcenter.org/vbulletin/showpost.php?p=6656470&postcount=1), dessen Bandbreite mit 5GB/s bi-direktional deutlich zu gering war.
Nimmt man nun eine 384-Bit GPU an, die in der Dual-GPU-Konfiguration 128-Bit für den HS-IC nutzt, hat man hier bei einem Speichertakt von 2GHz schon 64GB/s.
Bleibt nun die Frage, ob man sich diesen Ansatz einfach nur sichern wollte oder ob er schon in kommenden GPUs (nächsten 2 Jahre) angewandt wird.
Interessante Auszüge:
Although the performance of the private bus compares favorably to that of a PCIE bus, 25 GB/s may still be somewhat limiting for certain graphics applications such as video games or animation programs. Therefore, several features have been added to the invention which further improve the effective bandwidth of the private bus or reduce the memory traffic between the two GPUs to increase the overall performance of the two GPU system. These features include render target interleaving; render-to-texture data duplication; data compression; variable length packet protocol; using non-data signals for transmitting data; storing all vertex data, geometry data and geometry-related push buffer commands in the frame buffer of each GPU; and utilizing each GPU to perform all geometry processing. The invention and each of these features are described in the following paragraphs.
One advantage of the disclosed graphics system 200 is that its cost may be approximately 87.5% of the cost of the graphics system 100 , due to the smaller frame buffers 202 , 204 relative to the frame buffers 104 , 108 , while its performance may be between 135% and 150% of the performance of the graphics system 100 , based on the aforementioned features of the graphics system 200 . Thus, the cost-performance ratio of graphics system 200 may be much more favorable than that of similar multi-GPU SLI systems, such as graphics system 100 . Another advantage of the disclosed graphics system 200 is that it is more scalable that an SLI system because the graphics bandwidth through the PCIE bus has been greatly reduced through use of private bus communications between pairs of GPUs. This architecture enables two GPUs to behave as one faster GPU, allowing the use of split-frame rendering and its problems to be avoided.
Patent vom November 2009, was beschreibt wie ein Teil des Speicherinterfaces für einen High-Speed-Interconnect zwischen zwei GPU genutzt werden kann. Dadurch sollte wohl dann besser skalierendes SFR möglich sein.
Das erinnert doch etwas an den Crossfire-Sideport des RV770 (http://www.forum-3dcenter.org/vbulletin/showpost.php?p=6656470&postcount=1), dessen Bandbreite mit 5GB/s bi-direktional deutlich zu gering war.
Nimmt man nun eine 384-Bit GPU an, die in der Dual-GPU-Konfiguration 128-Bit für den HS-IC nutzt, hat man hier bei einem Speichertakt von 2GHz schon 64GB/s.
Bleibt nun die Frage, ob man sich diesen Ansatz einfach nur sichern wollte oder ob er schon in kommenden GPUs (nächsten 2 Jahre) angewandt wird.
Interessante Auszüge:
Although the performance of the private bus compares favorably to that of a PCIE bus, 25 GB/s may still be somewhat limiting for certain graphics applications such as video games or animation programs. Therefore, several features have been added to the invention which further improve the effective bandwidth of the private bus or reduce the memory traffic between the two GPUs to increase the overall performance of the two GPU system. These features include render target interleaving; render-to-texture data duplication; data compression; variable length packet protocol; using non-data signals for transmitting data; storing all vertex data, geometry data and geometry-related push buffer commands in the frame buffer of each GPU; and utilizing each GPU to perform all geometry processing. The invention and each of these features are described in the following paragraphs.
One advantage of the disclosed graphics system 200 is that its cost may be approximately 87.5% of the cost of the graphics system 100 , due to the smaller frame buffers 202 , 204 relative to the frame buffers 104 , 108 , while its performance may be between 135% and 150% of the performance of the graphics system 100 , based on the aforementioned features of the graphics system 200 . Thus, the cost-performance ratio of graphics system 200 may be much more favorable than that of similar multi-GPU SLI systems, such as graphics system 100 . Another advantage of the disclosed graphics system 200 is that it is more scalable that an SLI system because the graphics bandwidth through the PCIE bus has been greatly reduced through use of private bus communications between pairs of GPUs. This architecture enables two GPUs to behave as one faster GPU, allowing the use of split-frame rendering and its problems to be avoided.