NVIDIA GPUDirect 2.0

NVIDIA GPUDirect is designed to reduce the bottleneck that the CPU poses by eliminating unnecessasry system memory copies and CPU usage. Third party devices, including another graphics card, can also directly access the GPU memory.

CUDA 4.0 will introduce a new, improved GPUDirect. Its most important improvement, as far as performance is concerned, is that it permits peer-to-peer data transfers between the GPUs, bypassing the system memory and CPU.

Before GPUDirect 2.0, data that needs to be transferred from GPU2's memory to GPU1's memory must first be copied to system memory.

With GPUDirect 2.0, the data can now be transferred directly from GPU2's memory to GPU1's memory via the PCI Express interconnect. It can also support load and store, where GPU1 can pull and use data directly from GPU2's memory.

In both cases, the system memory and CPU are bypassed, thereby eliminating their associated latencies and bandwidth limitations. In addition, the traffic on the PCI Express interconnect is reduced by half.

Support Tech ARP!

If you like our work, you can help support out work by visiting our sponsors, participate in the Tech ARP Forums, or even donate to our fund. Any help you can render is greatly appreciated!