Network cache injection for coherent GPUs

A graphics processing unit (GPU) compute node comprising:

• a GPU comprising a GPU cache;
• a network interface controller (NIC) configured to receive data for processing on the GPU;
• the NIC further configured to send the data to a main memory of the GPU compute node;
• the NIC further configured to send coherence information to a coherence directory of the GPU compute node based on the data, the coherence information indicating that a cache entry of the GPU cache is invalid;
• the GPU configured to receive the coherence information;
• the GPU further configured to determine, based on the coherence information, whether the data includes a network header or a command packet;
• the GPU further configured to load the data into the GPU cache from the main memory of the GPU compute node responsive to the data being determined to include the network header or the command packet.

A method for inputting memory from a network interface controller (NIC) of a graphics processing unit (GPU) compute node to a cache of a GPU of the GPU compute node, the method comprising:

• receiving, by the NIC, data for processing on the GPU;
• sending, by the NIC, the data to a main memory of the GPU compute node;
• determining, by the NIC, whether the data includes a network header or a command packet;
• responsive to the data being determined to include the network header or the command packet, sending coherence information to a coherence directory of the compute node based on the data, the coherence information indicating that a cache entry of the cache is invalid;
• and
• responsive to the coherence information, writing the data into the cache from the main memory of the GPU compute node.