Accelerating distributed reinforcement learning with in-switch computing

A programmable switch includes an input arbiter to analyze packet headers of incoming packets and determine which of the incoming packets are part of gradient vectors received from worker computing devices that are performing reinforcement learning. The programmable switch also includes an accelerator coupled to the input arbiter, the accelerator to: receive the incoming packets from the input arbiter; asynchronously aggregate gradient values of the incoming packets, as the gradient values are received, to generate an aggregated data packet associated with a gradient segment of the gradient vectors; and transfer the aggregated data packet to the input arbiter to be transmitted to the worker computing devices, which are to update local weights based on the aggregated data packet.