Abstract: Apparatus, systems, methods, and articles of manufacture to manage power of deep learning accelerator systems are disclosed. An example apparatus includes a power manager and a power controller. The power manager is to generate a power table to allocate power frequencies between an accelerator and memory based on a ratio of compute tasks and bandwidth tasks in a first workload; update the power table based on a request to at least one of add a second workload or remove the first workload; and determine an index into the power table. The power controller is to determine a power consumption based on the power table; determine whether to update the index based on a power budget and the power consumption; and allocate power to the accelerator and the memory according to the power table.

Abstract: Apparatus, systems, methods, and articles of manufacture to manage power of deep learning accelerator systems are disclosed. An example apparatus includes a power manager and a power controller. The power manager is to generate a power table to allocate power frequencies between an accelerator and memory based on a ratio of compute tasks and bandwidth tasks in a first workload; update the power table based on a request to at least one of add a second workload or remove the first workload; and determine an index into the power table. The power controller is to determine a power consumption based on the power table; determine whether to update the index based on a power budget and the power consumption; and allocate power to the accelerator and the memory according to the power table.