bu-icsg / dana

Dynamically Allocated Neural Network Accelerator for the RISC-V Rocket Microprocessor in Chisel
Other
207 stars 36 forks source link

Support for configurations larger than the cache #8

Open seldridge opened 8 years ago

seldridge commented 8 years ago

Add support for paging in of data into the cache in the event that an NN configuration is too large to fit in cache. While a straight "get from memory the data you need" approach would work, a better approach would be to page in the weights as needed since the weights are the asymptotically dominant component of the NN configuration.