SambaNova RDU (Reconfigurable Dataflow Unit)
Realizes: AI training and inference dataflow graphs
Reconfigurable Dataflow Units implement granular dataflow graphs by combining configurable tiles with per-tile scheduling and streaming data paths. Each tile bundles compute arrays, SRAM buffers, and intra-tile routers, enabling DataScale to keep thousands of MACs busy while mapping compiled tensor graphs to the dataflow fabric.
Examples
🔗
SambaNova DataScale RDU deployment
SambaNova DataScale orchestrates racks of RDUs, compiling large language models into streaming dataflow programs and pushing tensors through configurable tiles with fine-grained scheduling and pipelined memory access.
[MATRIX_MULTIPLY,AI_INFERENCE]
high-throughput
data-center accelerator
≈65 pJ per multiply-accumulate