Is FuriosaAI's chip architecture truly innovative?

FuriosaAI's RNGD adopts an innovative TCP (Tensor Contraction Processor) architecture to simultaneously achieve performance, programmability, and power efficiency.

Existing Systolic Arrays are highly efficient for matrix operations, as data flows through a fixed grid structure. However, because the structure is rigid, resource utilization drops sharply when the size of the tensor being computed or the batch size does not match the fixed grid.

TCP was designed to solve this problem, with the following distinguishing features:

• Dynamic reconfiguration: Instead of a fixed grid, it uses small computational units called 'slices', which are dynamically reconfigured as needed to accommodate various tensor shapes.
• Efficient data movement: The Fetch network allows data to be broadcast simultaneously to multiple slices.
• Spatiotemporal data reuse: Beyond spatial reuse, it also supports data reuse in the temporal dimension by leveraging buffers and sequencing logic.

Thanks to this flexibility, TCP maintains high resource utilization and efficiency even in Inference workloads with small batch sizes and variable tensor sizes.