Whitepapers

Generative-AI workloads evolve on a quarterly cadence, with novel attention variants, quantization formats, and routing schemes appearing faster than fixed-function neural-processor silicon can be respun. This paper presents the General-Purpose Neural Processor (GPNPU): a single programmable processor architecture, delivered as configurable IP, that unifies scalar, vector, and matrix execution in one instruction-set architecture (ISA) and one programming model accessible from standard C++ and Python.

A graph compiled to a fixed sequence length spends the same work on every prompt. On a Chimera GPNPU, prompt length is a value the running program carries, so one binary serves every length up to the capacity you set, and the work follows the prompt.

SIMT earns its generality by deferring decisions to run time. Edge inference is compiled ahead of execution, so that machinery draws power to absorb an uncertainty the workload does not contain. A Chimera core removes the machinery and keeps the generality.

On a machine built from a neural accelerator, a vector DSP and a host CPU, deciding where to cut the graph costs you on every inference for the life of the product. On a Chimera core there is nothing to cut.