Proton NPUCerebral ChipsEvery machine should think.
Block 01 / instruction ownership

64-bit RISC-V core

The CPU that executes the program, controls scalar memory accesses, and coordinates vector and matrix work.

1 coreRV64Six-stage pipeline

The main processor

The scalar core owns the program counter, scalar registers, control flow, exceptions and retirement. The vector unit and matrix engine receive instructions from this CPU; neither runs a second software program.

RISC-V scalar pipeline and tagged accelerator dispatch
A conceptual pipeline view. Accelerator completion can take multiple cycles; the first version deliberately serializes matrix commands. Open full-size diagram ↗
PropertyCurrent implementation
Core count / word sizeOne core, 64-bit scalar registers
PipelineSix stages; single issue; in-order retirement
Scalar executionInteger arithmetic and scalar floating point
L1 caches4 KiB instruction cache; 8 KiB data cache
Instruction interfaceTagged accelerator request/response interface (implementation-specific)
Matrix opcodesCustom-1, opcode 0x2B; mzero / mmacc
Command orderingAccelerator dispatch waits for retirement slot 0; younger issue waits for matrix completion

Why the ordering rules matter

A fence can flush and refetch younger instructions. Issuing a matrix command before that fence retires could execute the same accumulation twice. Our integration waits until the command is at the first retirement slot, and carries its transaction ID through completion.

Vector memory bookkeeping counts accepted requests—valid and ready—so bus backpressure cannot accidentally count one request several times. The original matrix-disabled configuration remains available.

Where to read the code

The open-source core is fetched at the pinned revision in sources.lock.json. Integration patches and upstream credits make its implementation and provenance inspectable.