PROTON PLATFORMTHE PROTON PLATFORM
Research

One system from model evidence to edge execution.

Proton connects workload analysis, developer software, architecture simulation, evidence-gated engineering, and accelerator hardware into one measurable path for local AI.

PUBLIC STATUS BOUNDARY

Product family in architecture, prototype, and development stages

SYSTEM VIEW
CEREBRAL CHIPS
P-01Proton NPUprototype
P-02Proton LPUresearch
P-03Proton SDKprototype
P-04ProtonSimprototype
P-05Proton Forgeprototype
SHARED CONTRACTWorkload → evidence → decision
01SYSTEM OVERVIEW

The platform is the product—not a disconnected chip and toolchain.

Modern edge inference depends on model formats, numerical behavior, compilers, runtimes, firmware, memory movement, and compute architecture working as one system. Proton is organized around that entire path.

  • 01Workload evidence informs architecture
  • 02Architecture constraints reshape software
  • 03Verification evidence determines what advances
02PROTON NPU / PUBLISHED RTL

The working scalar, vector and matrix reference platform.

An open hardware reference platform that runs scalar, RVV 1.0 vector and custom INT8 matrix instructions in one bare-metal program, verified in RTL simulation.

Proton NPU is the working reference platform for scalar, vector and matrix integration. The longer-term Proton LPU research explores quantization, memory movement and language-model execution beyond this baseline.

03PROTON LPU

Quantization-native accelerator architecture.

Proton LPU-E0 is the current single-tile research direction: RISC-V control, vector and matrix execution, explicit local memory, DMA, and a practical host software path.

Research architecture; no production silicon exists.

04PROTON SDK

From model graph to matrix instructions.

An open software stack that runs Stories260K end to end on Proton NPU in RTL simulation, using bare-metal IREE and INT8 matrix kernels.

Verified in Verilator on the existing RTL. Batch one, bounded context, model-specific graph. These results do not establish silicon performance or general model support.

05PROTONSIM

Choose the machine from workload behavior.

ProtonSim is intended to connect operator mix, tensor shapes, quantization, memory traffic, and execution regimes to traceable architecture decisions before RTL is frozen.

06PROTON FORGE

Agents accelerate iteration; evidence gates acceptance.

Forge coordinates specification, generation assistance, compilation, differential checks, randomized testing, coverage, regression analysis, and human review.

07SHARED EVIDENCE LAYER

Every product produces inputs for the next decision.

Workload contracts, architecture models, compiled artifacts, reference comparisons, verification results, and review notes form the connective tissue across Proton.

  • 01Traceable assumptions
  • 02Versioned artifacts
  • 03Independent checks
08STAGE LEGEND

Status labels are part of the engineering contract.

Research, prototype, in-development, planned, and future labels separate current work from direction. They are intentionally visible throughout the site.

CONTINUE THE SYSTEM

Explore Proton NPU

Continue