Comma AI, George Hotz's company, has introduced chestnut — a compute module for the comma four system that scales power consumption from 10W to 100W via an open PCIe Gen4 x4 to USB4 bridge. Along with the device, a new openpilot 0.11.2 control model with 1 billion parameters was released — 30 times larger than the previous on-device model. The product is already available for purchase for $249 to $799.

image
image
image

What happened

Comma AI has released chestnut — a compute module for comma four based on the ASMedia ASM2464PD controller, implementing a PCIe Gen4 x4 to USB4 bridge. The device's firmware is fully open and written in C (the tinygrad/asm2464pd-firmware repository on GitHub). The device connects to the car's USB port and scales the system's compute power from ~10W to ~100W, comparable to the Tesla HW4 level. Two configurations have been released: "tiny chestnut" — just the dock for $249, and "chestnut ready to drive" — a dock with an AMD Radeon RX 9060 8GB graphics card, cables, and mounts for $799. Installation is performed in the passenger footwell or under the seat. Along with chestnut, openpilot 0.11.2 was released with a new 1-billion-parameter control model — it requires 100 times more FLOPs than the previous model. The product is sold in the Comma AI store with a 30-day return policy and a one-year warranty.

Context

For ten years, openpilot has developed under the paradigm of squeezing models into a 10-watt chip: compression, quantization, and distillation for small processors. Chestnut challenges this approach — instead of optimizing the model for limited hardware, the team is moving to desktop GPUs via an open USB4 bridge. The project uses the tinygrad framework instead of PyTorch, making the stack completely free of proprietary dependencies. The PCIe to USB4 bridge firmware is public — a rare precedent in the automotive segment, where the industry traditionally works with closed eGPU solutions. The team noted that the 1-billion-parameter model is the first step toward chestnut-class models, which will become the main direction of autonomy development.

Why this matters for the industry

Chestnut demonstrates an alternative architecture for autonomous driving: a desktop GPU instead of a custom ASIC. At a cost of $799 for the full "hardware + firmware + model" stack, a compute level comparable to Tesla HW4 becomes available to the community — this creates a niche for startups and independent developers. The open ASM2464PD firmware proves that public PCIe to USB4 bridge code is feasible in practice, challenging the industry of closed eGPU docks. Using tinygrad instead of PyTorch for training and inference of an automotive ML model is the first known precedent, which may stimulate the development of alternative ML ecosystems.

Why this matters for users

For openpilot users, this is the most significant hardware update to the project in a long time. The device can already be purchased and installed in a car — this is not a concept, but shipping hardware. A model with 30 times more parameters offers the potential for a qualitatively new level of autonomous control, although real benchmark data has not yet been published. For ML developers: the open firmware repository is available for study and modification, and the model's on-device operation allows testing new UX patterns, including visualization of the system's decision-making.

What is still unknown / limitations

The statement about a "fundamental shift" is confirmed for the hardware architecture but is not supported by ML benchmarks. A 100-fold increase in FLOPs does not guarantee a proportional increase in quality — scaling laws for end-to-end vehicle control models are not yet defined. Data on production latency, thermal stability during long trips, and the reliability of the PCIe-USB4 bridge in automotive conditions have not been published. The model's readiness for production deployments is premature without these metrics.

Sources

Author

Look at AI, editorial team