George Hotz never shied away from bold bets. The founder of Comma.ai just dropped a hardware package that ties his autonomous driving software directly to desktop-class graphics cards. Called Chestnut, the new dock turns any USB4-equipped laptop or comma four device into a host for full-size PCIe GPUs. And it does so with open-source firmware that developers can inspect and modify at will.
The timing feels deliberate. Comma.ai’s openpilot software has spent years squeezing smarter driving models into tiny power envelopes. On-device chips topped out around 10 watts. Now the company wants to run models that demand far more. A new 1B-parameter driving model arrives in openpilot 0.11.2. It carries 30 times more parameters and 100 times more FLOPs than the latest on-device version. Comma.ai’s own blog post makes the claim without hesitation.
Chestnut solves the power and compute gap. Users mount an AMD Radeon RX 9060 8GB GPU inside the passenger footwell or under the seat. A USB4 cable runs forward to the comma four. Power comes from the car’s cigarette lighter. The whole setup mimics the compute level of a Tesla HW4, according to the announcement. But the dock works beyond cars too.
Buyers can grab the tiny chestnut version for $249. It ships with standoffs, an ATX power adapter, GPU power cable and a one-meter USB4 cord. The ready-to-drive bundle costs $799. That package adds the RX 9060 GPU, a three-meter USB3 cable, car-specific power adapter and mounting strips. Both carry a 30-day money-back guarantee and one-year warranty. Details sit on the Comma.ai shop page.
Compatibility spans platforms. On a desktop or laptop with USB4, Chestnut acts like any other dock. Windows users stick to USB3 connections for full function. Mac owners follow separate instructions to enable supported GPUs. AMD cards deliver the broadest support across USB2, USB3 and USB4. NVIDIA cards work but only over USB4. The underlying chip is an ASM2464PD USB4 to PCIe bridge. Tinygrad, Hotz’s machine-learning framework, already taps the GPU over those lower-speed links.
Open-source firmware sets this product apart. The C-based code for the ASM2464PD controller lives on GitHub. It ships with a fallback to stock firmware so users never lose basic function. Phoronix first broke the news hours after the announcement and highlighted the open nature of the project. Their article noted the dock’s PCIe Gen4 x4 bandwidth and its appeal for both automotive and general computing users.
Developers on X reacted quickly. The tiny corp account, tied to Tinygrad, posted pricing and called it “the nicest GPU dock ever made.” Users shared plans to add Chestnut to Rivian trucks and Mac workstations. One post highlighted that the same USB-to-PCIe board powers both the bare dock and the car-ready version. Discussions in the #hw-chestnut Discord channel already focus on installation quirks and model loading.
This launch builds on earlier experiments. Back in March, CommaGuide explained how comma 3X and comma four could already connect external GPUs over USB4. Tinygrad had just enabled the necessary support. Inference then moved to the discrete card while the on-device processor handled lighter tasks. Models grew larger than 58MB because the GPU removed DSP and memory bottlenecks. An RX 9070 XT delivered performance on par with Tesla’s HW5 during internal tests. That report treated the capability as real and shipped even before a polished consumer dock existed.
Comma.ai’s data engine feeds the hunger for bigger models. The fleet logs nearly one million driving minutes daily. Fifteen percent of those drives feed a scaled training cluster. Over the past two years the company rewrote its stack so additional compute and data translate directly into better performance. Hotz’s team now plans to focus autonomy work on chestnut-class models. Development velocity should rise as a result.
Installation takes seven steps. Thread standoffs into the base. Connect power first. Seat the GPU in the PCIe slot. Secure it. Hook up the 8-pin cable. Mount the assembly with included strips. Finally run cables to the car and comma four. Airflow matters. The instructions warn users to leave space around the card.
Yet limits remain. First batch production is small. Windows support stops at USB3. Some NVIDIA features stay locked behind full USB4. And while the dock runs stock firmware as a safety net, custom firmware development will require comfort with low-level C code and PCIe bring-up.
Still, the implications stretch past openpilot. Any developer who needs portable GPU acceleration now has an open platform. Data scientists can attach high-end cards to laptops that lack discrete graphics. Linux users gain another path to eGPU setups without proprietary Thunderbolt headaches. The open firmware could spawn forks that add features the original team never considered.
Hotz positioned Chestnut as the logical next step after a decade of squeezing intelligence into 10-watt envelopes. The company once shipped fully autonomous experimental drives and point-to-point demos on phone-class hardware. Those days produced impressive intelligence density. Now the bet shifts toward raw scale. A 1B-parameter model marks the start. Future versions will push parameters and FLOPs higher.
Industry watchers have seen similar moves before. Tesla scaled its Dojo supercomputer and vehicle inference together. Other autonomy players chase ever-larger foundation models. Comma.ai takes a different route. It sells the hardware that lets customers bring their own GPU and run the same software stack at home or on the road. The approach keeps the barrier to entry low while giving serious users a clear upgrade path.
Whether Chestnut becomes a hit depends on real-world results. Early adopters will test how reliably the 1B model drives in rain, at night and during edge cases. They will measure actual power draw from the car outlet and thermal behavior inside a closed cabin. Developers will probe the firmware for new use cases. If the community embraces the open design, Chestnut could outlive its original purpose.
For now the product exists. It ships tomorrow for those who order today. The code is public. The models are coming. And the line between a consumer driver-assistance device and a rolling AI workstation just blurred. Short step. Big compute. Open source. The combination feels very much like a Hotz production.