Chestnut: The Open-Source eGPU Dock from comma.ai

George Hotz and the team at comma.ai have released a new piece of hardware called Chestnut. While designed for automotive AI, it represents a milestone for the external GPU (eGPU) community: a dock with transparent, readable firmware.

💰 Pricing and Packages

Chestnut is available in two primary configurations, both of which include a one-year hardware warranty and a 30-day money-back trial.

---

🚗 The Automotive Motivation

For the last decade, comma.ai has worked to fit openpilot driving models into a very tight power envelope. The previous limit was roughly 10 W, which eventually became a performance ceiling.

By utilizing the Chestnut dock to attach a desktop-class GPU to a comma four device, they have increased the power budget to approximately $100\text{W}$.

Performance Gains

The company aims to match the compute capabilities of Tesla's HW4. To leverage this, they are introducing a 1B-parameter driving model in openpilot 0.11.2. Compared to previous on-device models, this "chestnut-class" model features:

• $\approx 30\times$ the parameters.
• $\approx 100\times$ the FLOPs.

Data Pipeline:

• Daily Logs: $\approx 1,000,000$ driving minutes.
• Training Set: $15\%$ of all logged drives are integrated into the training cluster.

---

🛠️ Technical Architecture

At its core, Chestnut is a PCIe Gen4 x4 to USB4 bridge. The magic lies in the firmware, written in C and hosted publicly on GitHub under the tinygrad organization.

The Bridge Chip

The dock utilizes the ASM2464PD (an ASMedia USB4/Thunderbolt-to-NVMe controller). For PC enthusiasts, this is revolutionary because eGPU enclosures are typically "black boxes" with proprietary firmware blobs.

From the README: The documentation is candid about performance. A control message for reading/writing PCIe TLPs manages $3.6\text{ MB/s}$ (write) and $1.8\text{ MB/s}$ (read). While slow, it allows access to any point on the PCIe BAR.

Additional Data Paths:

• SRAM Access: A DMA path into the chip's $512\text{ KB}$ of SRAM.
• Throughput: Reaches $\approx 700\text{ MB/s}$ via $10\text{ Gbps}$ USB3.

---

💻 Why This Matters for PC Builders

While marketed for cars, comma.ai confirms that Chestnut works with standard computers. It transforms the eGPU experience from "blind trust" to "auditable hardware."

Comparison: Traditional vs. Chestnut

• Traditional eGPUs: Closed firmware $\rightarrow$ Trust the vendor $\rightarrow$ Forum-based troubleshooting.
• Chestnut: Open C source $\rightarrow$ Audit the code $\rightarrow$ Reflash the bridge.

// Conceptual representation of the open-source nature
void bridge_init() {
    printf("Initializing ASM2464PD with open-source firmware...\n");
    // Now you can actually see what's happening here!
}

The Reality Check

It is important to note that this is not a magic bullet for gaming performance:

1. Bandwidth: Gen4 x4 over USB4 is significantly narrower than the x16 slot the RX 9060 expects.
2. Target Audience: This is built for AI researchers and tinygrad workloads, not as a direct replacement for a Razer Core.

Final Note: While you can mount the GPU in your passenger footwell for the full comma.ai experience, doing so remains entirely optional.