Case study

RF straight into GPU memory over PCIe

Edge AI · RF · 2025 RF transceiver PCIe GPU Linux driver

Four transmit and four receive streams from an RF transceiver to a GPU host over PCIe, with the register interface, interrupts and a custom Linux driver.

The problem

Running AI models on radio signals at the edge means getting raw RF samples in front of a GPU, continuously, with nothing lost on the way. A radio doesn't pause while software catches up. If the receive path falls behind, samples are dropped and the model sees a gap. If the transmit path runs dry, the radio sends garbage. With several channels running in each direction at once, those failure modes multiply.

PCIe is the obvious link between the radio side and a GPU host, but a link is only part of the answer. The host needs a way to configure and monitor the hardware, which means a register interface. It needs to know when a buffer is full or empty without burning a CPU core asking, which means interrupts. And it needs software in the kernel that ties those together and hands buffers to user space, which for custom hardware means a driver that doesn't exist yet.

None of these pieces is exotic on its own. The difficulty is that they all have to agree: the FPGA, the driver and the application share one picture of where every buffer is and who owns it at each moment. Mismatches show up as rare corrupted or missing data, the hardest kind of bug to find.

What I built

The work covered the data path from the RF transceiver to the GPU host, and the pieces that let software drive it.

Eight streams over PCIe. Four transmit and four receive streams carrying samples between the RF transceiver and the GPU host.

The register interface. The control and status map the host uses to configure the hardware and see what it is doing.

Interrupts. Signalling from the hardware to the host so software reacts when data is ready instead of polling for it.

A custom Linux driver. The kernel side that brings the hardware up, handles its interrupts and gives applications access to the streams, so the people writing the AI side work with data, not with registers.

Tech

RF transceiver 4 TX · 4 RX streams PCIe Interrupts Register interface Linux kernel driver GPU host

Related services: PCIe and GPU streaming · Embedded firmware and drivers · FPGA design

Working on something similar?

Tell me what you're building and where you're stuck. I'll reply within one business day.

Start a project