FPGA PCIe DMA into GPU hosts
I build the whole path from an FPGA to a GPU host: the PCIe DMA engine on the FPGA, the register interface and interrupts, the Linux driver, and the host software that proves the data arrives fast enough. It's for teams whose sensor, RF or measurement data needs to reach a GPU for processing, with numbers to show for it.
What I do
FPGA-side streaming
PCIe endpoint and DMA design, buffering between your data source and the link, and multiple transmit and receive streams sharing one card.
Register interface and interrupts
A register map the host can actually use, with interrupts for completion and errors, documented so the software side isn't guesswork.
Custom Linux drivers
A driver that maps buffers, handles interrupts and moves data by DMA into GPU memory, plus Python control libraries for the people running the system.
Host tuning and measurement
Tools that measure latency and bandwidth end to end, and the host tuning that closes the gap between what the link can do and what you're getting.
For the firmware and processor side of a system, see embedded firmware. For the rest of the FPGA design, see FPGA design.
Typical problems I'm brought in for
- DMA throughput is a fraction of what the PCIe link should carry.
- The PCIe link won't train, trains at the wrong width or speed, or the host doesn't see the card.
- Data goes through system memory and an extra copy before the GPU sees it, and latency suffers.
- Streams drop samples under load, and nobody can say where.
- The FPGA team and the software team each think the bug is on the other side.
- You have a vendor reference design and need a production driver your team can maintain.
Related work
Quantum computing · 2026
FPGA-to-GPU streaming over PCIe
An FPGA accelerator card that streams measurement data into an embedded GPU host by DMA, with host tuning and tools to measure latency and bandwidth.
Edge AI · RF · 2025
RF straight into GPU memory
Four transmit and four receive streams from an RF transceiver to a GPU host over PCIe, with a custom Linux driver.
Software-defined radio · 2012–2016
The smallest SDR modules of their day
FPGA for a family of tiny radios, with DDR, PCIe, 10GbE and RF data converters.
How engagements work
Design review
$2,500
fixed price · delivered in three days
I look at your data path from FPGA to GPU and send a written assessment with next steps. Credited in full toward a sprint.
Sprint
$15,000
fixed price · two weeks
An agreed scope and a working, tested deliverable: a DMA path, a driver or a throughput fix, with the repo, tests and docs.
Retainer
$25,000 / month
month to month
Ongoing development with priorities set week by week, for teams that need a senior engineer on it continuously.
Prices in US dollars. Tell me about your project and I'll suggest the right fit.
Questions
Do you do both the FPGA and the driver?
Yes, and that's the point. The FPGA DMA design, the register interface, the Linux driver and the measurement tools come from one engineer, so there's no gap between teams for bugs to hide in.
Which GPU hosts have you streamed into?
Embedded GPU hosts fed by FPGA accelerator cards and RF transceivers over PCIe. Tell me your platform and I'll tell you what I know about it.
Will you sign an NDA?
Yes. I'm happy to sign your NDA before we get into the details of your system.
How do you use AI, and who owns the code?
AI agents working inside my rule sets draft RTL, drivers, tests and docs. Golden models, simulation in CI and hardware-in-the-loop tests catch what generation gets wrong, and I review and own everything that ships. You get the repo, scripted builds, tests and docs.
Data not reaching the GPU fast enough?
Tell me what you're building and where you're stuck. I'll reply within one business day.
Start a project