GPUDirect + SDR: How to Move One Billion Samples per Second over PCIe
summary
Epiq integrates GPUDirect with its PCIe SDRs so digitized IQ streams land directly in GPU memory via an updated PCIe DMA driver, tested on a Sidekiq X4 feeding an Quadro P2000 and an Nvidia Xavier platform under Linux. Results show PCIe Gen2 x4 caps the bus (~1.7 GB/s), while Gen3 x8 yields ~3.2 GB/s conventional versus ~4.7 GB/s with GPUDirect (≈45% improvement, ~1.17 Gsamples/s at 16‑bit IQ). CPU still handles DMA descriptors and metadata, but the pipeline works.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:17SDR system components
- 3:43GPUs for signal processing
- 5:08Classic SDR data flow
- 6:11NVIDIA GPUDirect overview
- 7:38GPUDirect SDR hardware
- 9:45GPUDirect DMA transfers
- 10:28Initial test results
- 13:41PCIe Gen 3 upgrade
- 15:06NVIDIA Xavier testing
- 15:47Achieving over one billion samples per second
- 16:30When GPUDirect matters
- 17:13GPU benefits beyond throughput
- 18:39Further exploration
- 20:25Acknowledgments
- 21:06Q&A
related talks
12mGetting GPUDirect to Work with Broadcom NICs
32mSponsor Talk – Haydn Nelson
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
33mReaching Ludicrous Speed
17mVDUSE - vDPA Device in Userspace
31mBPF programmable netdevice
33mAsynchronous Rails
16mGPU Accelerated Containers on Apple Silicon with libkrun and podman machine