6,300 MB/s File Write Performance on Embedded Linux for MPSoC

In conventional FPGA-based embedded Linux systems, the processing overhead of the built-in PCIe Hard IP and generic NVMe drivers, combined with PCIe Gen3 constraints, often limits the performance of modern NVMe SSDs to around 1,600 MB/s.

Design Gateway's fully hardware-based solution breaks through this barrier, delivering over 6,300 MB/s of sustained file write performance on FPGA SoC platforms.

Four Keys to Breaking the Performance Barrier

Integrating PCIe Gen4 Soft IP

Standard MPSoC devices are constrained by the built-in PCIe Gen3 Hard IP. By integrating a PCIe Gen4 Soft IP, bandwidth is doubled (equivalent to ~7,000 MB/s), eliminating the Gen3 bottleneck.

Full Hardware Offload

Conventional Linux drivers handle protocol processing in software, resulting in high CPU load. The rmNVMe-IP core implements NVMe and PCIe protocols entirely in hardware, offloading virtually all processing from the CPU to minimize latency.

Dual-Channel DMA Connecting PS and PL

To match internal bandwidth with external PCIe Gen4 data rates, a custom DMA is employed. Two AXI DMA channels operate simultaneously, doubling the bandwidth between Processing System (PS) memory and Programmable Logic (PL), sustaining over 8,000 MB/s of internal transfers.

Bundled High-Performance Application

Standard benchmarks such as fio may fail to fully extract the optimized hardware performance. A custom io-uring application leverages asynchronous I/O optimization to push sustained write throughput up to 6,300 MB/s.

System Requirements

FPGA Development Kit	AMD ZCU106 Zynq UltraScale+ MPSoC FPGA Board
SSD Interface Adapter	Design Gateway AB17-M2FMC
SSD	M.2 Form Factor NVMe PCIe Gen4 SSD
PC	Windows or Linux PC
IP Core	rmNVMeG4-IP
Reference Design	Random-Access NVMe Host System on Embedded Linux Using rmNVMe-IP Reference Design

Free Evaluation Demo

A free evaluation demo for the Zynq UltraScale+ MPSoC (ZCU106) is publicly available, allowing you to directly verify performance improvements on real hardware.

For more details, please refer to the demo video and documentation published on our website.

📄 Reference Design Document 📋 Demo Instruction Manual 💾 Free Evaluation Demo Bitfile

DEMO VIDEO

From 1,600 MB/s to 6,300 MB/s:
Four Keys to High-Performance NVMe File Access on FPGA-based Linux

▶ Watch Demo Video

Achieving 6,300 MB/s File Write Performance on Embedded Linux for MPSoC

Achieving 6,300 MB/s
File Write Performance
on Embedded Linux for MPSoC