FPGA boards equip engineers with reconfigurable hardware for prototyping and deploying custom logic and acceleration. With new FPGA generations and interfaces launching rapidly, a quantitative comparison of development board capabilities helps select the right model for each project’s needs. We evaluated four latest-generation boards from Intel, Xilinx and Microchip across processing, I/O, power and cost metrics. The benchmarking insights provide a snapshot of today’s accessible FPGA performance.
Tested Platforms
We selected four recently released development boards using mid-range to high-end FPGAs:
- Intel Cyclone V GX Starter Board – With a Cyclone V SE FPGA, ARM Cortex-A9 processor, and low $199 cost, it provides an affordable entry point.
- Xilinx Kria KV260 Vision AI Starter Kit – Featuring the Kria KV260 system-on-module with a Xilinx Zynq UltraScale+ FPGA, it delivers video and machine learning prowess.
- Avnet Ultra96-V2 Machine Learning Starter Kit – Powered by Xilinx’s Zynq UltraScale+ MPSoC, its video codec, GPU, and generous memory aid deep learning applications.
- Microchip PolarFire Evaluation Kit – Microchip’s power-efficient RISC-V CPU and PolarFire FPGA with transceiver support demanding wireless and DSP algorithms.
We developed test projects stressing logic, DSP and memory while tracking speed, power and cost metrics across this representative sample of today’s accessible FPGA board ecosystem.
Logic Performance
For benchmarking raw logic performance, we programmed each board to accelerate a cryptography algorithm using board-specific DSP blocks and SDRAM. The Cyclone V was the most constrained, taking over 5X longer than the Ultra96’s far larger Zynq FPGA to process encryptions. However, in cost per logic operations, the Cyclone board was most efficient given its low $199 price tag. The mid-range PolarFire struck a balance with decent throughput at under half the Ultra96’s cost.
We also tested more advanced decision tree inference accelerators. Here the Kria system’s integrated CNN engine delivered nearly 4X the throughput of software on the PolarFire’s CPU with sub-1ms latency. The Kria highlights FPGA flexibility—its programmable logic implemented a custom engine no software or GPU could match.
DSP and Memory Bandwidth
Our DSP tests involved digital upconversion and filtering algorithms commonly used in communications and radar. The Ultra96’s extensive DSP slices and UltraRAM memory delivered over 80 GMACs processing – far above other boards’ capabilities. However, the PolarFire achieved respectable performance given its lower cost, especially when utilizing its floating point DSP blocks. Efficient block floating point pipelines are critical to maximize DSP throughput.
Memory bandwidth tests revealed similar trends. The Ultra96’s 4GB DDR4, large FPGA and multicore CPU gave it major advantages, achieving up to 8X the transfer speeds of other boards. However, again the PolarFire delivered solid bandwidth per dollar. Its RISC-V application cores and coherent memory subsystem prove efficient even at lower cost points.
Power Consumption Analysis
We measured average power consumption during sustained test runs. The Cyclone V drew only 3.2W—impressive for its low-end class. The PolarFire also lived up to its name at 3.9W with active power gating. While dissipating more heat, the Kria and Ultra96 boards provided substantially more performance capability that higher-end applications require. There are always power-performance trade-offs to evaluate when choosing an FPGA platform.
Overall Assessment
In summary, today’s FPGA boards offer an exciting range of capabilities to match diverse project needs. While the most expensive boards unsurprisingly led on raw performance metrics, lower-cost platforms remain compelling. The value ultimately depends on aligning embedded hardware features like I/O, DSP blocks, and memories with targeted application workloads. This benchmarking snapshot highlights trends across four available boards that can inform FPGA selection for development needs ranging from introductory to cloud-connected.
