Application mapping is the first step of system design planning using Versal devices. During application mapping, you map various parts of the system into the adaptive SoC hardware, depending on your performance, latency and system cost requirements.
Embedded applications typically include an accelerator block that performs the compute intensive processing. They have a DRAM for data storage. An interconnect switch moves data from DRAM to accelerator hardware. On-chip memory provides faster intermediate access. An embedded processor controls the data flow and performs less compute-intensive tasks.
Data center applications typically comprise a PCIe® -based interface for transferring data from host to card DRAM. They also have an interconnect switch. An accelerator block performs compute-intensive functions. On-chip memory provides faster accelerator access. An embedded processor core controls the data movement in the Versal device, performs less compute-intensive functions, and communicates with the host processor.
As you map your application, consider which of the following Versal device series best suits your system application:
- Versal Prime Series Gen 2, Versal Prime, Premium and HBM Series
- The Versal Prime Series Gen 2, Versal Prime, Premium, and HBM Series include programmable logic (PL). This PL includes CLBs, RAMs, and DSPs. Also, the Versal Prime Series Gen 2, Versal Prime, Premium, and HBM Series include network on chip (NoC), PCIe interfaces, and processing subsystems.
- Versal AI Core, Versal AI Edge Series Gen 2, AI Edge, and Premium Series
- The Versal AI Core, Versal AI Edge Series Gen 2, AI Edge Series, and Versal Premium VP2502 and VP2802 devices include adaptable AI Engines. These devices include PL (CLBs, RAMs, DSPs), NoC, PCIe interfaces, and processing subsystems.
For example, you can map a compute-intensive accelerator function to either the DSP Engines or AI Engines, depending on compute, power, and latency requirements. If the function is more compute-heavy and requires multiple parallel processing, use AI Engines. Use DSP Engines if the function has small compute requirements and latency is critical.
To identify the best compute mapping for your design, refer to the following tutorials available from the AMD Technical Information Portal:
- 2D-FFT in the Vitis Tutorials: AI Engine.
- FIR Filter in the Vitis Tutorials: AI Engine.