Performance - Performance - AM027

Versal Adaptive SoC AIE-ML v2 Architecture Manual (AM027)

Document_ID
AM027
Release_Date
2025-10-23
Revision
1.1 English
The AIE-ML v2 array has a single clock domain for all the tiles and array interface blocks. The performance targed of the AIE-ML v2 array for the –1 speed grade devices is 1 GHz with VCC_AIE at 0.725V. In addition, the AIE-ML v2 array has clocks for interfacing to other blocks. The following table summarizes the various clocks in the AIE-ML v2 array and their performance targets.
Table 1. AIE-ML v2 Interfacing Clock Domains
Clock Target for –1L Source Relation to AIE-ML v2 Clock
AIE-ML v2 array clock 1 GHz AIE-ML v2 PLL N/A
NoC clock 1000 MHz NoC clocking Asynchronous, clock domain crossing (CDC) within the NoC
PL clocks 500 MHz PL clocking Asynchronous, CDC within AIE-ML v2 array interface
NPI clock 300 MHz NPI clocking Asynchronous

The following table shows the minimum bandwidth of the AXI4 memory-mapped interface by target block, under the following conditions:

  • The AXI4 memory-mapped manager operates ideally.
  • Access occurs in bulk and contiguously.
  • Bandwidth measures at the AIE-ML v2 array AXI4 NoC subordinate interface (NSU).
  • AIE-ML v2 array clock runs at 1 GHz.

The actual bandwidth varies depending on:

  • Internal bottlenecks and other non-idealities of the AXI4 memory-mapped manager (for example, PS, DMA in PMC, custom AXI4 manager in PL).
  • Contention in the NoC.
  • The AXI4 memory-mapped burst size and data access pattern.
  • The AIE-ML v2 array clock frequency.
Table 2. AIE‑ML v2 Internal AXI4 Memory-mapped Bandwidth at an AIE‑ML v2 Clock of 1 GHz
  128-bit Access 32-bit Access
Block Target Write BW (GB/s) Target Write BW (GB/s) Target Write BW (GB/s) Target Write BW (GB/s)
Tile and memory tile data memory 3.0 1.1 0.80 0.32
Tile program memory 2.9 1.1 0.32 0.32
Tile, memory tile, and array interface tile stream-switch and BD registers 3.0 2.0 1.7 0.80