arxiv : https://arxiv.org/abs/2504.09688

GAMA is an open-source framework delivering state-of-the-art acceleration for General Matrix-Matrix Multiplication (GEMM) workloads on AMD’s AIE-ML architecture (second generation), specifically optimized for machine learning applications. GAMA achieves up to 165 TOPS (85%) for int8 and 83 TBFLOPS (86% peak) for bfloat16 precisions, outperforming existing AIE1 (first generation) frameworks.

• Comprehensive GEMM acceleration for AMD Versal AIE-ML (VEK280)
• Supports int8 and bfloat16 precisions
• Maximizes the memory utilization upto 100%
• Custom buffer placement algorithm that minimizes bank conflicts reducing memory stalls
• Staggered kernel placement pattern across AIE arrays to maximize AIE array utilization upto 94%

GAMA operates on the AMD Versal VE2802 device, leveraging 304 AIE-ML cores (8×38 grid) and supporting PLIO interfaces for communication with Programmable Logic (PL). The framework implements designs at three levels:

• Single AIE: Maximizes memory utilization by selecting optimal kernel sizes for different precisions
• Pack of AIEs: Chains AIEs leveraging the cascade interface to solve a bigger problem size
• Complete Array: Replicates packs with staggered placement to avoid routing congestion and ensure scalability

GAMA's design can be scaled using these hyperparameters

Eg: Y=2, G=4, X=2

The workload is mapped to the AIEs using multiple packs as shown below

This design gets mapped to the grid as shown below

• AMD/Xilinx Vitis 2024.1
• AMD/Xilinx Versal VEK280 Platfrom (xilinx_vek280_base_202410_1.xpfm)

source /mnt/vault1/kmhatre/Software/AMD/Vitis/2024.1/settings64.sh
export XILINX_VERSAL=/opt/AMD/common_files/xilinx-versal-common-v2024.1/
source $XILINX_VERSAL/environment-setup-cortexa72-cortexa53-xilinx-linux

export ROOTFS=/opt/AMD/common_files/xilinx-versal-common-v2024.1/rootfs.ext4  
export IMAGE=/opt/AMD/common_files/xilinx-versal-common-v2024.1/Image
export PLATFORM_REPO_PATHS=/mnt/vault1/kmhatre/Software/AMD/Vitis/2024.1/base_platforms/

source /opt/xilinx/xrt/setup.sh

make -f Makefile_VEK280 aiesim AIE_SRC=<path to the aie directory> TARGET=hw Y=8 G=4 X=9 BUFF_OPT=0
eg: make -f Makefile_VEK280 AIE_SRC=aie/aie_constr_buf_array_scaling/aie_MKN_48x240x48_MMUL_4x8x8_i8_i32  TARGET=hw Y=8 G=4 X=9 BUFF_OPT=0 aiesim

@inproceedings{Mhatre2025GAMA,
 title={GAMA: High-Performance GEMM Acceleration on AMD Versal ML-Optimized AI Engines},
 author={Mhatre, Kaustubh and Taka, Endri and Arora, Aman},
 booktitle={2025 International Conference on Field-Programmable Logic and Applications (FPL)},
 year={2025},
 organization={IEEE},
}