Activation Compression of Graph Neural Networks using Block-wise Quantization with Improved Variance Minimization