Comment by LTL_FTC

Take a look at AMD’s 12-channel memory servers. The newer Epycs are up to 16-channels now, 1.6TB/s. Pretty great for inference.