The Inference Alpha: Maximizing Frontier Models on AMD

DigitalOcean reports major inference speed gains for frontier LLMs on AMD MI350X/MI355x GPUs using Wafer-assisted, custom kernel and stack optimizations. Kimi 2.5 rose from 22.5 to 255.2 tok/s, DeepSeek V3.2 improved single and concurrent throughput, and 774B GLM-5 reached 151.1 tok/s with 17.8 ms ITL on one 8-GPU node.