AMD released a new serving strategy for the GLM-5.3-FP8 model on its MI300X GPU servers on October 14, 2026, saying it improves throughput for long-context and high-concurrency workloads. It is the company's first major software update in the AI serving category since its earlier optimizations for large language models.
AMD reported a 147% increase in output throughput at max concurrency 128 with the DP8+EP8 mode, measured on a 96K-in / 32K-out workload. That compares with the previous tensor parallelism (TP=8) approach, which was faster on shorter workloads.
The DP8+EP8 method is built on the ROCm 7.2.3 platform and targets complex tasks like coding and planning, which require long context and high concurrency. Availability begins with the release of vLLM v0.28.0 and is initially for developers and enterprises using AMD's MI300X hardware.
"DP8+EP8 combines data parallelism with expert parallelism across the eight GPUs," said Pradeep Sakhamoori, lead engineer. "This approach allows each GPU to manage its own request stream and latent KV cache, improving overall throughput."
The announcement follows AMD's earlier release of the MI300X GPU, which was framed as a significant step in accelerating AI workloads. AMD did not say how the DP8+EP8 method will scale to more GPUs, and it raised the question of how the model's native rope length of 1,048,576 tokens will affect performance on different workloads.
Source: amd
