At Advancing AI 2026, @rapt_ai demonstrated six different models sharing a single AMD Instinct MI300X GPU — automatically, using fine-grained compute-unit and memory slices, with no manual configuration.
The same GPU went from 278 to 650 tokens per second, a 2.4× increase, at
AMD Data Center: powering AI, HPC and cloud with EPYC CPUs, Instinct GPUs, Pensando DPUs and open software. Support: bit.ly/AMDServer_


