The AMD Instinct MI350 Series, announced in June 2025, is presented as a major generational step for data center computing. Built on 4th Gen AMD CDNA architecture and fabricated with TSMC’s 3nm process, the MI350 family includes MI350X and MI355X GPUs with 288 GB of HBM3E memory and up to 8 TB/s bandwidth. AMD and GIGABYTE materials highlight up to 4x generational Artificial Intelligence compute improvement and a claimed 35x uplift in inference performance for certain workloads, positioning the MI350 series for training large models, high-speed inference, and demanding HPC tasks.
GIGABYTE has integrated the AMD Instinct family into a portfolio of servers optimized for density, cooling, and scale. The vendor offers UBB-based 8-GPU configurations and OAM modules such as MI325X and MI300X on universal baseboards inside G-series systems. Product examples include 4U liquid-cooled designs and 8U air-cooled enclosures including the G4L3 and G893 series, plus multi-socket systems such as the G383 series that host the MI300A APU in a four-LGA-socket configuration. Cooling options range from passive to direct liquid cooling, and systems support PCIe Gen5, Infinity Fabric links, and virtualization partitioning to address varied Artificial Intelligence and high-performance computing requirements.
The software stack emphasizes AMD ROCm 7.0 and ecosystem readiness. ROCm 7 is described as providing day-0 framework support for PyTorch, TensorFlow, JAX, ONNX Runtime, Triton, and vLLM, plus transformer kernels and distributed inference optimizations for OCP-FP8/MXFP formats. AMD testing cited from May 15, 2025 reports ROCm 7 preview delivering up to 3.5x faster inference and 3x faster training versus ROCm 6 on an eight-MI300X configuration across Llama 3.1-70B, Qwen 72B, and Deepseek-R1 models. Together, GIGABYTE hardware and AMD accelerators are framed as an open, high-density solution for enterprises and research centers pursuing large-scale Artificial Intelligence and HPC deployments.
