AMD Advancing AI 2025: Inside the New Instinct MI350 GPU and Helios AI Rack

AMD Advancing AI 2025 showcases AMD's vision in artificial intelligence, announcing the next‑generation AMD Instinct GPU and the latest open‑source AI software stack ROCm 7, and presenting an end‑to‑end open‑standard rack‑scale AI infrastructure built around AMD Instinct MI350 series accelerators, the fifth‑generation AMD EPYC “Venice” CPUs and AMD Pensando “Pollara” NICs.

Instinct MI350 series GPU

Based on the new CDNA 4 AI acceleration architecture, the MI350 series features an enhanced matrix engine, supports new data formats and mixed‑precision arithmetic. Built on a 3 nm process with 185 billion transistors and HBM3E memory, it introduces FP4 and FP6 AI data formats. The lineup includes the air‑cooled MI350X and liquid‑cooled MI355X, delivering up to 4× AI compute performance and 35× inference performance over the MI300X, and the MI355X offers up to 40 % more tokens per dollar than competing solutions.

ROCm 7 software stack

ROCm 7 is designed to meet the growing demands of generative AI and high‑performance computing workloads, significantly improving the developer experience. Compared with ROCm 6, inference performance improves on average by 3.5× and training by 3×. It adds broader framework support, expanded hardware compatibility, and introduces new development tools, drivers, APIs and libraries.

Next‑generation AI rack “Helios”

The Helios rack will be built on the upcoming AMD Instinct MI400 series GPUs, the Zen 6‑based AMD EPYC “Venice” CPU and AMD Pensando “Vulcano” NICs, with a planned 2026 launch of the full MI400 product line that promises up to a ten‑fold performance boost for expert mixed‑model inference.

“Venice” EPYC processor

The next‑generation EPYC “Venice” CPU uses a 2 nm process and Zen 6 architecture; AMD disclosed performance highlights but did not provide detailed specifications.