AMD confirms next-generation MI500 GPU built on 2nm node with CDNA 6 architecture and HBM4E memory.

AMD has confirmed details about its next-generation Instinct MI500 series AI accelerators, which will launch in 2027. Following a shift to an annual cadence similar to NVIDIA's approach, AMD will offer both standard and "Ultra" variants of its Instinct GPUs. These accelerators will power next-generation AI racks and are expected to deliver a disruptive uplift in overall performance as the industry accelerates its push for more capable AI infrastructure.

The MI500 accelerators will be fabricated on an advanced 2nm process technology at TSMC. While the existing MI400 series also utilizes a 2nm process, the technology employed for MI500 represents a more advanced iteration with various enhancements. This manufacturing advancement will support the architectural and memory improvements AMD is integrating into the new generation.

The MI500 will leverage the new CDNA 6 architecture, in contrast to the CDNA 5 architecture used by the MI400 series. Rather than transitioning to the UDNA architecture naming convention as previously reported, AMD will maintain the Instinct GPU naming structure. The accelerators will incorporate HBM4E memory, offering significantly higher speeds and memory bandwidth compared to the 19.6TB/s achieved by HBM4-powered MI400 accelerators.

AMD has emphasized the substantial performance gains expected from the MI500 series, positioning the company on a trajectory to achieve over 1000x AI performance improvement in just four years. This aggressive performance scaling is critical to address growing AI demand and maintain competitiveness in a market where rivals are similarly accelerating their development cycles. With MI500's 2027 launch, AMD aims to deliver the performance uplift necessary to compete in an increasingly demanding AI infrastructure landscape.