AMD is gearing up to launch its next-generation Instinct MI355X AI accelerator, promising substantial performance improvements and enhanced energy efficiency. The new chip, built on an advanced 3nm process node, is set to leverage AMD's latest CDNA 4 architecture and introduce support for next-gen HBM3E memory.

According to AMD's roadmap, the MI355X is slated for release in the second half of 2025. The accelerator is expected to deliver up to 9.2 PFLOPS of compute power in both FP6 and FP4 formats, marking a significant leap in AI processing capabilities. This represents an 80% increase in performance compared to its predecessor, the MI325X, in FP16 and FP8 modes.

"CDNA 4 achieves a 35-times performance improvement over CDNA 3, with AI compute capabilities increasing sevenfold."— AMD statement

The MI355X is designed to push the boundaries of memory capacity and bandwidth, offering up to 288GB of HBM3E memory. This represents a 50% increase in memory capacity compared to previous generations. Additionally, the new architecture is expected to improve data transfer speeds by 8 TB/s over the MI300X, further optimizing AI workload execution.

As part of AMD's broader AI strategy, the company is also developing the Instinct MI400 series, scheduled for release in 2026. This upcoming line will utilize the next-gen AMD CDNA-Next architecture, specifically tailored to enhance AI training and inference workloads. While detailed specifications remain undisclosed, early indications suggest a revolutionary design featuring two Active Interposer Dies (AIDs), each hosting four Accelerated Compute Dies (XCDs).

With these advancements, AMD aims to solidify its position in the competitive AI hardware market, challenging rivals like NVIDIA and Intel. As AI workloads continue to demand more computational power and energy efficiency, the race to develop more capable accelerators intensifies, promising exciting developments for the AI and high-performance computing industries in the coming years.