AMD Instinct MI350 and CDNA 4 Take the Stage

In response to NVIDIA’s Blackwell dominance, AMD officially launched its Instinct MI350 accelerators in 2025, built on the new CDNA 4 architecture. AMD claims the MI350 delivers up to 4× performance improvements in AI compute compared to the previous MI300 series. One of the key highlights is its efficiency in transformer inference, which positions it as a serious contender in large language model deployments. At Hot Chips 2025, AMD executives revealed deep architectural changes that allow the MI350 to scale more efficiently across massive datacenter workloads. The accelerators are paired with ROCm 7 software, which boasts significant improvements in both training and inference tasks. Hyperscalers are already piloting AMD hardware in multi-vendor AI clusters to reduce dependency on NVIDIA. Meanwhile, AMD teased its upcoming MI400 series, set for release in 2026, which is expected to push even further into the trillion-parameter era. Pricing strategies are being closely watched, as AMD’s ability to undercut NVIDIA could reshape the market. For developers, the expanding support ecosystem around ROCm is making it easier than ever to adopt AMD hardware. While competition is fierce, the MI350 marks a turning point for AMD in the AI accelerator wars.