AMD could make significant changes to its next generation of Instinct GPUs, which are designed for professional use. Currently, all models in the series are intended for both HPC (high-performance computing) and AI (artificial intelligence), but there are rumors that AMD may release separate models for each use case in its next generation.
The information comes from SemiAnalysis (thanks Tom’s Hardware), which suggests that we will see an Instinct MI450X model aimed at AI and an MI430X model for HPC, expected to be released in the second half of 2026. This move would allow AMD to better optimize each model for its specific function, offering lower precision computing (FP4, FP8, and BF16) for AI, and high precision computing (FP32, FP64) for HPC.
Improved Efficiency with Separate Models
The current Instinct MI300 series includes models that serve both purposes. By eliminating unoptimized models for specific uses, AMD will have more room to maximize the use of the die in each technology in the MI400 series.
In addition to these changes, another key update to the new generation, according to SemiAnalysis, will be the introduction of UALink interconnections. In addition to Infinity Fabric, AMD may use this new type of interconnection to compete with NVLink.
Will UALink Come Too Soon for AMD?
If the SemiAnalysis information is correct, AMD will be one of the first companies to offer GPUs for both AI and HPC with UALink. However, this move may come too early.
It seems that the development of switches for this technology has been slow, and AMD does not plan to create its own solutions. As a result, by the second half of 2026, the MI400 series may arrive with a limited selection of products capable of fully using this interconnection technology.
Rumors also suggest that AMD might offer its own rack solutions with the upcoming Instinct GPUs. These models could be the MI450X IF64 and the MI450X IF128, using Infinity Fabric, possibly through Ethernet.
We may get official information about these developments at Computex 2025 or at AMD’s own Advancing AI event in June.
