According to foreign media reports, AMD plans to launch its next-generation Instinct MI400 series of accelerators in the second half of 2026, featuring two models: the MI450X, targeting artificial intelligence (AI), and the MI430X, aimed at high-performance computing (HPC).
The MI400 series is based on AMD’s latest CDNA Next architecture. Unlike the current MI300 series, which supports both AI and HPC tasks, the MI300’s general-purpose design limits peak performance for both types of workloads. The MI400 series addresses this issue through differentiated positioning. The MI450X is specifically optimized for AI tasks, supporting low-precision computing formats such as FP4, FP8, and BF16, and removing FP32 and FP64 logic to maximize chip space for AI compute units. The MI430X, on the other hand, is designed for HPC tasks, supporting high-precision FP32 and FP64 computing, and removing low-precision AI logic to enhance HPC performance. This customized design allows the MI450X and MI430X to achieve higher efficiency and performance in scenarios such as AI training, inference, and scientific computing, respectively.
In terms of technical specifications, the MI400 series is expected to continue AMD’s advantages in memory capacity and bandwidth. Referring to the MI300 series, the MI300X is equipped with 192GB of HBM3 memory with a bandwidth of 5.3 TB/s, while the upcoming MI325X will be upgraded to 256GB of HBM3E memory with a bandwidth of 6 TB/s. The MI400 series may further adopt HBM3E or HBM4, providing up to 288GB of memory capacity and higher bandwidth to meet the demands of large-scale AI models and HPC applications. The MI300X achieves a theoretical peak performance of 2614.9 TFLOPS at FP8 precision, and the MI400 series, through architecture optimization and process upgrades (such as 3nm or more advanced nodes), is expected to significantly increase this figure.
Another highlight of the MI400 series is its support for UALink interconnect technology. UALink, jointly developed by AMD, Intel, Microsoft, and other companies, aims to provide a high-performance, scalable GPU interconnect solution, directly challenging Nvidia’s NVLink. UALink supports high-bandwidth, low-latency data transfer, making it suitable for building large-scale AI and HPC clusters. However, the commercialization of UALink currently faces certain challenges. Due to the difficulty for external suppliers (such as Astera Labs and Enfabrica) to provide mature switch chips before 2026, the MI400 series’ support for UALink may be limited to small mesh or ring topology configurations. AMD does not produce its own UALink switches and relies on partners, which increases deployment uncertainty. In contrast, the network solutions of the Ultra Ethernet Consortium are progressing faster, with commercial hardware already available, potentially providing an alternative scaling solution for the MI400 series.
In addition to UALink, the MI400 series will continue to support AMD’s Infinity Fabric technology, providing high-throughput, low-latency inter-chip communication. AMD plans to launch system-level solutions based on Infinity Fabric, such as the MI450X IF64 and MI450X IF128, supporting cluster configurations of 64 and 128 GPUs, respectively. These systems connect via Ethernet and target Nvidia’s rack-level platforms (such as the VR200 NVL144). Infinity Fabric has already demonstrated its advantages in the MI300 series; for example, the MI300A APU achieves a bandwidth of up to 5.3 TB/s through a unified CPU-GPU memory architecture, and the MI400 series is expected to further optimize this technology.
The chip design of the MI400 series also reflects AMD’s continued innovation in modular architecture. According to the latest information, the MI400 will adopt a chiplet design, including two Active Interposer Dies (AID), each AID integrating four Accelerated Compute Dies (XCD), totaling eight XCDs, which is larger than the two XCDs per AID in the MI300 series. In addition, the MI400 introduces a Multimedia IO Die (MID) to enhance data throughput and processing efficiency. This design not only improves performance but also reduces manufacturing costs and enhances product flexibility through modularity.
In terms of market positioning, the MI400 series will directly compete with Nvidia’s Hopper and Blackwell architectures. Nvidia’s H100 GPU offers a peak performance of 1978.9 TFLOPS at FP8 precision, while AMD’s MI325X has already surpassed this level. The MI400 series, through low-precision AI optimization and high-precision HPC support, is expected to further widen the gap in specific scenarios. Furthermore, AMD’s ROCm software platform will provide support for the MI400 series. The latest ROCm 6.2 version has improved inference and training performance by 2.4 times and 1.8 times, respectively, and supports key AI features such as FP8 and Flash Attention 3, ensuring the MI400 series remains competitive in the software ecosystem.
Of course, the shortcomings of the MI400 series should also be noted. In addition to the limitations of UALink, AMD’s brand influence in the AI market still lags far behind Nvidia. Nvidia dominates due to its CUDA ecosystem and early market positioning, and AMD needs to attract customers through its open ROCm platform and higher cost-effectiveness. In addition, the rapid development of the AI and HPC markets requires AMD to maintain rapid iteration. It is reported that the MI350 series (based on the CDNA 4 architecture) is expected to be launched in mid-2025 and will support FP4 and FP6 formats, potentially offering up to 2.3 PFLOPS of FP16 performance.
From an industry trend perspective, the demand in the AI and HPC markets continues to grow. Generative AI models (such as Llama 3.1 70B) place higher demands on memory capacity and computing performance, while HPC applications (such as climate simulation and drug discovery) require high-precision computing and large-scale cluster support. AMD’s differentiated strategy with the MI400 series precisely responds to these needs. At the same time, the development of open interconnect standards (such as UALink and Ultra Ethernet) will drive the industry towards more flexible and scalable architectures, and AMD, as a major participant, is expected to benefit from this trend.
The AMD Instinct MI400 series, through its customized design, advanced interconnect technology, and modular architecture, demonstrates its competitiveness in the AI and HPC fields. The launch of the MI450X and MI430X will provide users with more targeted solutions, while the support for Infinity Fabric and UALink will enhance its potential in cluster deployments. Despite facing challenges in interconnect technology and market competition, the innovative design of the MI400 series and AMD’s rapid iteration strategy position it as a significant force in the data center GPU market in 2026.