AMD is preparing to shake up the accelerator market with the next-gen Instinct MI400 series, which is set to launch in 2026.
This new product will come with significant performance improvements over its predecessors, including a dramatic boost in computing power and memory capacity.
The MI400 accelerators will be capable of delivering up to 40 PFLOPs of compute power in FP4 (Floating Point 4) and 20 PFLOPs in FP8, effectively doubling the capabilities of the current MI350 series. This is a huge leap forward, offering a serious upgrade for those working on demanding AI training and inference tasks. The MI400 will also come with a 50% increase in memory capacity, offering 432GB of HBM4 memory, up from the MI350’s 288GB of HBM3e. Additionally, the MI400 series will take advantage of the HBM4 standard, which provides an impressive 19.6 TB/s memory bandwidth – more than double the MI350’s 8 TB/s.
The architecture of the MI400 will see a significant change with up to four XCDs (Accelerated Compute Dies) per AID (Active Interposer Die), compared to the two XCDs in the MI300. Each AID will have separate Multimedia and I/O dies, along with dedicated MID tiles for efficient communication between the compute units and I/O interfaces. AMD will also use Infinity Fabric for inter-die communication, which is a feature that has been present in previous generations.
With these upgrades, the MI400 will be poised to dominate large-scale AI workloads and is expected to leverage AMD’s UDNA (Unified DNA) architecture. This shift could lead to a unified architecture strategy, combining the best aspects of AMD’s RDNA and CDNA models into a single, more powerful platform.
The MI400 series accelerators will launch in 2026, marking a major step in AMD’s push to compete head-to-head with Nvidia and Intel in the high-performance computing sector. If the company can deliver on these promises, it could change the landscape of AI and machine learning hardware.