AMD has officially previewed its next-generation Instinct MI400 GPU series and a comprehensive rack-scale solution called Helios at the ´Advancing AI 2025´ event, signaling an aggressive roadmap to challenge rivals in the Artificial Intelligence hardware sector. The MI400 chips, scheduled for early 2026, represent a massive leap with twelve HBM4 stacks per card. This configuration delivers a staggering 432 GB of on-package memory and approaches 19.6 TB/s of memory bandwidth, thanks to each early HBM4 module running at around 1.6 TB/s. On the computational side, the MI400 boasts 20 PetaFLOPS of FP8 throughput and 40 PetaFLOPS FP4, outstripping the sparse-matrix performance of the company’s current MI355X lineup by twofold.
The advancements extend beyond raw performance. AMD’s upgraded Infinity Fabric now delivers 300 GB/s per link between GPUs, nearly double the bandwidth seen in previous MI350 cards. This innovation removes previous scaling bottlenecks, making it possible to connect far more GPUs per rack without falling back on slower Ethernet networking. The increased interconnect bandwidth is a cornerstone for Helios, AMD’s fully integrated rack solution. Each rack merges tomorrow’s EPYC ´Venice´ server CPUs with MI400 accelerators and custom networking hardware, presenting an out-of-the-box deployment model aimed squarely at data center operators and cloud service providers looking to scale Artificial Intelligence workloads efficiently.
Performance comparisons underscore AMD’s ambitions. A single Helios rack containing 72 MI400 GPUs yields an estimated 3.1 ExaFLOPS in tensor operations and houses 31 TB of HBM4 memory. Against NVIDIA’s upcoming Vera Rubin system—with identical GPU counts but less memory per card—AMD’s offering surpasses in memory bandwidth and overall capacity, if by a slim margin in raw ExaFLOPS. Looking further ahead, AMD is already developing the MI450X IF128 system, anticipated in late 2026. This future design will interconnect 128 GPUs directly over Infinity Fabric at a staggering 1.8 TB/s bidirectional per device, laying the foundation for exceptionally large-scale clusters dedicated to next-generation Artificial Intelligence training and inference workloads.
