AMD Instinct MI400: The new AI accelerator doubles power and memory | 5 basic components of cpu | Hardware restaurant | Cpu definition and function | Turtles AI

AMD previews future Instinct MI400 accelerator: 40PFLOP FP4, 20PFLOP FP8, 432GB HBM4 at 19.6TB/s, twice as fast and with more memory than the MI350. Expected to debut in 2026 together with the “Helios” rack.

Key Points:

• 40PFLOPs in FP4 and 20PFLOPs in FP8 for the MI400
• 432GB of HBM4, +50% compared to the previous generation
• 19.6TB/s memory bandwidth, 2.5× MI350
• “Helios” system rack with 72 GPUs, open standards UALink/UltraEthernet

In San Jose, during the “Advancing AI” event on June 12, 2025, AMD’s Lisa Su raised the curtain on the next-generation hardware roadmap: while the first Instinct MI350 systems debut in Q3 2025, the real news is called MI400, coming in 2026. Based on the CDNA-Next architecture (also known as “UDNA”), this GPU is presented with a theoretical power of 40 petaflops in FP4 and 20 in FP8, that is, a net doubling of performance compared to the MI350X series. The leap is not only computational: HBM memory goes up to 432GB (+50%) with a bandwidth of 19.6TB/s — about double the previous 8TB/s.

The GPU will adopt a multi-chiplet design: up to four XCD (Accelerated Compute Die), two AID (Active Interposer Die), and dedicated die for multimedia and I/O, each AID with MID tiles to increase the efficiency of internal communication. Connectivity between chips will maintain the Infinity Fabric, improved in its fourth generation.

The platform is not limited to a single GPU: the concurrently announced “Helios” rack architecture will house 72 MI400s per rack, integrated with EPYC “Venice” (Zen 6) CPUs and Pensando “Vulcano” NICs, using open-standard UALink and UltraEthernet interconnects. Scale-up bandwidth will reach 300GB/s per GPU, with racks capable of delivering up to 2.9 exaflops FP4.

AMP has already entered into strategic partnerships with OpenAI, Meta, Oracle, Microsoft, xAI, and Crusoe to support the MI400 platform. OpenAI has formalized its adoption of the future MI400/MI450 GPUs and participation in the development of the MI450, while Crusoe has announced a $400 million investment.

AMD’s stated goal is to counter Nvidia’s closed ecosystem (NVLink, NVSwitch) by offering open standards and interoperability. In addition, the company has focused on expanding the ROCm software to achieve full parity with CUDA. Despite the ambitious debut, AMD stock fell between 1% and 2% on the day of the announcement, while analysts remain cautious about the potential immediate impact.

In parallel, rumors about further developments are emerging: the MI450X accelerators with IF64/IF128, expected in 2026, could push rack-scale configurations up to 128 GPUs, increasing scalability with Infinity Fabric via Ethernet.

A further piece of AMD’s strategy is the full integration of the rack infrastructure: after the acquisition of ZT Systems, the company is aiming for complete hardware-software solutions, as already implemented with the MI350 racks and the upcoming Helios.

This preview paints a picture of an AI infrastructure future centered on powerful GPU modules, ultra-fast memory, and open racks, laying the foundation for the advanced ecosystem of the coming years.