- AMD’s MI400 GPU expected to offer 432GB HBM4 across 12 memory stacks
- 256-core EPYC Venice CPU to debut with PCIe Gen6 and 2nm process
- AMD’s Helios rack aims for 10x performance with double-wide AI architecture
AMD recently revealed its Instinct MI350 series of GPUs, but the bigger news is what the chip giant has planned for 2026.
The company is preparing a next-generation AI platform which includes the Instinct MI400 GPU, EPYC “Venice” CPU, and a major shift in rack design with its Helios infrastructure, an expansive, double-wide configuration aimed at scaling performance and bandwidth.
The MI400 GPU is expected to ship with up to 432GB of HBM4 memory, built using 12 stacks of 36GB HBM4, based on numbers shared by Micron and memory per-GPU estimates from AMD’s rack capacity.
A direct response to Nvidia
This would be a considerable jump from the current MI350’s 8-stack setup, putting it on track to rival Nvidia’s upcoming Vera Rubin platform.
Alongside the MI400, AMD plans to debut its 256-core EPYC “Venice” server CPU, built on a 2nm process with PCIe Gen6 support and up to 1.6TB/s of memory bandwidth.
The platform will also feature the Vulcano 800G NIC and support open standards like UALink and Ultra Ethernet for improved scale-out connectivity.
AMD says the Helios rack and MI450 GPU will offer up to 10x the performance of the MI355X, positioning it as a direct response to Nvidia’s accelerated roadmap.
Sam Altman appeared on stage at the company’s recent Advancing AI event alongside AMD’s CEO Lisa Su to emphasize OpenAI’s early interest in the new platform, saying “I think it’s going to be an amazing thing.”
Although the hardware won’t arrive until 2026, AMD’s preview sets the tone for what could be a fierce race with Nvidia.
The Helios rack might not have an official name yet, but it’s clear AMD is thinking bigger – and wider! – as it moves toward next-gen AI infrastructure.
Via ServeTheHome
