AMD is really trying to flex on Nvidia with some monster AI chips. The company showed its first 2-nanometer processors, an EPYC server CPU called Venice, and an Instinct MI455X accelerator, built for a liquid-cooled system named the Helios AI rack. The CEO held up the new silicon, highlighting a configuration with one Venice CPU and four of the MI455X GPUs per node. That setup scales up to thousands of cores and an absurd amount of memory bandwidth across a full rack. AMD claims the entire Helios platform can deliver nearly three exaflops of AI compute.
The Instinct MI455X GPU is a massive chip with two main compute dies and sixteen HBM4 memory stacks. They are positioning it against Nvidia's upcoming rival hardware, claiming advantages in scale-out bandwidth and memory capacity. The accompanying EPYC Venice CPU uses a Zen 6C core architecture to cram up to 256 cores into a single package by utilizing eight compute chiplets. This design promises major performance and efficiency gains over current server parts. A more mainstream version with 192 cores will also be available.
Beyond the flagship, AMD outlined a broader data center portfolio using these new technologies. This includes a dense rack holding seventy-two GPUs for hyperscale companies and a smaller eight-GPU solution for enterprise clients. They are also preparing cache-enhanced CPUs and specialized GPUs for scientific computing. The company stated these 2nm chips are now in production, with shipments to customers expected later this year.
The Instinct MI455X GPU is a massive chip with two main compute dies and sixteen HBM4 memory stacks. They are positioning it against Nvidia's upcoming rival hardware, claiming advantages in scale-out bandwidth and memory capacity. The accompanying EPYC Venice CPU uses a Zen 6C core architecture to cram up to 256 cores into a single package by utilizing eight compute chiplets. This design promises major performance and efficiency gains over current server parts. A more mainstream version with 192 cores will also be available.
Beyond the flagship, AMD outlined a broader data center portfolio using these new technologies. This includes a dense rack holding seventy-two GPUs for hyperscale companies and a smaller eight-GPU solution for enterprise clients. They are also preparing cache-enhanced CPUs and specialized GPUs for scientific computing. The company stated these 2nm chips are now in production, with shipments to customers expected later this year.
