AMD Venice EPYC CPUs and Instinct MI455X GPUs Power Helios Rack Reveal at CES 2026

AMD showcased its next-gen server hardware including 256-core Venice CPUs, MI455X GPUs with 320B transistors, and the 2.9 EFLOPS Helios rack system during CES 2026 keynote demonstrations.

At CES 2026, AMD displayed its forthcoming data center components including Venice EPYC processors, Instinct MI455X accelerators, and the Helios rack-scale system. These products represent AMD's 2026 server roadmap with architectural advancements in packaging and density.

Venice EPYC: 256 Zen 6 Cores via 2.5D Packaging

CES 2026 AMD Venice Chip CES AMD Venice Chip Render

Venice marks AMD's first server CPU using 2.5D packaging, featuring dual I/O dies surrounded by eight Zen 6 CCDs. This configuration enables 256 cores per socket, manufactured on TSMC's 2nm process. The physical layout confirms AMD's shift from monolithic designs, though exact interconnect details remain undisclosed. Performance projections indicate significant throughput gains over current Turin processors, particularly in highly threaded workloads.

Key Venice specifications:

Core count: 256
Manufacturing process: TSMC 2nm
Package type: 2.5D with central I/O dies
Expected launch: Q4 2026

Instinct MI455X: 320B-Transistor AI Accelerator

CES 2026 AMD MI455X Chip AMD CES 2026 Keynote AMD Data Center Portfolio

The MI455X GPU features twelve chiplets fabricated on mixed 2nm/3nm nodes, surrounded by twelve 36GB HBM4 stacks. With 320 billion transistors, it targets rack-scale AI deployments. AMD confirmed three MI400-series variants:

Model	Use Case	Configuration
MI455X	AI Training	Rack-scale
MI440X	8-way Servers	Power-optimized
MI430X	HPC	High FP64 performance

The MI500 series roadmap (CDNA 6 on 2nm with HBM4e) was confirmed for 2027 release.

Helios Rack-Scale Deployment

AMD's 7000-pound Helios system integrates both Venice CPUs and MI455X GPUs in a double-width rack. Each compute tray contains four MI455X accelerators and one Venice processor, scaling to 72 GPUs and 18 CPUs per rack. Total FP4 AI performance reaches 2.9 EFLOPS with Pensando networking handling scale-out requirements.

System builders targeting AI infrastructure should evaluate Helios for large clusters, while traditional data centers may consider MI440X configurations for 8-way GPU nodes. Power efficiency metrics remain undisclosed pending final silicon validation.