AMD EPYC 9006 “Venice”: Everything Announced at Advancing AI 2026

Flagship Model Core Counts, Cache & Key Specs

The flagship SKU is the EPYC 9006 SP7. Its headline specifications represent a significant generational leap:

The Four Planned Variants

AMD detailed four distinct sub-families within the Venice lineup. Each is tailored to a different workload segment and has a separate availability timeline :

Naming note: The low-power variant is officially called the EPYC 9006 LP by AMD. Some media outlets use "Verano" as its code-name . "Venice-X" is the official name for the 3D V-Cache model (EPYC 9006X SP7) .

Performance Claims vs. Nvidia Vera and AMD Turin

AMD made several bold performance claims at the keynote, positioning Venice as a direct competitive response to Nvidia's data center CPU efforts :

• vs. Nvidia Vera CPU: AMD claimed the EPYC Venice delivers up to 2.2× the throughput of Nvidia's Grace-class "Vera" CPU in AI inference workloads .
• vs. AMD's own Turin (EPYC 9005 / Zen 5): AMD said Venice delivers up to 1.8× the performance of the previous-generation EPYC "Turin" (9005 series) across a broad set of data center workloads .
• Broader generational claim: Venice delivers 18× the throughput of the first-generation EPYC (Naples, from 2017) .

These figures are AMD's own benchmark results presented at the keynote . Independent verification is expected once samples are widely available.

Venice in the Helios Rack-Scale AI System

Venice CPUs are a core component of AMD's Helios rack-scale AI system, which was also unveiled at the event . Helios is AMD's first fully integrated rack-scale platform, designed to compete with Nvidia's rack-scale offerings.

Helios rack composition per AMD's official specs :

• 72× AMD Instinct MI455X GPUs (with HBM4 memory, ~31 TB total HBM capacity)
• 18× 6th Gen EPYC Venice CPUs
• AMD Pensando "Vulcano" networking (UALink-based)
• Aggregate performance: 2.9 EFLOPS (FP4) per rack; a double-wide version scales to 3 AI exaflops
• Aggregate memory bandwidth: 1.4 PB/s per rack

Customer and shipping status:

• Microsoft announced on July 20, 2026 that it will deploy Helios racks in Azure starting H2 2026 for AI inference workloads .
• Other named customers include Meta, OpenAI, and Oracle .
• At the keynote, AMD described Helios as "already shipping" to these customers .