AMD officially launched its sixth-generation EPYC server processors, codenamed "Venice," on Tuesday. Built on the new Zen 6 architecture, these processors aim to secure the company’s position in data centers increasingly focused on agentic AI.
The new lineup, which includes the flagship EPYC 9996, is designed to handle dynamic workflows where CPUs must manage autonomous agents, language models, and data retrieval. "AI significantly expands the role of the CPU at every layer of the data center," Madhu Rangarajan, Corporate Vice President of Compute and Enterprise AI at AMD, told Gizmologi.id.
Performance Benchmarks Against Industry Rivals
AMD claims the EPYC 9996 provides up to 3.7 times the performance of systems using Nvidia Vera CPUs, according to data cited by AMD. When compared to the Intel Xeon 6980P, the company reports a 2.2-fold performance lead and a 2.0-fold improvement in performance-per-dollar.
These processors can reach boost frequencies of 5.3 GHz, offering up to 1.8 times the large language model (LLM) throughput of previous generations, AMD reported. For sandbox environments, AMD states the new chips are roughly 3.3 times more efficient in agents-per-watt than Nvidia Vera and 2.5 times more efficient than the Intel Xeon 6980P.
Venice Processors Utilize SP7 Platform and TSMC 2 Nm Process
The Venice processors utilize the SP7 platform and are manufactured using TSMC’s 2 nm process. The flagship EPYC 9996 features 256 cores and 512 threads, with a default power consumption of 600 watts.
AMD is also pushing rack-scale deployments through its Helion ecosystem, which offers liquid-cooled, air-cooled, and inference-optimized configurations. Some setups reportedly support up to 73,728 threads in a single rack, a density designed to handle massive AI agent deployments. Supermicro has already begun shipping H15 series servers utilizing these processors, which set multiple SPEC CPU world records, according to Supermicro.
Microsoft Azure and Oracle Cloud Integrate New Chips
Microsoft Azure and Oracle Cloud Infrastructure (OCI) are integrating the new chips into their cloud services. Azure plans to deploy the processors across its HDv2, HXv2, and ND MI455X v7 instances.
Despite these capabilities, AMD clarified that Zen 6 was not originally conceived for agentic AI when development began four years ago. "We did not specifically predict the emergence of agentic computing, but we understood from the beginning that the CPU would remain a key component in running AI software," said Mike Clark, AMD’s lead architect for Zen, as reported by Gizmologi.id.
https://t.co/FjdQPFenlV
EPYC 9006 Series Offers Diverse Core Counts and Pricing
The lineup includes models ranging from 8-core configurations up to the 256-core EPYC 9996. The processors utilize a mix of Zen 6 cores for single-core performance and Zen 6c cores for higher density, allowing for a range of configurations depending on whether the workload prioritizes individual operation speed or massive parallel processing, according to AMD.
Pricing for the EPYC 9006 series varies significantly based on core count. The flagship EPYC 9996 is priced at $14,904 per unit for orders of at least 1,000 units, while the entry-level 8-core EPYC 9016 is priced at $700, according to AMD.
AMD is developing an EPYC Verano processor based on LPDDR5X memory, which is expected to launch in 2027. The company anticipates that the ratio of CPUs to GPUs in AI infrastructure may shift toward a 1:1 balance in the future.
Related reading