AMD's 6th-gen EPYC ships with up to 256 cores per socket
Edited and reviewed by AI Ping Editorial Team
3-line summary
- AMD announced general availability of sixth-generation EPYC SP7, with server systems now shipping.
- Up to 256 cores and 16 DDR5 channels support agent execution and data preparation.
- Azure HDv2 opened preview requests, with cloud general availability planned to follow.
AMD announced general availability of its sixth-generation EPYC SP7 server processors on November 10. The family, codenamed Venice, provides up to 256 cores and 512 threads per socket. Systems from server manufacturers are shipping, while Microsoft opened requests for access to its Azure HDv2 virtual machine preview on the same day.
The launch gives enterprises a choice of new server systems to deploy and an early cloud program to evaluate. AMD identifies three main CPU roles around agentic AI: executing agents and their tool calls, supplying work and data to GPU hosts, and running the databases and business applications those agents use. An agent's decision can lead to a program running or a database being queried, creating work for the CPU infrastructure around the model.
More cores and wider data paths
The range includes models aimed at running many jobs and others aimed at high per-core frequency. The EPYC 9996 has 256 cores and a maximum boost of 4.1GHz. The 96-core EPYC 9686F reaches up to 5.3GHz. AMD defines maximum boost as the highest frequency a single core can achieve under normal server operating conditions, a different specification from sustained all-core speed.
The platform supports 16 DDR5 memory channels per CPU. With MRDIMM memory running at up to 12,800MT/s, AMD lists maximum theoretical bandwidth of about 1.6TB/s per socket. Expansion devices connect over PCIe 6.0, with twice the bandwidth per lane of the previous generation's PCIe 5.0 connection. Those specifications expand both memory traffic and the links to other devices.
AMD describes agent execution as a set of concurrent tasks. The CPU hosts the environments that run code and tools, prepares retrieval data and handles database requests. While a GPU runs the model that decides what happens next, the host system supplies the resources and data to carry out that action. AMD is pairing core density with data movement for those roles.
Servers and rack configurations
Dell lists its PowerEdge R9825 as a dual-socket, 3U system with up to 512 cores using two EPYC 9006 SP7 processors. Its 32 DDR5 slots support up to 8TB with RDIMMs or 4TB with the faster MRDIMMs. Dell lists both air cooling and optional direct liquid cooling. The memory capacity and cooling arrangement depend on the system configuration.
AMD is also applying the processors to its Helion reference rack design, introduced in October. Helion Liquid uses liquid cooling for dense CPU deployment, while Helion Air targets existing air-cooled data centers. Helion Air Inference combines CPU execution and local GPU inference within one rack. The open blueprint offers different CPU, GPU and cooling arrangements for a site's infrastructure.
Azure begins with a preview for data preparation
Microsoft's HDv2 preview focuses on collecting, cleaning and filtering data for AI training. Its published VM specifications include more than 500 physical CPU cores, 3.8TB of DDR5 memory and 30TB of local NVMe storage. A 400Gbps network connection handles incoming raw data and outgoing prepared datasets.
Microsoft also describes running many isolated environments for agent code in parallel. Its published throughput comparisons are preliminary and use an integer-rate benchmark as a proxy for agent tasks. Customers can request access to evaluate the preview. Microsoft plans to support services including Azure Kubernetes Service and Azure Databricks when HDv2 reaches general availability.
AMD says general availability of Azure and Oracle Cloud instances will follow the server shipments. Oracle Compute E7 is also set to use the new processors. AMD's November 10 release gives no firm dates for those cloud launches.
