AMD has launched an edge-centric subset of its enterprise-grade CPUs.

The EPYC Embedded 4005 series is designed for network security devices and entry-tier industrial edge servers, with the potential to enhance AI inference performance and efficiency.

Based on AMD’s Zen 5 x86 architecture, the processor features up to 16 cores and 32 threads in a 4 nm chiplet arrangement. AMD claimed this helps to deliver high clock speeds and up to 128 MB of L3 memory cache, designed to support latency-critical applications, including industrial real-time control systems and network packet processing.

The chip giant also highlighted their configurable thermal design power (TDP), ranging from 65W to 170W, with which users can tailor the balance between performance and energy efficiency for specific embedded applications, including next-generation firewalls (NGFWs).

Other features include double data rate 5 (DDR5) memory at 5600 million transfers per second with a 128-bit width and sideband error-correcting code (ECC), to help ensure reliable high-speed data access. Also touted was Peripheral Component Interconnect Express (PCIe) Generation 5 support with 28 lanes and 11 root ports, to aid system designers in connecting networking, storage, and accelerator components efficiently.

Embedded use cases

With a lifecycle of seven years’ planned manufacturing availability, AMD is targeting long-haul embedded and telecom use cases, touting built-in reliability, availability, serviceability (RAS) features such as DRAM ECC, on-chip ECC/parity, and PCIe error detection in order to help ensure optimum operational uptime.

The EPYC Embedded 4005 also targets compact, low-power use cases with the AM5 single-socket, in an alternative to AMD’s hyperscale-and-HPC catered multi-socket EPYC 9004 processors. The AM5 is also compatible with various generations of EPYC processors.

The Embedded 4005 also supports the AVX-512 instruction set for calculations with a 512-bit data path, potentially accelerating and optimizing AI inference workloads, in which trained machine learning (ML) models generate predictions and outputs from input data.

The chips are marketed mainly for edge over AI use cases, and are comparable in specs to Intel’s Xeon D and Xeon 6 offerings, the latter of which have been used by Nokia to power 5G network applications.

AMD’s offering comes after the mainline series of 4005 CPUs were unveiled in May, targeting enterprise markets, and deployed by the likes of cloud computing provider Vultr.

In more recent months, the company was reported to have agreed to pay the U.S. government 15% of the revenue generated from China chip sales as a condition for obtaining export permissions.

AMD also partnered with semiconductor system IP company Arteris in August to improve chiplet design using the latter’s FlexGen Network-on-Chip interconnect IP.