AMD Ryzen Embedded 8640U socket: FP8

What the Embedded 8640U is

The Ryzen Embedded 8640U is a six-core, twelve-thread embedded processor that AMD released in April 2024 with the Ryzen Embedded 8000 series for industrial AI. It uses Zen 4 cores from the Hawk Point die at 3.5 GHz base with boost to 4.9 GHz. Cache is 16 MB of shared L3 plus 6 MB of L2. Default power is the 28 W class inside a 15 to 30 W configurable band, built on TSMC 4 nm.

Graphics come from 8 RDNA 3 cores at up to 2600 MHz, the 760M class that mirrors the mobile Ryzen 5 8640U. AMD embedded pages label it plainly as Radeon Graphics without the 760M brand, but core count and clocks match. An XDNA NPU adds up to 16 TOPS for vision and automation models at the edge, with up to 39 TOPS of combined throughput across NPU, CPU and GPU.

Against the 8-core AMD Ryzen Embedded 8840U the 8640U gives up two cores and four GPU cores for lower cost and easier thermals. Against the 45 W AMD Ryzen Embedded 8645HS it trades base clock and sustained power for the 28 W envelope that fanless and compact enclosures need. ECC memory is supported where the board wires it, which separates this line from consumer mobile parts.

Socket and motherboard compatibility

The 8640U ships as a soldered BGA part in the FP8 family grouping used by this site. AMD embedded ordering pages list the OPN infrastructure as FP7r2 in a lidless package, sharing the Hawk Point die with FP8 mobile parts such as the AMD Ryzen AI 7 450. There is no socketed board market and no processor upgrade: system makers solder the APU to application boards with long-life firmware.

Boards for this series target industrial temperature ranges, extended availability and locked BOMs rather than consumer features. Display outputs, networking and serial links follow the carrier design, and AMD positions last-time-buy support years out for embedded customers. Integrators qualify the exact board and BIOS instead of shopping chipsets.

Shoppers comparing inside the line meet the same footprint across all four SKUs in this batch, from the 8640U through the AMD Ryzen Embedded 8845HS. Core count, GPU size and power class differ while the package idea and memory generation stay constant.

Memory, PCIe and cooling

Memory support is DDR5-5600 across two channels with ECC where the platform supports it. Error correction matters for unattended machines that run months between reboots, catching single-bit faults before they corrupt control loops or logs. Matched pairs feed the GPU as well as the cores, so dual-module layouts remain the recommended fit even at modest capacities.

PCIe provides 20 lanes of Gen4 for NVMe storage, frame grabbers, network cards and AI accelerators on the carrier. Lane splits follow the board design rather than a consumer chipset table. Builders adding capture or inference cards should read the carrier manual for slot widths and sharing rules.

Cooling is board specific across the 15 to 30 W band, spanning passive heatsinks in sealed boxes to fanned sinks in kiosks. No cooler ships with the tray APU. Thermal validation belongs to the enclosure: sustained clocks under vision pipelines depend on ambient limits and sink mass more than on the headline boost figure.

Upgrade paths

Soldering removes any in-place CPU upgrade. The performance ladder inside the Embedded 8000 line runs from this 6-core 8640U through the 6-core 45 W AMD Ryzen Embedded 8645HS to the 8-core AMD Ryzen Embedded 8840U and AMD Ryzen Embedded 8845HS, chosen when the board is specified. Storage and I/O mezzanines are the parts that change over the product life.

A platform move beyond Hawk Point means requalifying on a newer embedded family with its own carrier. Consumer FP8 laptops such as Strix Point systems share the package idea but not the longevity or ECC story, so they are poor substitutes for industrial duty. The Socket vs chipset: what decides whether a CPU works guide covers consumer chipset logic that does not govern these soldered embedded designs.