AMD Ryzen AI Embedded P185i socket: FP8

What the Ryzen AI Embedded P185i is

The Ryzen AI Embedded P185i repeats the 12-core P185 specification with industrial temperature qualification. Four Zen 5 cores plus eight Zen 5c cores give 24 threads with boost to 5.1 GHz, cache is 12 MB of L2 with 24 MB of L3, graphics are 16 RDNA 3.5 compute units to 2900 MHz, and the XDNA 2 NPU offers up to 50 TOPS inside 80 total system TOPS. Default TDP is 28 W in the 15 to 54 W configurable range on TSMC 4 nm.

The i grade qualifies the junction from minus 40 to 105 degrees instead of 0 to 105, which admits cold start cabinets, vehicles and outdoor enclosures. Clocks, cache, graphics size, memory support and power match the standard AMD Ryzen AI Embedded P185 exactly, so performance planning transfers and only thermal validation changes. One noted difference sits in memory ceilings: commercial boards use LPDDR5X to 8533 MT/s while industrial variants are specified to 8000 MT/s.

The 10-year lifecycle with last time buy in 2036 covers both grades, and sampling began with the March 2026 announcement ahead of third quarter production. Gorgon Point silicon is the Strix Point refresh with higher clocks and a faster NPU, so the platform inherits a mature board base.

Socket and board compatibility

The chip solders down in FP8, the 25 × 40 mm BGA shared with Phoenix through Gorgon Point designs. The joint is permanent, modules come from vendors such as Advantech, congatec and Kontron, and each build is qualified for its SKU across the industrial span including joint and interface.

Board reuse helps: the package and power envelope match the rest of the P100 high end, but firmware must know the SKU. Designs that stay within commercial limits can use the standard P185 with the same throughput and simpler margins.

Below this flagship the industrial ladder continues with 10-core and 8-core i models on the same package, and above it only the Halo class FP11 X100 series offers more cores.

Memory, PCIe and cooling

Memory is DDR5-5600 modules or soldered LPDDR5X, dual channel with ECC and Link-ECC options, with industrial boards specified to LPDDR5X-8000 against 8533 on commercial builds. The slight ceiling difference reflects validation margins at temperature extremes rather than a different controller. Capacity freezes at manufacture on soldered layouts.

Expansion gives 16 PCIe Gen4 lanes, two USB4 ports, USB 3.2 and 2.0, and four display outputs to 7680x4320. Lanes serve NVMe, networking and vision hardware per board manual, with optional B650 or X670 chipset expansion on designs that need more. The 16 compute unit graphics engine covers display and parallel compute without a discrete card.

Cooling designs to the configured TDP between 15 and 54 W with materials rated for the full minus 40 to 105 span. Interfaces that stiffen or pump out in the cold fail industrial validation, so the board thermal specification governs every choice. No boxed cooler exists.

Platform, software and lifecycle

AMD targets the industrial flagship at factory automation, transport and outdoor robotics that need 12 threads of deterministic x86 beside graphics and AI. Linux with Xen separation, FreeRTOS for real-time tasks and ROCm for GPU math form the stack, and the 10-year supply to 2036 protects certified lines.

Product families often carry both grades on one board design validated twice, which keeps software and spares common while covering indoor and outdoor installs from a single program.

Upgrade paths

Soldered parts cannot be upgraded. Downsizing means a new build around a smaller industrial SKU on FP8, and growth beyond 12 cores means the Halo X100 series. Desktop Gorgon class silicon uses AM5.

Size the SKU and the memory for the lifetime workload at project start, since neither changes afterward.