AMD Ryzen AI Embedded X188i socket: FP11

What the Ryzen AI Embedded X188i is

The Ryzen AI Embedded X188i is a 12-core, 24-thread embedded APU that AMD launched in July 2026 as part of the X100 series, the industrial branch of the Strix Halo silicon. It pairs 12 uniform Zen 5 cores with 32 RDNA 3.5 graphics compute units and an XDNA 2 NPU rated up to 50 TOPS. The base clock is 3.2 GHz and the maximum boost is 5.0 GHz, with 12 MB of L2 and 64 MB of L3 cache behind the cores.

The letter i marks the industrial temperature grade. Where the standard X188 is rated for junction temperatures from 0 to 105 degrees, the X188i is qualified from minus 40 to 105 degrees for equipment that cold starts outdoors or runs in hot enclosures. The silicon, clocks, cache, graphics configuration and 55 W default TDP are otherwise the same as the standard model. AMD backs the whole X100 series with a 10-year embedded lifecycle and a last time buy date of 2037.

Strix Halo is a chiplet design: one or two Zen 5 CCDs sit beside a large I/O die that holds the graphics engine, the NPU, a 32 MB MALL cache for the GPU and the 256-bit memory controller. That layout is why the package is so large, and why the chip needs the FP11 ball count. The 4 nm process keeps the 55 W default inside a 45 to 120 W configurable window that the board maker sets for the product.

Socket and board compatibility

The X188i is soldered to its board in the FP11 ball grid array, so the first compatibility fact is that there is no socket to plug into and no later processor swap. The 2,077-ball joint is made once at the factory, and a design that needs this chip qualifies it for the full production run. Computer-on-modules in COM-HPC form and AMD's own Kria AI system-on-module carry the same silicon for buyers who prefer a module to a chip-down layout.

Because the package is shared with the whole X100 series, the board around it looks the same for the X199, X188 and X168: soldered LPDDR5X memory, a system heatsink or cold plate, and firmware that knows the SKU. Moving between SKUs means a different board build, not a BIOS setting. Partner modules from vendors such as congatec, Arbor, iBase, IEI, Sapphire and Seavo were announced for production in the fourth quarter of 2026.

The industrial grade is the reason to pick this SKU over the standard AMD Ryzen AI Embedded X188. Anything rated only for commercial temperatures risks failing qualification in rail, outdoor robotics or factory equipment that sees frost. When the deployment stays indoors within commercial limits, the standard model gives the same performance for less qualification effort.

Memory, PCIe and cooling

Memory is soldered LPDDR5X up to 8533 MT/s on eight channels, which is a 256-bit bus shared as one unified pool by the CPU, the GPU and the NPU. Configurations reach 128 GB, and embedded boards add Link-ECC. The unified pool removes the copies between separate CPU and GPU memories, which is what lets large vision models stay resident for robotics. Capacity and speed are fixed when the board is built, so they join the SKU as permanent choices.

Expansion comes from 16 PCIe Gen4 lanes plus two USB4 ports, USB 3.2 and three USB 2.0 ports, with display outputs covering HDMI 2.1, DisplayPort 2.0 and embedded DisplayPort for up to four screens at 7680x4320. The lanes typically serve NVMe storage, networking and sensor cards, since the 32 compute unit integrated graphics already handle display and parallel compute. The exact lane split is set per board, so the board manual governs.

Cooling is a system responsibility across the 45 to 120 W configurable range, and the minus 40 degree rating cuts both ways: the heatsink must still make proper contact when cold, and the thermal interface must survive the full span. The lidless package carries a stiffener ring, and mounting pressure follows the board specification. There is no boxed cooler and no retail heatsink list for a soldered embedded APU.

Platform, software and lifecycle

AMD positions the X100 series for physical AI: robots, machine vision, medical imaging, drones and industrial automation that run perception, planning and control beside AI inference. The software stack covers Linux, ROCm, PyTorch, ONNX and TensorFlow with Xen virtualization for mixed criticality designs, and the deterministic x86 cores suit real-time control loops that do not belong on an accelerator. Sampling began in June 2026 with production availability planned for the fourth quarter.

The 10-year lifecycle with last time buy in 2037 is aimed at equipment with long certifications. A medical device or rail controller requalifies at high cost, so a processor that stays available matters more than a small benchmark margin. The industrial temperature rating extends the same logic to the operating environment.

Designs that need fewer cores but the same graphics can look at the AMD Ryzen AI Embedded X168i, which keeps 32 compute units with eight CPU cores. The smaller FP8 package covers the P100 series for lower power budgets.

Upgrade paths

There is no CPU upgrade path on a soldered package. When the workload outgrows 12 cores, the step inside FP11 is the 16-core AMD Ryzen AI Embedded X199i, which needs its own board build with the same memory and power design scaled up. When the workload shrinks, the 8-core industrial model is the corresponding step down.

Moving off FP11 means a new platform. The FP8 P100 series serves lower power embedded designs with up to 12 smaller cores, and desktop builders meet related silicon on socketed boards. Because memory is soldered on FP11, even a memory size change means a new board, so capacity planning belongs at the start of the project.