FP11 socket: CPUs, chipsets and cooler fit

A soldered package, not a socket to plug into

On FP11 there is no socket in the desktop sense. The processor arrives soldered to the motherboard or to a computer-on-module such as a COM-HPC board, with 2,077 balls joining it to the board. That joint is made once at the factory and cannot be reworked in the field. Buyers used to socketed desktops should reset their expectations: model selection, memory size and thermal design are all final at purchase.

The practical result is a qualification mindset. AMD validates the embedded X100 SKUs for continuous operation and offers them for up to 10 years with last time buy dates out to 2037, and the industrial temperature variants cover minus 40 to 105 degrees. A product team picks the SKU whose core count, GPU size and TDP fit the worst case workload over that whole period, because there is no later CPU swap to fix an undersized choice. Sampling of the X100 series began in June 2026 with production availability planned for the fourth quarter, so early designs run on the same silicon that later ships in volume.

Memory, expansion and cooling

The 256-bit LPDDR5X bus is the reason FP11 exists. Eight channels feed the CPU cores, the up to 40 GPU compute units and the NPU from one pool, which removes the copying between separate CPU and GPU memories that discrete designs need. Embedded X100 configurations reach LPDDR5X-8533 with Link-ECC and up to 128 GB, which is what lets large vision and language models stay resident for robotics workloads. Consumer Strix Halo boards use LPDDR5X-8000 with the same width.

For expansion the SoC gives 16 PCIe Gen4 lanes alongside USB4 ports, USB 3.2, HDMI 2.1, DisplayPort 2.0 and embedded DisplayPort outputs, driving up to four displays at 7680x4320. The lanes typically serve NVMe storage, networking and camera or sensor cards rather than a discrete GPU, since the integrated Radeon graphics already match performance segment discrete cards of its generation. Board designers split the 16 lanes to suit the product, so the exact slot layout always comes from the board manual.

Cooling follows the system, not a retail heatsink catalogue. The embedded SKUs default to 55 W inside a 45 to 120 W configurable window, and the board needs a heatsink or cold plate validated across that range and across the rated temperature span. The package is lidless with a stiffener ring, so the thermal solution mounts to the board and presses on the die area through the specified interface material.

Where FP11 boards are used

AMD aims FP11 boards at physical AI: robots, medical imaging, machine vision, drones and industrial automation that need CPU, GPU and NPU in one long lived component. The Kria AI system-on-module built around the X199 pairs the SoC with an FPGA carrier for cameras, motors and industrial networking, and partner modules in COM-HPC form come from houses such as congatec, Arbor, iBase, IEI, Sapphire and Seavo. Software support covers Linux, ROCm, PyTorch, ONNX and TensorFlow with Xen virtualization for mixed criticality designs.

The 10-year embedded lifecycle is the other half of the offer. Consumer Strix Halo laptops turn over yearly, but an industrial controller or medical device cannot requalify a processor every two years. AMD lists last time buy in 2037 for the X100 SKUs, which lets a design certified once stay in production. That promise is why the socket page matters even though nothing plugs into it: it records which silicon, memory and I/O a decade of boards will share.

When the product needs less than Halo, the smaller FP8 package covers Strix Point, Hawk Point, Phoenix and Gorgon Point APUs in 25 × 40 mm with a narrower memory bus. FP11 is the step up for designs where 16 cores, 40 GPU compute units and 128 GB of unified memory justify the larger board and the higher power budget.