AMD Ryzen AI Embedded X199 socket: FP11

What the Ryzen AI Embedded X199 is

The Ryzen AI Embedded X199 is the top model of the X100 series that AMD launched in July 2026, carrying the full Strix Halo configuration of 16 Zen 5 cores and 40 RDNA 3.5 graphics compute units. Threads total 32, the base clock is 3.0 GHz and the maximum boost reaches 5.1 GHz. Cache runs to 16 MB of L2 and 64 MB of L3, with a further 32 MB of MALL cache reserved for the graphics engine. The default TDP is 55 W inside a 45 to 120 W configurable window.

All 16 cores are full performance Zen 5 cores with simultaneous multithreading and AVX-512 support, and AMD makes a point of the uniformity: there are no efficiency cores to schedule around, which suits real-time control code. The integrated GPU runs to 2900 MHz and the XDNA 2 NPU adds up to 50 TOPS, so CPU, graphics and AI acceleration share one die and one memory pool. The process is TSMC 4 nm.

As an embedded part the X199 adds what the consumer Ryzen AI Max+ 395 lacks: qualification for continuous operation, a 10-year supply program with last time buy in 2037, and module options such as the Kria AI system-on-module. Customer sampling started in June 2026 ahead of production availability in the fourth quarter. The standard temperature grade covers junction temperatures from 0 to 105 degrees, while the AMD Ryzen AI Embedded X199i extends the low end to minus 40.

Socket and board compatibility

The X199 solders to its board in the FP11 ball grid array, and that fact shapes every compatibility statement on this page. With 2,077 balls on a 37.5 × 45 mm substrate, the joint is permanent, so board, memory size and thermal design are chosen once for the life of the product. Chip-down layouts and COM-HPC modules are the two vehicles, with partner modules announced from congatec, Arbor, iBase, IEI, Sapphire and Seavo.

Firmware must know the SKU, so a board built for another X100 member does not accept the X199 by update alone. The package, power envelope and memory wiring match across the series, yet each build is validated for its specific chip. Reference designs include AMD's Maple board and the Kria AI robotics developer platform, which pairs the SoC module with FPGA logic for cameras and industrial networking.

Designs that cannot use the full 16 cores stay inside the same package: the AMD Ryzen AI Embedded X188 offers 12 cores with 32 graphics compute units on FP11. Anything smaller belongs on FP8, where the P100 series serves lower power budgets.

Memory, PCIe and cooling

The memory subsystem is the heart of this chip: eight channels of soldered LPDDR5X up to 8533 MT/s form a 256-bit bus with up to 128 GB shared coherently by CPU, GPU and NPU, with Link-ECC on embedded boards. Sharing one pool removes the traffic between separate memories and lets large models for vision and language stay resident, which robotics needs for low latency responses. Speed and capacity are frozen at board build time.

For I/O the SoC supplies 16 PCIe Gen4 lanes, two USB4 ports, USB 3.2 and 2.0, and display outputs for up to four screens at 7680x4320 over HDMI 2.1, DisplayPort 2.0 and embedded DisplayPort. The lanes feed NVMe storage, networking and sensor interfaces rather than a discrete graphics card, since the 40 compute unit iGPU already covers parallel compute. Lane allocation differs per board and lives in the board manual.

Thermal design spans the 45 to 120 W configurable range around the 55 W default, using a system heatsink or cold plate against the lidless package with its stiffener ring. Continuous 24/7 operation is part of the qualification, so the cooling margin must hold at sustained load, not just bursts. No cooler ships with the chip.

Platform, software and lifecycle

AMD aims the X199 at physical AI workloads where one x86 APU replaces a CPU plus discrete GPU plus accelerator card: humanoid robots, autonomous machines, medical imaging and machine vision. The stack runs Linux with ROCm, PyTorch, ONNX and TensorFlow, plus Xen virtualization for mixed criticality, and the CPU headroom runs perception, planning and control beside inference. AMD's own testing configured a consumer Max+ 395 to X199 settings against Intel and Jetson class hardware, so treat vendor numbers as indicative rather than guaranteed.

Long availability is the second reason to choose embedded over client silicon. The 10-year program with last time buy in 2037 lets certified equipment stay in production without requalification, which dominates lifetime cost in medical and industrial markets. Commercial temperature grade fits indoor cabinets, and the industrial i variant covers the rest.

Upgrade paths

A soldered flagship has no drop-in upgrade. If 16 cores prove too many, the 12-core X188 is the step down on the same package and memory design. If the design later needs still more, there is no bigger Strix Halo die, so growth means a second board or a future platform such as the Gorgon Halo refresh.

Leaving FP11 means leaving Halo: the FP8 P100 series covers 4 to 12 smaller cores for lower power, and desktop Gorgon Point variants live on AM5. Memory size deserves early attention in particular, because soldered LPDDR5X cannot grow after manufacture.