AMD Ryzen AI Max 390 socket: FP11

What the Ryzen AI Max 390 is

The Ryzen AI Max 390 is a 12-core, 24-thread mobile APU that AMD launched in January 2025 in the middle of the consumer Strix Halo lineup. Its twelve uniform Zen 5 cores run at 3.2 GHz and boost to 5.0 GHz, backed by 12 MB of L2 and a large 64 MB of L3 cache. Graphics come from AMD Radeon 8050S Graphics with 32 compute units at up to 2800 MHz, and an XDNA 2 NPU rated at up to 50 TOPS covers local AI work.

The 390 is the balanced Halo. It matches the core count, clocks and cache of its PRO twin, the AMD Ryzen AI Max PRO 390, while leaving fleet management features to the commercial stack. Against the 16-core flagship above it, the 390 gives up four CPU cores and 8 graphics compute units, which lowers cost and cooling demand while keeping most of the multithreaded pace. Total platform AI throughput is rated at up to 110 TOPS across CPU, GPU and NPU.

Strix Halo is a chiplet design. Two Zen 5 core chiplets sit beside a large I/O die that holds the graphics engine, the NPU and the 256-bit memory controller, and AMD builds the compute silicon on the TSMC 4 nm FinFET process. The system maker sets power and cooling inside a 45 to 120 W configurable window, with 55 W as the default TDP for this SKU.

Buyers reach the 390 when twelve threads of headroom matter but the flagship GPU is more than the display needs. Creation tools that split across cores use the full thread count, while games run on the 32 compute unit GPU fed by wide LPDDR5X. The January 2026 refresh later added a 12-core sibling with the full 40 compute unit GPU for buyers who want the same CPU with wider graphics.

Socket and motherboard compatibility

AMD ships the Ryzen AI Max 390 only as a soldered part for FP11, which means the processor arrives fixed to the board and it cannot be replaced or upgraded in the field. Assembly solders all 2,077 balls of the 37.5 × 45 mm package in one factory step, and the thermal design is validated for that single configuration through the service life of the machine.

There is no board or chipset choice in the retail sense. The vendor selects the soldered memory capacity, tunes the firmware and validates cooling for the 45 to 120 W window. Moving to 16 cores or down to 8 cores means a different board build, not a part swap, even though the Strix Halo SKUs share the FP11 package. Readers new to soldered packages can review LGA, PGA and BGA: CPU package types explained for how a ball grid array differs from a socketed desktop part.

The FP11 footprint is shared with the commercial branch of the same silicon, including the PRO twin AMD Ryzen AI Max PRO 390 and long lifecycle embedded parts such as the AMD Ryzen AI Embedded X188i. The smaller FP8 package covers lower power laptops with narrower memory buses. Halo boards stay distinct from those mainstream designs by memory width, package size and power class.

Memory, PCIe and cooling

Memory is soldered LPDDR5X up to 8000 MT/s on a 256-bit bus, with up to 128 GB in one unified pool shared by the CPU cores, the 32 compute unit graphics engine and the NPU. Up to 96 GB of that pool can be reserved for graphics workloads on the 300 series platform. Capacity and speed are fixed when the board is built, so they rank with the SKU as permanent choices that deserve planning before purchase.

Expansion comes from 16 PCIe Gen4 lanes from the SoC, which typically serve NVMe storage and high speed peripherals, plus native USB4 ports for docks and external devices. Display and parallel compute run on the integrated Radeon graphics, so the lanes stay free for storage and I/O duties. The exact lane split and port list are set per machine design, so the vendor specification governs.

Cooling is a system responsibility across the 45 to 120 W configurable range. Twelve Zen 5 cores plus 32 graphics compute units make this a demanding SKU to cool in a thin chassis, so chassis choice decides how long the chip holds peak clocks under combined load. There is no boxed cooler and no retail heatsink list for a soldered mobile APU.

Upgrade paths

There is no CPU upgrade path on a soldered package. When twelve cores stop being enough, the step inside Strix Halo is a different class of machine with a wider CPU and GPU, since this SKU sits in the middle of the stack. A board replacement is the only route, and it carries new memory and firmware with it.

Moving off FP11 means a new platform. When the workload shrinks, machines built around the 8-core class of the same silicon keep the memory platform with fewer threads. Socketed desktop builders meet related Zen 5 silicon on boards such as AM5. 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 rather than at the end.