AMD published the article on September 24. Quenton Hall in the embedded product group wrote it. AMD first named the Ryzen AI Embedded X100 family on July 23. Samples went to customers in June. Volume chips are still planned for October through December (AMD, 2026).
Think of X100 as a laptop-class processor built to sit in a sealed box on a factory floor. The top part has 16 main processor cores, a strong graphics block, and a smaller block that is good at running models. All three share one memory pool of up to 128 gigabytes. Some versions are rated from minus 40 to 105 degrees Celsius and for ten years of all-day use. Module partners include Arbor, Congatec, iBase, IEI, Sapphire, and Seavo.
I wrote about mimik in May next to Sony and TSMC cameras and again next to Xiaomi phones that already carry their own models. Fay Arjomandi's company treats nearby devices as places work can run without sending every request to a distant cloud. The September 24 article is AMD attaching that idea to a named factory chip and a count of how many small workers one box can hold.
The camera box now has to do more than mark a defect
Hall's factory example is easy to picture. The old box takes a photo, scores it, and raises a flag. The newer box is supposed to spot the fault, look at earlier steps on the line, message the technician who owns that station, and recommend a pause. Each of those steps is a small software worker. A plant can put hundreds of them on one computer, each watching one sensor or one rule. They wake when something happens, fetch what they need, call a tool, and pass the result along.
The picture model still runs on the graphics and model blocks. The main processor does the traffic work: which worker runs next, what data it may see, and what happens when a step fails.
mimik's operating software, sold here as mimOE Embedded Edition, keeps one copy of a model in memory and lets many workers use it. Memory then grows with how many different models you load, not with how many workers you start. AMD says that pairing, in a mimik study of 455 job mixes, held up to 2.3 times more workers at once than NVIDIA Jetson T5000-class boards. AMD paid for the paper. mimik published it on July 23. It is a model of spare processor, graphics, and memory room. It is not a published result from a live plant (mimik, 2026; AMD, 2026).
The score on the picture is only one step. The plant also pays for the messages and the decision that follow it.
One memory pool keeps the photo and the next action close
Those workers keep a lot of working notes on the box: the photo, the prompt, the tool reply, the log. If the main processor and the graphics chip keep separate memory, the box has to copy that data back and forth. X100 uses one shared pool. AMD says that pool can move up to 273 gigabytes a second, and that a memory test ran 1.3 times faster than an Intel Core Ultra Series 3 part set up as a stand-in (AMD, 2026). That test measures memory speed. It does not measure factory workers. It does show why AMD wants the decision software and the picture model on the same chip.
Hall also leans on familiar PC software. Most tools a worker will call still ship first for the same processor family used in office PCs. AMD says an internal conversion test kept 83 percent of a set of NVIDIA graphics programs when they were moved to AMD's own tools. A team still has to test the real application on the real board. The claim for a factory designer is that you can add these workers without rewriting the whole control program.
Ask how many workers the box must hold next year
Buyers still arrive with a sheet that only rates how fast the box can score pictures. Hall is asking them to count the workers, the models that stay loaded, and the other software that already runs on that computer. You can use that question on any vendor. A board that is strong at one large vision model can stall when the main processor has to run a hundred small jobs. An X100 box can stall too if every group loads its own copy of the model.
I expect more chip companies to ship this kind of software with the silicon. mimik already sells that layer. If plants start buying boxes by how many workers they can run, Qualcomm, NVIDIA, and Intel will not leave that layer to someone else. Software on the chip is how a decision happens while the part is still on the belt.
Count the workers you want on one cabinet in 18 months, including the message to the technician. Ask the supplier whether those workers share one copy of the model. The 2.3 times figure is a paid study, not a floor result. Run your own mix on the module you will buy, at the temperature of that room, before you change the parts list. Volume chips are still due in the fourth quarter.
Hall, Quenton. "Beyond TOPS: AMD Ryzen AI Embedded X100 at the Edge." AMD, 24 Sep. 2026, https://www.amd.com/en/developer/resources/technical-articles/2026/amd-ryzen-ai-embedded-x100-at-the-edge.html.
AMD. "AAI 2026: AMD Delivers Leadership Heterogeneous Compute for Physical AI." 23 Jul. 2026, https://newsroom.amd.com/news/aai-2026-ryzen-ai-embedded-x100/.
mimik. "Architectural Fit for Production-Scale Agentic AI on Heterogeneous SoCs." 23 Jul. 2026, https://www.mimik.com/agentix-compute-benchmarking.
Bellamkonda, Shashi. "The Missing Layer: Sony and TSMC Complete the Physical AI Stack." shashi.co, 8 May 2026, https://www.shashi.co/2026/05/the-missing-layer-sony-and-tsmc.html.
Bellamkonda, Shashi. "The Xiaomi Stack: When the Phone Knows the Model and Makes the Chip." shashi.co, 25 May 2026, https://www.shashi.co/2026/05/the-xiaomi-stack-when-phone-knows-model.html.
