SANTA CLARA, September 25 — AMD says it has built the first x86 client processors that can run a 300-billion-parameter language model entirely on a single machine, with no cloud connection at all. The chips are the Ryzen AI Max PRO 400 Series, and the company’s answer to a question plenty of developers have run into: what happens when the model you actually want to use is too big to fit on the hardware you can afford? AMD’s bet is that the fix isn’t a faster GPU.
It’s more memory.
One pool of memory, shared by everything
Here’s what it means in plain terms. A typical high-end gaming graphics card carries its own dedicated pool of video memory, usually capped at 16GB or 24GB. A large language model has to fit inside that pool to run locally, which is why so many people end up working with smaller 7B or 13B models instead of frontier-scale ones.
AMD’s design takes a different route. The PRO 400 Series uses a unified memory architecture, sharing one large memory pool among the CPU cores, an integrated GPU, and a neural processing unit — all on the same chip package.
Think of it like a single pantry the whole kitchen can reach into, rather than separate cupboards each cook has to stock alone. Both models in the series, the Ryzen AI Max PRO 485 and the Ryzen AI Max+ PRO 495, list an identical maximum memory figure of 192GB, with up to 160GB configurable as GPU VRAM. That’s the number doing the heavy lifting behind AMD’s 300-billion-parameter claim.
What AMD is actually arguing
The company’s position is blunt: memory capacity, not raw GPU horsepower, is the primary bottleneck for local AI inference. If that’s right, the practical upshot is that a workstation chip with a big shared pool can handle models that would choke a more powerful discrete card with a smaller dedicated one. The PRO 400 Series builds on AMD’s earlier Ryzen AI Max chips, known by the code name Strix Halo, but raises the maximum memory ceiling and adds “PRO” manageability and security features aimed at commercial and enterprise buyers.
ASUS, HP, and Lenovo are expected to ship systems based on the new silicon in the third quarter of 2026, according to AMD’s announcement. The target audience is specific: developers, researchers, and enterprise IT departments that want to run frontier-scale models locally rather than renting cloud GPU time or settling for smaller 7B or 13B models.
AMD is positioning the line for business and enterprise use, not the consumer market.
The open questions
AMD has not disclosed detailed specifications such as clock speeds, core counts, or TDP for the new chips. Independent benchmarks and real-world performance data have not yet been released either. That leaves memory bandwidth, actual inference speeds, power consumption, and pricing all unresolved.
The 300-billion-parameter claim rests on memory capacity, and actual performance will depend on factors including memory bandwidth and model optimization. The announcement positions the PRO 400 Series against competing local AI hardware efforts, including Nvidia’s DGX Spark and offerings from Qualcomm and Apple in the broader AI PC race.
The success of AMD’s approach will depend on software support, developer adoption, and whether unified memory can deliver competitive performance for large-model inference. What to watch next is straightforward. The third-quarter systems from ASUS, HP, and Lenovo will be the first real test, and the independent benchmarks that follow will show whether a big shared memory pool translates into the kind of local inference speed enterprises are willing to pay for.
