AMD Ryzen AI Max 400: New APU Unlocks 192GB Unified Memory

AMD had a number of announcements today in the lead-up to Computex, but the one we're most excited about is the announcement of the Ryzen AI Max 400 series processors, code-named Gorgon Halo. "Excited" might be the wrong word; these are, fundamentally, a refresh of the Ryzen AI Max 300 series processors. Those chips are still pretty dang cool, though, and so it goes that the Ryzen AI Max 400 series is pretty cool too.


Strictly speaking, what AMD's announcing today is the Ryzen AI Max PRO 400 series; the non-PRO versions are "coming soon." However, as is typically the case with AMD's PRO processors, we expect that the "amateur" versions of the chips (because they're not "PRO," see?) will be identical to these PRO models. So there's three SKUs this time, just like with the launch of the Ryzen AI Max 300 series: the Max+ PRO 495 with 16 Zen 5 CPU cores and 40 RDNA 3.5 GPU cores, the Max PRO 490 with twelve CPU cores and 32 GPU CUs, and then the Max PRO 485 with eight cores and the same Radeon 8050S Graphics as the 490.


Specifications-wise, there are almost no changes from the 300 series; the top-end model gets 100 MHz of extra clock rate on both CPU and GPU. The biggest difference in these parts is actually to do with the platform; AMD has verified the Ryzen AI Max 400 series for up to 192GB of RAM. That's a 50% bump over the Ryzen AI Max 300 series, and while you might think it an irrelevant difference, AMD is pushing these chips hard as AI workstation processors, and the extra RAM is actually meaningful in that context.


Indeed, AMD says this is the "world's first x86 client processor to run a 300B+ LLMs". Grammar aside, AMD is really stretching the definition of "client" here considering we're talking about a chip that has sixteen full-fat CPU cores and a 256-bit memory bus. Still, the claim is legit, lthough we wonder what kind of performance you can really expect out of such a configuration; we didn't have great experiences testing AI on the Ryzen AI Max+ 395 in the past. That's been some eight months past now, though; AMD is said to have made major strides in its AI software support since then.


The chipmaker suggests that switching to a Ryzen AI Max series machine running a local LLM could potentially save you $750 a month versus running Claude Sonnet 4.5. The difference of a local model versus Claude Sonnet in terms of performance is notable, but many people have had good results running local models to power applications like OpenClaw, so AMD's probably onto something here.


The other big announcements today primarily concerned the AMD Ryzen AI Halo developer platform, which is essentially AMD's first-party answer to the NVIDIA DGX Spark. AMD won't prefer us to describe it that way, but that's exactly what it is. This machine is slightly under 1 liter in size, yet packs in a Ryzen AI Max+ 395 with 128GB of unified memory, Wi-Fi 7, a 2TB PCIe 4.0 SSD, and your choice of Windows or Linux, which is arguably a killer feature versus NVIDIA's Linux-only box. AMD also dunks on Apple's Mac Mini with the M4 Pro, noting that the smaller memory capacity on Apple's machine limits the kinds of workloads you can run on it.

The Ryzen AI Halo box is going to start at $3,999 USD, and pre-orders will open at some point next month; a version based on the Ryzen AI Max+ PRO 495 is also on the way at some point, although AMD didn't offer any indication as to when devices with those chips would be available. We'll likely see some announcements at Computex.