AMD ROCm with Ryzen 400 support
At CES 2026, AMD spoke about ROCm 7.2.2 as its next client-facing update, with added support for Ryzen AI 400 “Gorgon Point” processors. One slide also labels ROCm 7.2 as a common release for both Linux and Windows. Of course this shouldn’t come as a surpirse given that Ryzen 300 series is already supported and 400 series are nothing but a refresh.
For AMD, however, this update is important because it also means easier setup for local image generation. The company referenced ComfyUI integration, including a download path via ComfyUI.org, and introduced it as part of a broader push to make ROCm less “datacenter only” and more consumer-friendly. AMD said its supported Ryzen and Radeon product list doubled over the last year.
Source: AMD
On Windows, AMD focused on making ROCm a normal dev stack. The company called out WinML readiness, PyTorch on Windows, and an ONNX workflow for inference and training. AMD also referenced TheRock, its open build and packaging effort around HIP and ROCm components.
For performance, AMD used CES slides to show what it claims is the recent pace of improvement. One chart compares ROCm 7.1 against ROCm 6.4 and claims a 5x uplift for ComfyUI, plus model-specific gains such as 2.6x faster SDXL and 5.2x faster Flux S on Ryzen AI Max systems. Those figures were shown alongside ROCm 7.1.1 callouts on the demo slides, not as a direct ROCm 7.2.2 benchmark.
AMD also used the stage to argue that local inference is moving toward larger, “cloud-class” models on client hardware. Demo slides listed SDXL (6.6B parameters), Z Image (6B), and FLUX 2 (32B) running on Ryzen and Radeon platforms under ROCm 7.1.1. Another chart showed token throughput examples on Ryzen AI Max+ systems, including 4.9 tokens/s for Llama 3 70B and 48 tokens/s for “GPT-OSS 120B,” as labeled on AMD’s slide.
Now, if only everyone could afford a Ryzen MAX+ system, that would be even better. Meanwhile, Ryzen 400 series laptops (and apparently desktops) are set to launch later this year starting from this quarter.
Source: AMD
