AMD's Ryzen AI Developer Center: Local AI Setup Without the ROCm Grind

AMD's Ryzen AI Developer Center, bundled with Ryzen AI Max Plus 395 machines, eliminates manual ROCm and driver setup for local AI workloads, offering guided playbooks for ComfyUI, LM Studio, and Unsloth fine-tuning. The unified 128 GB LPDDR5X memory pool enables running large models like GPT OSS 120B at around 30 tokens per second, with dense 9B models reaching nearly 40 tokens per second. A Pro 395 variant with up to 192 GB memory has been announced for larger models.

AMD's Ryzen AI Developer Center: Local AI Setup Without the ROCm Grind AMD's Ryzen AI Developer Center ships with playbooks for ComfyUI, LM Studio, and Unsloth fine-tuning, cutting out manual ROCm and driver setup. What is the AMD Ryzen AI Developer Center? The AMD Ryzen AI Developer Center is a piece of software bundled with AMD’s Ryzen AI Max Plus 395 based machines marketed as Ryzen AI Halo that handles the setup work local AI builders usually do by hand. Instead of partitioning a Linux drive, installing ROCm, and matching driver versions to PyTorch builds, the machine ships with the stack already configured, and the developer center gives you a dashboard to see what’s installed, adjust memory allocation between CPU and GPU, and launch guided “playbooks” for tasks like image generation, local LLM serving, and fine-tuning. TL;DR - The Ryzen AI Max Plus 395 chip pairs 16 CPU cores with a Radeon 8060S GPU around 60 teraflops FP16 and an NPU near 50 TOPS, all sharing one pool of memory instead of using a discrete GPU with separate VRAM. Unified memory means the system can allocate up to 128 GB of LPDDR5X between CPU and GPU tasks, letting users load models like GPT OSS 120B or larger Qwen 3 variants that wouldn’t fit on a 32 GB discrete card.- The developer center’s out-of-box setup removes the need to manually install ROCm, configure drivers, or partition drives, a common pain point on prior AMD Linux AI workstations. Playbooks are organized by skill level: beginner guides cover ComfyUI and LM Studio, intermediate ones cover fine-tuning with Unsloth, and advanced ones cover multi-machine clustering and writing custom GPU kernels.- Real-world testing showed dense 9B models running near 40 tokens per second, a mixture-of-experts model Qwen 3.6, 3B active parameters exceeding 50 tokens per second, and GPT OSS 120B running around 30 tokens per second, speeds workable for chat, RAG, and lighter agent tasks. - A Pro 395 variant with up to 192 GB of memory has already been announced by AMD, aimed at users who want even more headroom for large models. - The software supports remote access via SSH , so the machine can run headless as a dedicated inference box while being driven from another computer. - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. How does unified memory change local AI setup? Traditional discrete GPU setups split memory into two pools: fast VRAM on the card and slower system RAM. Once a model exceeds VRAM capacity, layers get offloaded to system RAM, and token generation speed drops sharply. A 32 GB card, for instance, simply can’t hold a 120 billion parameter model no matter how aggressively it’s quantized. The Ryzen AI Max Plus 395 avoids this by giving the CPU and GPU access to the same 128 GB pool of LPDDR5X memory. Users can decide how much of that pool goes to the GPU versus the CPU. Because most AI applications don’t need much system RAM, the majority of the pool, in testing referenced at 75%, can be dedicated to the GPU for loading models. That’s what makes running very large models locally feasible on a machine this size, something a 32 GB discrete card architecture can’t match regardless of quantization tricks. The tradeoff is that this unified GPU won’t match a high-end discrete card’s raw VRAM bandwidth. But for a wide range of workloads, from chat models to image and video generation, it’s fast enough to be usable, and it opens the door to models that simply couldn’t run at all before. What do the playbooks actually cover? The playbooks are organized as a tiered library of guided setups, viewable by operating system Windows or Linux and by difficulty: Beginner playbooks cover getting ComfyUI running with pre-installed models, and serving LLMs through LM Studio. These are built so the ROCm version and driver compatibility are already handled, meaning tools “just work” without version hunting. Intermediate playbooks include fine-tuning with Unsloth, a widely used tool for running basic fine-tunes on open models. These playbooks also cover how to configure graphics memory allocation for different kinds of training or inference tasks. Advanced playbooks cover more specialized work: clustering multiple Ryzen AI Halo machines together, and writing custom GPU kernels, complete with a walkthrough of what the kernel code does and how to configure the environment to run it. The library isn’t exhaustive. Advanced playbooks are still sparse compared to the beginner tier, but the coverage already spans the tasks most people buying a machine like this actually want to do: run local LLMs, generate images and video, and fine-tune models without fighting the underlying stack. Is the Ryzen AI Halo actually fast enough for real work? Speed depends heavily on model architecture and size. In testing, a dense 9B model with a multi-token-prediction drafter ran at just under 40 tokens per second. Switching to a mixture-of-experts model Qwen 3.6 with roughly 3B active parameters and a drafter pushed speeds above 50 tokens per second, since MoE architectures only activate a fraction of total parameters per token. GPT OSS 120B, a much larger mixture-of-experts model that comes pre-installed on the machine, ran around 30 tokens per second. Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. For context, 30 tokens per second is workable for chat and retrieval-augmented generation, though it’s on the slower side for latency-sensitive agent workflows, where a smaller, faster model is often the better fit. None of this matches a high-end discrete GPU workstation on raw speed, but the point isn’t raw speed. It’s the ability to load models that a 32 GB discrete card physically cannot hold, and still get usable throughput. Image and video generation also proved workable through ComfyUI, which comes pre-installed with a model called Image Z. In one test using the Qwen image model alongside pose-based control nets, individual image renders took roughly 19 to 20 seconds, fast enough to iterate through many prompt variations in a single session rather than requiring an overnight batch run. Is the developer center worth it compared to manual setup? For anyone who has manually configured ROCm, matched PyTorch versions, and debugged driver conflicts on a Linux AI box, the value proposition is straightforward: the developer center removes that entire process. The machine arrives with the stack pre-installed, a dashboard for adjusting memory allocation between CPU and GPU, and a curated set of walkthroughs for common tasks. It won’t cover every possible workflow out of the box, and advanced playbooks remain limited. But for the common cases, running local LLMs, generating images, doing basic fine-tunes, it turns what used to be hours of environment setup into a guided, repeatable process. That matters most for people who want to spend their time building on top of the stack rather than debugging it. Frequently Asked Questions What chip powers the AMD Ryzen AI Halo? It runs on the Ryzen AI Max Plus 395, which combines 16 CPU cores, a Radeon 8060S GPU rated around 60 teraflops at FP16, and an NPU rated near 50 TOPS. How much memory can the Ryzen AI Halo allocate to the GPU? The system has 128 GB of unified LPDDR5X memory shared between CPU and GPU, with the split configurable by the user. A Pro 395 variant supporting up to 192 GB has also been announced. Does the developer center support remote use? Yes. It supports SSH-based remote access, so the machine can run headless as a background inference server while being controlled from another computer. Is ROCm still something users need to install manually? No. The developer center machine ships with ROCm and PyTorch already configured, and the dashboard shows installed versions rather than requiring manual installation. Can the NPU be used for LLM inference today? Not widely yet. Most current LLM stacks and applications primarily use the GPU rather than the NPU, though that’s expected to change as software support matures.