# AMD Teases Ryzen AI Halo, a ROCm Dev Mini-PC With Windows and Linux, Launching Q2 2026

> AMD previews Ryzen AI Halo, a compact ROCm dev box on Strix Halo with Zen 5 and RDNA 3.5 iGPU-up to 60 TFLOPS. Ships Q2 2026 with Windows or Linux and up to 128GB LPDDR5X.

- Type: AI news
- Page: https://completeaitraining.com/news/amd-teases-ryzen-ai-halo-a-rocm-dev-mini-pc-with-windows/
- Published: 2026-01-07
- Updated: 2026-02-19
- Author: Jeroen Erne
- Jobs: IT and development
- Reading time: 3 min

### AMD teases Ryzen AI Halo: a compact ROCm-first box for AI developers

At CES 2026, AMD previewed the Ryzen AI Halo, a high-end mini-PC aimed at teams building and testing models on the ROCm stack. A full launch is planned for Q2 2026, with AMD positioning it as a standard, ready-to-go dev box inside the AMD ecosystem.

The headline: it's built on Ryzen AI Max+ (codename Strix Halo). That means Zen 5 CPU compute paired with an integrated RDNA 3.5 GPU and a wide LPDDR5X memory bus for serious local workloads.

#### What we know so far

- **APU:** Ryzen AI Max+ (Strix Halo) with Zen 5 CCDs and a high-performance I/O die.
- **GPU:** Integrated RDNA 3.5 Radeon with 40 CUs, quoted at "up to" 60 TFLOPS.
- **Memory:** "Up to" 128GB LPDDR5X on a 256-bit bus for high bandwidth and large model footprints.
- **SKUs:** The "up to" language suggests lower-cost, lower-spec options will be offered.

That combination-high bandwidth iGPU + large shared memory-has already proven useful for local LLM work with current partner-built systems. Halo looks like AMD's official, standardized take on the same formula.

#### Software environment

- **ROCm-first:** AMD says the Halo will ship pre-loaded with ROCm-optimized models and apps.
- **OS options:** Both Windows and Linux are supported. That's a notable shift from the typical Linux-only dev boxes.
- **Day 0 models:** AMD is committing to day 0 support for leading models, which implies first-class ROCm support for RDNA 3.5 (different from the CDNA architecture used in Instinct accelerators).

Whether AMD rolls a custom Linux image or layers ROCm on a mainstream distro is still an open point. Either way, expect a turnkey setup focused on fast iteration.

#### Networking and scaling

NVIDIA's DGX Spark made 200GbE a headline feature using an integrated ConnectX-7 NIC, enabling small clusters and scale-out testing. AMD has the Pensando Pollara NIC in its portfolio, but hasn't confirmed if Halo includes high-speed networking by default.

If you're planning multi-node training or distributed inference tests, this detail will matter. It could also be a major price swing.

#### How it compares to NVIDIA's developer boxes

Conceptually, Halo mirrors the DGX "developer box" pattern: standardized hardware, tuned software, fast start. The key differences will be the networking story and how well RDNA 3.5 iGPU paths are optimized across common frameworks under ROCm.

- DGX systems emphasize out-of-the-box scale-out and NV/NIC integration.
- Halo emphasizes shared-memory iGPU performance and ROCm coverage on Windows and Linux.

For context on the NVIDIA side, see [NVIDIA DGX](https://www.nvidia.com/en-us/data-center/dgx/?ref=completeaitraining.com). For AMD's software stack, see [AMD ROCm](https://www.amd.com/en/technologies/rocm?ref=completeaitraining.com).

#### Who should care

- Teams prototyping LLMs locally, where 64-128GB shared memory and high bandwidth matter more than discrete GPU VRAM.
- Engineers targeting ROCm for deployment or cost reasons, who want a reference box with predictable behavior.
- Windows-first devs who've been blocked by Linux-only hardware kits.

#### Open questions to watch

- **Networking:** Will there be 100/200GbE options for cluster tests?
- **SKU tiers:** What are the concrete CPU/GPU clocks and memory sizes per tier?
- **Thermals/acoustics:** Can the box sustain "up to 60 TFLOPS" under load without throttling?
- **Software image:** Custom AMD Linux vs. a standard distro with ROCm pre-loaded?
- **Framework coverage:** How smooth is PyTorch/TensorFlow/ONNX Runtime on RDNA 3.5 via ROCm out of the box?

#### Practical next steps for engineering teams

- List your core workloads (token throughput targets, context lengths, batch sizes) and map expected memory needs to 64GB vs. 128GB.
- Decide if you need high-speed networking for multi-node tests; if yes, wait for AMD's NIC details.
- Validate your stack on ROCm now (kernels, graph compilers, custom ops) to reduce migration friction.
- Plan for dual-boot or containerized dev if your team splits across Windows and Linux.

If your team needs structured learning paths to get productive on ROCm and modern AI workflows, consider the [AI Learning Path for Training & Development Managers](https://completeaitraining.com/learningpath/ai-for-training-and-development-managers/) at Complete AI Training.

#### Timeline

AMD says the Ryzen AI Halo will get a full reveal and launch in Q2 2026. With Ryzen AI Max+ hardware already shipping in volume via partners, expect a shorter ramp from announcement to availability.

---

This is the Markdown form of https://completeaitraining.com/news/amd-teases-ryzen-ai-halo-a-rocm-dev-mini-pc-with-windows/. The whole site is mapped at https://completeaitraining.com/llms.txt.
