It's been over a year since the launch of the RTX 50 and RX 90 series GPUs, and 2026 is set to end with no new GPU launches. We have the global data center demand and consequent memory shortage to thank for it. So, it's not surprising that the most interesting piece of hardware from Team Red this year is not a GPU but an AI mini PC. The Ryzen AI Halo Developer Platform is a $3,999 compact workstation powered by the Ryzen AI Max+ 395, the most powerful consumer APU on the market, at least in the x86 world. For this price, you're looking at a 16-core, 32-thread processor, a 40-CU RDNA 3.5 iGPU, and up to 128GB of unified memory. This Strix Halo mini PC is expressly targeted at making local AI workloads on AMD hardware significantly faster and more seamless. With its 128GB of unified memory, this mini PC can beat fan-favorite local AI GPUs like the RTX 3090 and even the RTX 4090 and RTX 5090 when it comes to chewing through massive LLMs. It does have to compete with Nvidia's DGX Spark, but it has price and availability on its side, although AMD's ROCm platform still lags behind Nvidia's CUDA.

The Ryzen AI Halo Developer Platform is a local AI beast

High-end LLM development in a box

AMD's most powerful mini PC might not be as interesting as a new RX 90 series GPU to most consumers, but it excels at what it's made for. Packed with the Strix Halo flagship in the form of the Ryzen AI Max+ 395, Radeon 8060M graphics, 80MB of combined cache, a 650 TOPS NPU, and 128GB of unified memory, the Ryzen AI Halo Developer Platform isn't kidding around. As of now, this $3,999 SKU is the only one AMD is offering, but more affordable models will be available later. Consumers who don't want a $3,999 mini PC for local AI can look to AMD's Strix Point or Gorgon Point laptops for more reasonably priced options.

The Strix Halo mini PC will absolutely devour most large models you can throw at it. The 128GB of unified memory can run 200B parameter models, embarrassing even the mighty RTX 5090 and its 32GB of VRAM. The memory runs at a theoretical maximum of 256 GB/s, which is much slower than what you see on top-end discrete GPUs. However, if you can't even fit a massive model on your GPU, the memory bandwidth advantage doesn't help you. You can also leverage the 40-CU iGPU and NPU on the Strix Halo mini PC for specialized workloads, and boot both Windows and Linux.

AMD aims to eliminate the software layer with this compact workstation, allowing users to avoid the lengthy setup usually associated with AMD's ROCm stack for local AI. The Ryzen AI Developer Center software environment is built into the mini PC, making AI playbooks and validated model packages available on first launch. The objective is to enable local AI professionals to test, run, and develop LLMs without dependencies. As for the comparisons with Apple's unified memory machines, AMD was sure to highlight how its solution has more memory and broader support across Windows and Linux. In terms of running cost, you're looking at somewhere around $16 per month, and that's at a sustained power draw of around 150W when used for 8 hours a day. AMD made it clear that this mini PC makes sense only for serious local AI professionals, but it does make a pretty strong case within that niche.

Nvidia's DGX Spark had better watch out

A more interesting Nvidia vs. AMD battle

AMD is positioning the Ryzen AI Halo Developer Platform as a high-end local AI workstation, similar to Nvidia's DGX Spark mini PC. The latter is powered by Nvidia's GB10 chip featuring a 20-core Arm CPU, 6144 CUDA cores, 128GB of unified memory running at 273 GB/s, and up to 1 PFLOP of performance. The two competing compact PCs are mostly well-matched, with performance likely to favor one or the other, depending on the specific workload, although AMD's dev box is currently $700 cheaper. Nvidia had earlier announced the DGX Spark for the same $3,999 price, but later increased it to $4,699, making AMD the "value" option yet again.

The great thing for AMD is that the company has already shown that its Gorgon Point flagship, i.e., the Ryzen AI Max+ Pro 495, will surpass the Strix Point flagship. It will offer up to 192GB of unified memory and will be capable of running 300B parameter models. Manufactured with the same Zen 5 architecture, the Gorgon Halo flagship will deliver even more impressive results later this year. However, Nvidia's DGX Spark does hold a slight advantage against AMD's current AI box in terms of memory bandwidth, and its tensor cores are likely to perform better in raw compute and prompt processing. The fact is that local AI professionals can't really go wrong with either machine, but the decision will still hinge on AMD's software stack, which is where things get more complicated.

AMD's AI mini PC beats high-end GPUs, but ROCm still needs work

The gap is not as huge anymore

Whether you choose Nvidia's DGX Spark or AMD's Strix Halo mini PC might depend on the CUDA vs. ROCm debate. For the longest time, "Just buy Nvidia" used to be true for local AI workloads, thanks to CUDA being a more mature platform and the preferred choice for new models and software packages. ROCm had the reputation of getting delayed support and being riddled with more tinkering than CUDA. However, users who don't mind a bit of tinkering, llama.cpp supports both Vulkan and ROCm backends, and AMD fully supports Ollama, LM Studio, and ComfyUI. The setup that used to take a weekend on AMD hardware just works now with a simple installation.

While ROCm has improved substantially with PyTorch 2.12 and AMD's HIP, rough edges still exist. The ROCm development that works so well on AMD's Instinct GPUs in data centers needs to perform just as well on consumer iGPUs like Strix Halo. Quantization libraries are not CUDA-exclusive anymore, but CUDA is still the default, with ROCm support arriving later, and requiring more manual intervention. Even on Strix Halo hardware, Ollama can still hit a timeout when searching for the GPU, forcing you to manually tell it where to look. While ROCm has stopped being a dealbreaker for local AI, it's still not CUDA. Nvidia's hardware might be more expensive, but many professionals might just skip the manual tinkering required on AMD.

AMD's high-end AI mini PC has solid potential, and ROCm might be the only hurdle

While Nvidia's DGX Spark arrived first, AMD's Ryzen AI Halo Developer Platform is currently cheaper for equivalent hardware. It supports almost every local AI tool most people actually use, and ROCm has come a long way. AMD's Gorgon Halo will further bolster its local AI capabilities in the high-end market. What remains is the gap between ROCm and CUDA, and closing it might take AMD a bit longer than it hopes.