AMD Launches Ryzen AI Halo Mini PC With 192 GB of Memory for $7300

AMD has priced its reference Ryzen AI Halo mini PC at $7299.99, nearly double the price of its predecessor. The machine, built on the Ryzen AI Max+ Pro 495 (codename Gorgon Halo), is already on sale in US retail. The case, just over a liter in volume, holds 192 GB of LPDDR5X-8533 unified memory and a 2 TB SSD. Doing the math naively, that comes to about $7300 per liter of volume.

What exactly AMD is selling

Ryzen AI Halo is positioned as a developer platform, not a home computer. It is a reference system. AMD uses it to show what a workstation for running models locally on its hybrid processors should look like, and it sells the machine ready-made. The main feature is the memory architecture: the CPU cores and the graphics share a single pool. The GPU can claim an amount of memory that owners of consumer graphics cards, with their 16–32 GB, can only dream of.

Key specs from the announcement:

• Processor: Ryzen AI Max+ Pro 495 (Gorgon Halo)
• Memory: 192 GB of unified LPDDR5X-8533
• Storage: 2 TB SSD
• Case: just over 1 liter in volume
• Price: $7299.99, sold in US retail

Why memory matters more than clock speeds

Two parameters decide how well local language models run: how many weights fit in memory, and how fast they can be read. The previous Ryzen AI Max+ 395 topped out at 128 GB. Now the ceiling is 192 GB. As a rough estimate, a 4-bit quantized model with 300 billion parameters takes up about 150 GB, which leaves headroom for context and the system. This is an estimate, not a measurement, and the real headroom depends on the quantization format and the context length.

Bandwidth is trickier. AMD does not state it in the material. If the bus is still 256-bit, as on the predecessor, LPDDR5X-8533 delivers about 273 GB/s. For a dense model weighing 100 GB, that puts the ceiling at roughly 2–3 tokens per second during generation. Mixture of Experts architectures, where only a portion of the weights is active for each token, fare noticeably better. That is where this kind of memory pays off.

Price and competitors

The near doubling of the price looks steep. The source data does not give the reasons. The obvious candidate is the larger memory capacity, but that is a hypothesis, not a confirmed fact. Buyers have two alternatives: build a tower with several discrete GPUs, or buy a compact system with a single memory pool. The first offers more compute but takes up more space, uses more power and is noisy. The second wins on capacity per watt and per liter.

Who it suits and what to check before buying

The platform makes sense for people who build agents and applications on top of large open models and do not want to send data to the cloud. Another scenario is testing on unified memory before deploying to servers. For everyone else, $7300 is too much.

Before ordering, check four things:

• Real generation speed on your own models. Factory benchmarks with short prompts say little.
• The state of the software stack. Support for ROCm and popular inference engines on this chip is critical.
• Cooling and noise. A one-liter case with a powerful APU typically runs into thermal limits under sustained load.
• Availability outside the US. For now, only US retail is mentioned, and prices and delivery times in other regions are unknown.