The first desktop systems built on AMD's Gorgon Halo chip have arrived, offering up to 192 GB of unified memory for running large AI models locally, but at prices well above the previous generation, The Register reported Tuesday. GMKtec's EVO-X5 Pro, among the first to market, starts at $6,799.

What Happened

Gorgon Halo is the codename for the Ryzen AI Max+ 495, a system-on-chip that The Register described as essentially a factory-overclocked version of AMD's Strix Halo. It keeps the same basic layout: 16 Zen 5 CPU cores, an integrated GPU with 40 compute units and an XDNA 2 neural processing unit, with clock speeds raised by about 100 MHz across the board.

The bigger change is memory. Gorgon Halo supports up to 192 GB of LPDDR5x at 8,533 MT/s, up from Strix Halo's 128 GB at 8,000 MT/s. That capacity is shared between the CPU and GPU, which is what makes these machines attractive for running large language models.

RelatedGermany's BSI Urges Small Plant Operators to Take Controllers Offline After French Fridge Attack

At 4-bit precision, The Register estimated the system can run models of roughly 345 billion parameters under Linux and about 320 billion under Windows, where the GPU memory allocation tops out at 160 GB. That is enough for open-weight models such as DeepSeek V4 Flash, with 284 billion parameters, and Z.AI's GLM-5.3-Flash, with 320 billion.

Why It Matters

Capacity determines which models fit; bandwidth determines how fast they run. Text generation in large language models is largely limited by memory speed, and Gorgon Halo's bandwidth of 273 GB/s is only about 6.5% higher than Strix Halo's 256 GB/s. Buyers get the ability to load much larger models, but token generation speeds will improve only marginally, according to The Register.

Price is the larger obstacle. A year ago, a 128 GB Strix Halo box cost between $2,000 and $3,000, and by this summer that had risen to about $4,000, The Register said. The Register put GMKtec's starting price for the 192 GB model at about 64% above the earlier generation, and it expects premium brands such as Framework and HP, with its ZBook Ultra G3a, to charge more.

The Memory Squeeze

The price increases trace back to the AI data center boom. LPDDR5x is in short supply, and Nvidia's rack-scale systems consume it in bulk: each NVL72 rack contains between 36 and 54 TB of the memory, according to The Register. Samsung has warned that supply will remain constrained through 2028, and The Register expects the squeeze to worsen as AMD adopts the same memory type more widely.

Nvidia's competing desktops are feeling the same pressure. Workstations based on its GB10 chip have roughly doubled in price over the past year, from $3,000 to $4,000 to between $6,000 and $8,000. Its RTX Spark systems for Windows, now launching from major PC makers as mini PCs and notebooks, top out at 128 GB of memory, giving Gorgon Halo an edge in capacity.

What to Watch

Framework and HP are expected to bring Gorgon Halo systems to market, and their pricing will show how much of the memory shortage vendors pass on. For small teams that bought unified-memory mini PCs to run open-weight models privately, the question is whether the ability to load 300-billion-parameter models justifies paying far more for a machine that is only marginally faster.