AMD's 192GB 'Gorgon Halo' chip powers a $7,099 mini-PC for local AI

Why it matters: according to The Register, Gorgon Halo is essentially a factory-overclocked version of AMD's existing Strix Halo APU, with meaningfully more memory and a modest clock speed bump.

The chip specs break down as follows:

• CPU: up to 16 Zen 5 cores.
• GPU: a 40-compute-unit integrated graphics unit.
• NPU: an XDNA 2-based neural processing unit for AI acceleration.
• Memory: 192GB of unified LPDDR5x running at 8,533 MT/s, versus Strix Halo's 128GB at 8,000 MT/s.
• Clock speed: roughly 100 MHz higher than Strix Halo.
• Memory bandwidth: 273 GB/s, about 6.5% faster than Strix Halo's 256 GB/s.

Unified memory means the CPU, GPU, and NPU all share the same pool of RAM rather than each needing its own dedicated memory, which is part of what makes running very large AI models on a compact system possible in the first place.

The big picture: at 4-bit precision, a common way of compressing AI models to use less memory, the Evo-X5 can theoretically run models with up to 345 billion parameters on Linux, or 320 billion on Windows.

That range puts genuinely large open models within reach of a single desktop machine. The Register specifically points to DeepSeek V4 Flash, a 284-billion-parameter model, as one that fits comfortably inside that ceiling. For businesses or individuals who want to run a frontier-scale model without sending data to a third-party cloud API, that's a meaningful capability shift.

In real life a small business handling sensitive client data could run a large language model entirely on-premises on a machine like this, instead of routing every query through an external AI provider's servers.

The catch: The Register cautions that LLM inference is "predominantly memory bound," meaning the speed at which memory can move data matters at least as much as how much of it is available. Extra capacity without matching bandwidth gains doesn't automatically translate into proportionally faster performance.

Background: Strix Halo, the platform Gorgon Halo builds on, isn't new. AMD revealed the silicon in 2024 and released it across a wide range of laptops through 2025, marketed as Ryzen AI Max. AMD's own developer-focused mini-PC built around the chip, the Ryzen AI Halo, launched at $3,999.99 with 128GB of memory, aimed squarely at local AI development work.

That price point kicked off a wave of copycat hardware. "Every mini PC vendor on the planet is now slapping 'Ryzen AI MAX+ 395' on something," as one review put it. The GMKtec EVO-X2, an earlier 128GB model from the same maker behind the new Evo-X5, ranged from $2,349 for 96GB up to $3,299 for the full 128GB configuration, well under AMD's own reference price.

That earlier 128GB generation was already capable of real local AI work. A 128GB Strix Halo system could handle 70-billion-parameter models comfortably, along with select larger mixture-of-experts models and even a constrained, low-bit route to running DeepSeek V4 Flash. The jump to 192GB with Gorgon Halo extends that ceiling further, rather than opening up an entirely new category of capability.

By the numbers: a year ago, 128GB Strix Halo systems, the previous generation, cost between $2,000 and $3,000. By summer 2026, those same class of machines had climbed to around $4,000. The new 192GB Gorgon Halo systems starting near $7,000 represent roughly a 64% increase over where the market sat just months ago.

That's not an AMD-specific problem. LPDDR5x memory remains in short supply industry-wide, squeezed by surging demand from AI data center infrastructure. Samsung has warned that memory supply constraints are likely to persist through 2028, according to reporting cited by The Register.

Nvidia isn't escaping the same pressure. The company's comparable GB10-based workstations have also roughly doubled in price over the past year, now landing in the same $6,000 to $8,000 range as the Evo-X5.

The rivals: AMD isn't the only company betting on unified memory for local AI. Apple's 2026 Mac Studio takes the same basic approach with its M5 series chips, and the comparison is instructive.

The M5 Max Mac Studio starts at $2,499 with up to 128GB of unified memory and 614 GB/s of bandwidth, well ahead of Gorgon Halo's 273 GB/s. The M5 Ultra configuration goes further still, supporting up to 512GB of unified memory at 1.2TB/s of bandwidth, enough to comfortably run 400-billion-parameter class models and even some 600-billion-parameter models at reduced precision.

That top configuration starts at $5,499, though reaching the full 512GB requires pairing it with the more expensive 80-core GPU option.

The catch: Apple's memory bandwidth advantage is substantial, more than double what Gorgon Halo offers at the high end. For memory-bandwidth-bound AI inference, that gap likely translates into meaningfully faster real-world performance per token generated, even before accounting for the memory capacity difference. AMD's platform still has the advantage of running Windows and Linux natively, which matters for developers whose tooling isn't built around macOS.

What's next: The Register's own analysis lands on a fairly blunt note of skepticism, that it's "awfully hard to cross shop when you can't afford either option." At this price point, the comparison isn't really Gorgon Halo versus a cheaper alternative. It's whether a buyer can justify a five-figure hardware purchase at all for local AI work that a cloud subscription might otherwise handle for a fraction of the upfront cost.

For a narrow set of buyers, researchers, businesses with strict data privacy requirements, or hobbyists serious about running frontier-scale open models at home, that math might still work out. For most people, hardware in this bracket remains firmly outside typical PC-buying territory, and this generation's pricing hasn't made that any easier, even as the underlying capability keeps genuinely improving.

Who's affected: the real audience for a machine like this is narrower than the marketing around "run AI at home" implies. Developers building and testing applications against large open-weight models, small businesses with data residency or privacy requirements that rule out cloud APIs, and researchers who need to experiment with model internals directly are the practical buyers.

Casual users experimenting with a local chatbot have far cheaper options, including smaller models that run fine on ordinary consumer hardware.