Glonce

AMD Gorgon Halo debuts with 192 GB memory, from $6,799 at GMKtec

productsThe Glonce editorsThe Register🔥69Glonce rating69Impact65Novelty70Relevance85Surprise45Players70Credibility82
AMD Gorgon Halo debuts with 192 GB memory, from $6,799 at GMKtec

The first systems built on AMD's Gorgon Halo system-on-chip platform have arrived, with up to 192 GB of unified memory. The Register says that is enough to put DeepSeek V4 Flash on your desk, if you can pay a hefty premium for the RAM. GMKtec's EVO-X5 Pro is among the first to use the top-specced Ryzen AI Max+ 495 SoC, which the article shortens to Gorgon Halo. Its regular pricing starts at $6,799.

The pitch is local AI. At 4-bit precision, Gorgon Halo has enough memory to run models of around 345 billion parameters on Linux or 320 billion on Windows. The gap comes from memory partitioning: on Windows you have to allocate 160 GB of memory to the GPU, while the Linux kernel and AMD's GPU drivers let users take advantage of nearly all the available memory. That puts the 284 billion parameter DeepSeek V4 Flash and Z.AI's 320 billion parameter GLM-5.3-Flash within reach. The Register suggests local AI enthusiasts and perhaps privacy-conscious small businesses might find the price worth it, to run larger, more capable models from the security of their homes and offices.

The chip itself is a modest step. Announced earlier this year, Gorgon Halo is, in the article's words, essentially a factory overclocked version of the Strix Halo APU that powers AMD's AI Halo workstation, which The Register reviewed this summer. Both can have up to 16 Zen 5 cores, a 40-compute-unit integrated GPU and an XDNA 2-based NPU. AMD has raised clocks by about 100 MHz across the board, and Gorgon Halo supports 192 GB of faster 8,533 MT/s LPDDR5x, against Strix Halo's 128 GB at 8,000 MT/s.

The authors do not expect the clock bump to make a meaningful difference to performance. Faster memory could help a little for local AI, since LLM inference is predominantly memory bound. But Gorgon Halo's 273 GB/s of memory bandwidth is only about 6.5 percent faster than Strix Halo's 256 GB/s. At most, the article expects a few more tokens a second on highly optimized models. The real attraction is the higher memory capacity.

That capacity is costly because of the ongoing memory shortage. A year ago, a 128 GB Strix Halo box cost between $2,000 and $3,000 depending on the OEM. DDR5 prices have since skyrocketed, and by the time AMD launched its AI Halo workstation this summer, prices had jumped to $4,000. With 50 percent more memory, Gorgon Halo products should be pricier still. The GMKtec unit is already 64 percent more expensive, and the article notes the brand is usually on the more affordable side. More premium brands like Framework, and notebooks like HP's ZBook Ultra G3a, will likely command a steep premium.

The Register argues there is little AMD can do. LPDDR5x is in short supply thanks to the AI boom: each Nvidia NVL72 rack is packed with between 36 and 54 TB of it, and that is just CPU memory. The situation is only going to get worse as AMD embraces LPDDR5x for its own AI-optimized Verano Epycs. Samsung warned earlier this year that memory supplies are likely to remain constrained through 2028.

Nvidia faces the same squeeze. Prices of AI workstations based on its GB10 SoC have roughly doubled over the past year, from $3,000-$4,000 to $6,000-$8,000 today. Nvidia's RTX Spark, announced at Computex late this spring, is a family of premium Windows systems from major OEMs built around the same GB10 and offering up to 128 GB of memory, as mini PCs and notebooks. The authors would not be surprised to see them command a premium over existing GB10 boxes. That leaves AMD in a position to undercut Nvidia by offering more memory for less, but it is awfully hard to cross shop when you can't afford either option.

Key facts

  • GMKtec's EVO-X5 Pro, among the first Gorgon Halo (Ryzen AI Max+ 495) systems, has regular pricing starting at $6,799.
  • Gorgon Halo supports up to 192 GB of 8,533 MT/s LPDDR5x, versus Strix Halo's 128 GB at 8,000 MT/s; at 4-bit precision that fits models of around 345 billion parameters on Linux or 320 billion on Windows.
  • Bandwidth is 273 GB/s, only about 6.5 percent above Strix Halo's 256 GB/s; the article expects at most a few more tokens a second on highly optimized models.
  • A 128 GB Strix Halo box cost $2,000 to $3,000 a year ago and $4,000 by summer; Samsung warned memory supply may stay constrained through 2028.
  • Nvidia's GB10-based AI workstations went from $3,000-$4,000 to $6,000-$8,000 in a year, and RTX Spark systems may cost more still.

Why it matters

Gorgon Halo raises the ceiling for running big models on a single desktop box: up to 192 GB of unified memory brings 284 billion parameter DeepSeek V4 Flash and 320 billion parameter GLM-5.3-Flash within reach at 4-bit precision. The Register's point is that capacity, not speed, is the draw. The chip is essentially a factory overclocked Strix Halo, and the memory bandwidth gain is only about 6.5 percent. The story is also a snapshot of how the AI boom's appetite for LPDDR5x is pushing up the price of consumer hardware.

Who it affects

The Register names local AI enthusiasts and perhaps privacy-conscious small businesses as the people who might find the price worth paying, since they can run larger models at home or in the office. Buyers who need a lot of memory right now are, in the article's words, in for a lot of pain. AMD and Nvidia both feel the memory squeeze, as Nvidia's GB10-based workstations have also roughly doubled in price over the past year.

How to use it

Memory allocation depends on the operating system. On Windows you need to allocate 160 GB of memory to the GPU, which caps models at around 320 billion parameters at 4-bit precision. On Linux, the kernel and AMD's GPU drivers let users take advantage of nearly all the available memory, for models of around 345 billion parameters. The GMKtec EVO-X5 Pro is available with regular pricing starting at $6,799. The article gives no availability dates or pricing for Framework or HP Gorgon Halo products.

How solid is it

This is a news and analysis piece from The Register, not satire. Hardware specs, memory figures and prices are stated plainly, and the price history is given as ranges. The article gives no benchmark results or measured tokens-per-second figures for Gorgon Halo; the performance view is the authors' expectation. The claim that the GMKtec unit is 64 percent more expensive does not state its baseline.

Risks and caveats

Prices are the main risk: Gorgon Halo products from pricier brands like Framework and HP will likely command a steep premium, and the article says the memory situation is not expected to improve anytime soon. Samsung warned earlier this year that supplies are likely to stay constrained through 2028, and AMD's own move to LPDDR5x for its Verano Epycs could make it worse. The performance gain over Strix Halo is small, so the case for buying rests on the extra memory alone. Nvidia's RTX Spark systems, at up to 128 GB, may also cost more than existing GB10 boxes, so the price gap the article sees in AMD's favor may not make either option affordable.

“Gorgon's main attraction is really the prospect of higher memory capacity.”

— The Register