AMD’s Gorgon Halo benchmarks show the promise, and limits, of running large AI models locally

AMD has published new AI benchmarks for its Ryzen AI Max+ Pro 495, pitching Gorgon Halo as a rival to Nvidia’s coming RTX Spark devices. The headline advantage is memory: AMD’s system can be configured with up to 192GB of unified memory, but more room for a model does not mean faster responses.

AMD Watch analysis

What happened

Tom’s Hardware reports that AMD tested generative-AI workloads in ComfyUI, comparing a top-spec Ryzen AI Max+ Pro 495 system with 192GB of memory against an Intel Core Ultra X9 388H system with 64GB. AMD’s claimed lead ranged from 1.1x to 32.2x; the largest figure came from a model the publication could not find on Hugging Face, making it a particularly shaky candidate for the headline.

For larger models, AMD reported peak speeds of 20 tokens per second on GLM 5.3 Flash, a 320-billion-parameter model with 18 billion active per token, using Unsloth’s UD-IQ4XS format. It also reported up to 42 tokens per second on Qwen 3.8 Flash Next, using dynamic 4-bit quantisation and multi-token prediction. These are AMD’s results as reported by Tom’s Hardware, not independent benchmark findings.

Why it matters

The extra memory could let a desktop run larger models locally, without sending every prompt to a cloud service. That is a meaningful option for people who value control over their data or want to experiment with heavyweight models at home. But capacity and speed are different things: AMD’s reported figures are “up to” results, and performance falls as context grows.

The comparisons also have limits. The AMD and Intel systems had different memory capacities, and the article says Gorgon Halo is largely a refresh of Strix Halo. Tom’s Hardware notes that its earlier testing found the Ryzen AI Max+ 395 behind Nvidia’s DGX Spark in both time to first token and tokens per second across three models. The RTX Spark itself has not yet been benchmarked in the report, so this is not a direct contest between the new devices.

Our read

The interesting story is not that AMD has declared victory before its rival arrives. It is that local AI PCs are being sold on a trade-off: fit larger models in memory, even if they run more slowly. Treat the new numbers as a preview of what AMD says its hardware can do, not a buying verdict. At roughly $7,000 for the top configuration currently discussed in the report, that distinction has a rather practical price tag.

What to watch

• Independent tests of Gorgon Halo against the RTX Spark on the same models and settings.
• How token speed changes with longer contexts, rather than at peak conditions alone.
• Final prices and configurations as more Gorgon Halo systems become available.

Discussion spark: For a local AI workstation, would you pay more to run larger models in memory, even if they respond more slowly, or is speed the deal-breaker?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.