AMD Positions Gorgon Halo for Running AI Models Locally
AMD highlights the Ryzen AI Max+ Pro 495, which supports up to 192GB of memory and achieved ComfyUI performance 1.1–32.2 times faster than the Intel Core Ultra X9 388H. However, the 32.2x figure is a particularly notable outlier compared with the others.
There are still no direct benchmark results against the RTX Spark, and AMD’s figures use a 192GB system compared with an Intel system with 64GB of memory, so they should be interpreted cautiously. Another point is that throughput decreases as context length increases, meaning that running large models in real-world workloads may be slower than the initial figures suggest.
Gorgon Halo and the New AI PC Battle
[ProductShot: Illustration from AMD]
Gorgon Halo devices have only recently started reaching the market, and some models cost more than $7,099, making them AI PCs primarily aimed at users who want to run large models locally.
AMD is positioning this chip family directly against the RTX Spark and calls this market “agentic PCs.” The selling point is therefore not just speed, but also the ability to support larger models.
When Running Large Models Locally Means Choosing Between Capacity and Speed
Imagine that you want to run a large AI model locally: the Ryzen AI Max+ Pro 495 supports up to 192GB of memory, providing more space for larger models. However, AMD reports a maximum speed of 20 tokens/sec for GLM 5.3 Flash using a specialized quantization method.
The real question is not simply “can it run?” but how much speed you are willing to trade for capacity. As context length increases, performance naturally declines. Even if you are willing to pay for a high-end system, using large models still requires accepting some latency.
Where Gorgon Halo Fits in AMD’s AI Chip Strategy
Gorgon Halo is a chip family AMD has positioned directly against the RTX Spark. Larger Apple M-series models are secondary competitors in the agentic PC market. This segment is still expanding and is not limited to conventional laptops.
From an architectural perspective, Gorgon Halo is not a major platform change, but rather a refresh of the existing Strix Halo family. Its primary selling point is therefore support for up to 192GB of memory for running large models locally, rather than a complete overhaul of the overall user experience.
What Changed from Strix Halo to Gorgon Halo
Gorgon Halo remains in the same agentic PC category as Strix Halo, but AMD is positioning it more directly against the RTX Spark. The main development is a refresh of the existing Strix Halo, not a major platform change.
| Factor | Strix Halo | Gorgon Halo |
|---|---|---|
| Product positioning | agentic PCs | agentic PCs focused on competing with RTX Spark |
| Development approach | Previous generation | Refresh of Strix Halo |
| Maximum memory | Not specified in the source | 192GB |
| Gaming and general application performance | No new data | The source does not expect major differences |
The key change is support for greater memory capacity for running AI models locally, rather than a major leap in gaming or general application performance.
The ComfyUI figures suggest that Gorgon Halo is well suited to local generative AI workloads, such as image generation or running memory-intensive workflows. Its support for up to 192GB provides room for more parameters than systems with 128GB of memory.
With GLM 5.3 Flash, which has a total of 320 billion parameters, the Ryzen AI Max+ Pro 495 achieves throughput of up to 20 tokens per second. Qwen 3.8 Flash Next, a multimodal MoE model, reaches up to 42 tokens per second, making it suitable for workloads that process both text and multiple data formats.
However, these figures represent maximum performance, and throughput decreases as context length increases. Long conversations or large documents may therefore receive slower responses than the figures shown on the chart.
Compared with RTX Spark, Intel, and Apple, Who Has the Advantage in Which Area?
| Factor | Gorgon Halo | RTX Spark / DGX Spark | Intel Core Ultra X9 388H | Larger Apple M-series |
|---|---|---|---|---|
| Maximum memory | 192GB | 128GB | 128GB | No data |
| Reported AI benchmark results | Up to 20 tokens/s in GLM 5.3 Flash and 42 tokens/s in Qwen 3.8 Flash Next | 64 tokens/s in GPT-OSS 120B on DGX Spark | AMD reported a 1.1x–32.2x advantage in ComfyUI | No data |
| Comparison status | No direct benchmark results against RTX Spark yet | No direct benchmark results against Gorgon Halo yet | Tested on a system with 64GB of RAM | Identified as a competitor in the same group |
Each set of figures uses different models, quantization methods, and conditions, so their performance cannot be compared directly. Gorgon Halo stands out for its memory capacity for running large models, while RTX Spark has throughput figures from DGX Spark. Apple has no benchmark results in this report.
Pros
- +Up to 192GB of memory helps run large models locally
- +ComfyUI benchmarks show a 1.1–32.2x advantage over Intel, while supporting large MoE models
Cons
- −There are still no direct benchmark results against RTX Spark
- −The 32.2x figure may be an outlier, and there is no gaming or general application data
- −Performance decreases as context length increases, so the “maximum” figures may not reflect real-world use
Costs and Conditions That Are Not Shown on the Price Tag
Some Gorgon Halo devices cost more than $7,099, and AMD’s reported figures also come from specific conditions. For example, GLM 5.3 Flash uses mixed 4-bit quantization, while Qwen uses dynamic 4-bit quantization together with MTP.
High memory capacity makes it possible to run large models locally, but this may come at the cost of lower performance. As context length increases, throughput also decreases, meaning that the “maximum” figures may not match real-world usage.
Is AMD Selling “Capacity” More Than Speed?
Gorgon Halo appears to be selling space for larger models as its main advantage, since it supports up to 192GB of memory compared with 128GB for RTX Spark. The selling point is therefore not just tokens per second, but how long it can run large models locally without suffering excessive performance degradation.
As context length increases, throughput naturally decreases. Therefore, the “maximum” figures do not tell the whole story. A final judgment will have to wait for RTX Spark benchmark results and general-use data for Gorgon Halo to determine which system maintains better speed and stability in real-world workloads.