19:55Machine-translated from Azerbaijani by AIRead in Azerbaijani
AMD Reveals AI Benchmarks for Gorgon Halo Chip With 192 GB of Memory
Nvidia is expected to unveil its RTX Spark devices this week. The announcement is expected to be made at Microsoft’s event on October 7. Ahead of the event, AMD has released the first benchmark results for its rival Gorgon Halo chips. The first devices are already on sale, with some versions costing more than 7099 dollars (approximately 12070 manat).
AMD tested generative AI performance in ComfyUI and calculated the average result across several models. The Ryzen AI Max+ Pro 495 with 192 GB of memory was compared with the Intel Core Ultra X9 388H with 64 GB of memory. However, these processors do not belong to the same device class. The Panther Lake platform supports up to 128 GB of memory. Depending on the test, the AMD chip’s advantage ranged from 1.1 to 32.2 times. The largest difference may have resulted from a compatibility issue or from the model being too large for the rival system. No new results have been released for games or everyday applications. Gorgon Halo is primarily an updated version of the previous Strix Halo family. The main change is an increase in shared memory from 128 GB to 192 GB. This memory, which is used jointly by the processor and graphics component, helps run larger AI models. The boost clock of the Ryzen AI Max+ Pro 495 has also been increased by 100 MHz. The core count and internal architecture remain the same as in the previous generation. Its main rivals are RTX Spark and, to some extent, Apple’s M series. In AMD’s test, GLM 5.3 Flash processed 20 tokens per second. A token is a small piece of text processed by an AI model. The higher the number, the faster the response is generated. The model has a total of 320 billion parameters, but 18 billion of them are used for each token. The Unsloth UD-IQ4_XS compression format was used to reduce the load on memory.
In the Qwen 3.8 Flash Next test, the Ryzen AI Max+ Pro 495 delivered up to 42 tokens per second. The model has 125 billion base parameters. Of these, 51 billion are allocated to the digital representation of text, while approximately 4 billion are allocated to calculating several tokens in advance. When each token is processed, 5 billion parameters are activated. AMD released the highest result from each test. However, performance may decline as the text and task become longer.
For comparison, DGX Spark processed 64 tokens per second with the GPT-OSS 120B model. On the previous Ryzen AI Max+ 395 processor, this figure was 56 tokens. Because the results were obtained with different models, they cannot be regarded as a direct speed comparison. In previous tests, the AMD chip lagged behind DGX Spark across 3 models in both time to first token and tokens processed per second. Gorgon Halo supports 192 GB of unified memory and can run larger models directly on the device. However, having more memory does not mean that those models will automatically run faster. According to AMD, more than 500 thousand devices have been shipped in the category of agent-based computers. The company says the total number of computers with AI features is in the tens of millions. One of the first Gorgon Halo devices, the Minisforum MS-S1 Max-P495, is already on sale. The most powerful configuration costs approximately 7000 dollars (11900 manat). The price range is expected to broaden as models from different manufacturers are released.
AMD tested generative AI performance in ComfyUI and calculated the average result across several models. The Ryzen AI Max+ Pro 495 with 192 GB of memory was compared with the Intel Core Ultra X9 388H with 64 GB of memory. However, these processors do not belong to the same device class. The Panther Lake platform supports up to 128 GB of memory. Depending on the test, the AMD chip’s advantage ranged from 1.1 to 32.2 times. The largest difference may have resulted from a compatibility issue or from the model being too large for the rival system. No new results have been released for games or everyday applications. Gorgon Halo is primarily an updated version of the previous Strix Halo family. The main change is an increase in shared memory from 128 GB to 192 GB. This memory, which is used jointly by the processor and graphics component, helps run larger AI models. The boost clock of the Ryzen AI Max+ Pro 495 has also been increased by 100 MHz. The core count and internal architecture remain the same as in the previous generation. Its main rivals are RTX Spark and, to some extent, Apple’s M series. In AMD’s test, GLM 5.3 Flash processed 20 tokens per second. A token is a small piece of text processed by an AI model. The higher the number, the faster the response is generated. The model has a total of 320 billion parameters, but 18 billion of them are used for each token. The Unsloth UD-IQ4_XS compression format was used to reduce the load on memory. In the Qwen 3.8 Flash Next test, the Ryzen AI Max+ Pro 495 delivered up to 42 tokens per second. The model has 125 billion base parameters. Of these, 51 billion are allocated to the digital representation of text, while approximately 4 billion are allocated to calculating several tokens in advance. When each token is processed, 5 billion parameters are activated. AMD released the highest result from each test. However, performance may decline as the text and task become longer. For comparison, DGX Spark processed 64 tokens per second with the GPT-OSS 120B model. On the previous Ryzen AI Max+ 395 processor, this figure was 56 tokens. Because the results were obtained with different models, they cannot be regarded as a direct speed comparison. In previous tests, the AMD chip lagged behind DGX Spark across 3 models in both time to first token and tokens processed per second. Gorgon Halo supports 192 GB of unified memory and can run larger models directly on the device. However, having more memory does not mean that those models will automatically run faster. According to AMD, more than 500 thousand devices have been shipped in the category of agent-based computers. The company says the total number of computers with AI features is in the tens of millions. One of the first Gorgon Halo devices, the Minisforum MS-S1 Max-P495, is already on sale. The most powerful configuration costs approximately 7000 dollars (11900 manat). The price range is expected to broaden as models from different manufacturers are released.