AMD benchmarks Gorgon Halo against Intel in the week Nvidia's RTX Spark is expected

The ratios come from a rig with unequal memory. AMD ran its top-spec 192GB Ryzen AI Max+ Pro 495 against a Core Ultra X9 388H system with 64GB, on an Intel platform that supports up to 128GB ^[9]. The AMD machine had three times the memory ^[22]. AMD averaged multiple ComfyUI runs across various models and compared total throughput ^[8]. On that setup, a ratio folds chip speed and model fit into one number. Tom's Hardware called the 32.2x top end a clear outlier ^[26]. It found no model named Yuve on Hugging Face and said the gap could be an optimization issue, or a model too big to run on the Panther Lake machine ^[27].

AMD would argue it compared top-of-stack parts; Tom's Hardware says the two chips are not in the same class ^[7]. Gorgon Halo mainly competes with RTX Spark, and to a lesser extent with Apple's larger M-series SoCs ^[11]. Nvidia's part has no published performance results beyond a few questionable Geekbench leaks ^[6]. Its launch is expected at Microsoft's event on Wednesday, October 7 ^[3].

The token figures need the same care. GLM 5.3 Flash has 320 billion parameters but activates 18 billion per token ^[17]. AMD's peak on it is 20 tokens per second in Unsloth's UD-IQ4_XS format ^[15]. Tom's Hardware's own DGX Spark peak, 64 tokens per second, came from GPT-OSS 120B at 4-bit ^[16]. That model activates about 5 billion parameters per token ^[17]. GLM activates 3.6 times as many ^[24], so the 20 and the 64 measure different jobs.

Qwen 3.8 Flash Next is the closer match, with 5 billion active parameters per token. AMD quotes up to 42 tokens per second on it, using Unsloth's dynamic 4-bit quantization and multi-token prediction ^[1]. The last-gen Ryzen AI Max+ 395 ran GPT-OSS 120B at 56 tokens per second in Tom's Hardware's testing ^[16], 14 more ^[21]. Model, quantization and lab all differ, so the gap supports no conclusion about the new chip. For any of these numbers to transfer to a buyer's workload, the model, the quantization and the context length would all have to match.

AMD published peaks ^[15][1]. Tom's Hardware wrote that in both cases the "up to" "carries a lot on its shoulders," and that performance drops at higher context lengths ^[18].

Gorgon Halo is largely a refresh of Strix Halo ^[10]. It sells into a category Tom's Hardware describes as far smaller than some of the RTX Spark hype suggests ^[19]. AMD told a press prebriefing it had shipped "10s of millions" of AI PCs, then clarified the figure as "over half a million" agentic PCs ^[12]. The number shrank by a factor of at least 20 between sentences ^[25].

Memory is the one axis with public numbers on both sides, and the matchup so far has turned mainly on it ^[20]. Gorgon Halo supports 192GB against RTX Spark's 128GB ceiling ^[13], 64GB more ^[23]. That capacity runs larger models locally, at lower performance ^[14]. I think the memory spec is the only part of AMD's pitch a buyer can check against Nvidia this week. A team whose models fit in 128GB is choosing on throughput, and that comparison arrives only with independent RTX Spark results. For that team I would hold an order for a machine that can cost upwards of $7,099 ^[5] until those results are out. A team that needs more than 128GB already has its answer in the spec sheet ^[13].