DGX Spark / GB10 vs Ryzen AI Max+ 395

NVIDIA GB10 Grace Blackwell (DGX Spark and partners)

Highly recommended

Editorial opinionOur verdict

Same memory (128 GB) and nearly the same bandwidth (273 vs 256 GB/s), so token generation is similar. The GB10 wins on software (CUDA, vLLM, fine-tuning) and prompt processing; Strix Halo wins on price and is a normal x86 PC. For chat and agents at home, Strix Halo is better value; for CUDA development, GB10.

Factual data	DGX Spark / GB10	Ryzen AI Max+ 395
Memory (largest)	128 GB	128 GB
Usable for models	115 GB	115 GB
Bandwidth	273 GB/sMore	256 GB/s
Largest popular model at 4-bit	Qwen3.5 122B A10B	Qwen3.5 122B A10B
Price from	$5,000	$2,200Advantage
Price per GB	$39	$19Advantage
Power	240 W	120 W
Compute stack	CUDA (Arm)	ROCm / Vulkan

Which models each one runs

Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.

Model	DGX Spark / GB10	Ryzen AI Max+ 395
Qwen3 1.7B	Runs≤ 246 tok/s	Runs≤ 231 tok/s
Qwen3.5 9B	Runs≤ 48 tok/s	Runs≤ 45 tok/s
Gemma 4 12B	Runs≤ 38 tok/s	Runs≤ 36 tok/s
gpt-oss-20b	Runs≤ 136 tok/s	Runs≤ 128 tok/s
Qwen3.6 27B	Runs≤ 16 tok/s	Runs≤ 15 tok/s
Qwen3.6 35B A3B	Runs≤ 148 tok/s	Runs≤ 139 tok/s
Llama 3.3 70B	Runs≤ 6 tok/s	Runs, tight≤ 6 tok/s
gpt-oss-120b	Runs≤ 100 tok/s	Runs≤ 93 tok/s
Qwen3.5 122B A10B	Runs≤ 45 tok/s	Runs≤ 42 tok/s
Qwen3.5 397B A17B	Not enough memory	Not enough memory

Speed ceilings are theoretical; see each page for measured results.