APPLE first released a M Pro 256-bit APU (en.wikipedia.org/wiki/Apple_M5: M5 Pro 307 GB/s), when AMD followed with Strix Halo (256 GB/s) (Gorgon Halo refresh: 273 GB/s) and then NVIDIA followed with Spark N1X (301 GB/s). Currently AMD offers with Gorgon Halo up to 192 GB unified memory|RAM and NVIDIA only offers in Spark only up to 128 GB RAM, so where is NVIDIA's enthusiasm?
And how about Linux support, not only a locked down version:
QuoteDGX Spark runs on Linux: NVIDIA says it uses DGX OS, a customized Linux distribution with built-in drivers, diagnostics, and a stable base for AI and analytics workloads.
? This is not very enthusiastic either.
But nevermind, because all have one thing in common: Relatively low memory speed/bandwidth of ~300 GB/s and the on it based GPU compute on the level of a 4060 Laptop to 5070 Laptop.
So, for comparison:
AI bang for buck performance alternativeFor more AI bang for buck performance and upgradable parts, get a consumer desktop PC 96 to 128 GB RAM (a consumer desktop PC uses 128-bit memory interface (aka dual-channel) and the RAM's MT/s is typically between 5600-6200 MT/s, which equals to ~96 GB/s (vs Spark's ~300 GB/s), but the GPU(s) will make up for it (see calculation below) (32 GB VRAM at ~1800 GB/s, when running 2 5070 Ti (each at 896 GB/s) in tensor parallel) + 1 GPU (at lest like a 5070 Ti (16 GB VRAM at 896 GB/s, it will have much faster prompt processing and you will have 128+16 GB of total memory) (a 5080 24 GB VRAM is rumored to be released in 2027 Q1)) or may get a motherboard that supports bifurcation (geizhals.eu/?cat=mainboards&xf=2833_AMD+AM5~2873_ab+2&pagesize=30&promode=true) and get 2 GPUs of the same GPU architecture and run them in tensor parallel mode, which will double the VRAM memory in size and speed (reddit.com/r/LocalLLaMa/search/?q=tensor parallel).
Let's calculate our arbitrary AI score (memory size times memory speed):
N1X Spark:
- 38528 = 128 GB unified memory * 301 GB/s.
Desktop PC build using two 5070 Ti and running them in tensor parallel:
- 69632 = 128 GB * 96 GB/s + 32 GB VRAM * 1792 GB/s.
- 66560 = 96 GB * 96 GB/s + 32 GB VRAM * 1792 GB/s.
And then there's of course the price of 6700 bucks for the 128 GB unified memory config. I think the suggested Desktop PC build will be cheaper and have more memory as well (128 GB vs 128 GB+32 GB) (2000ram+2*1250gpus+400 mobo+300cpu+300psu) or less if you already have some of the required parts. So, 6700 bucks for that is not very enthusiastic either.
That said, if you're still interested, the AI models that are best suited for these ~300 GB/s APUs (M Pro, Strix/Gorgon Halo, Spark N1X) are of MoE (as opposed to dense) architecture, see e.g.:
artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cglm-5-3-flash%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cnemotron-3-5-lightning%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cmuse-glimmer%2Cnvidia-nemotron-3-super-120b-a12b%2Cqwen3-8-27b-medium%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cinkling-small%2Csolar-open2-250b%2Cqwen3-6-35b-a3b%2Cqwen3-6-27b%2Cqwen3-8-flash-next%2Cqwen3-8-27b%2Cmotif-3%2Chy3%2Cg9v3-39a5b%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cmistral-medium-3-5%2Cling-3-0-flash%2Cgpt-oss-120b%2Cdeepseek-v4-flash%2Cnemotron-3-5-bolt-74b-a9b%2Cnemotron-3-5-super-vl&intelligence=artificial-analysis-intelligence-index
Qwen3.8-Flash-Next is basically the winner overall (highest AA performance per size) and has only 6B active parameters, perfect for ~300 GB/s devices.
NVIDIA even offers NVFP4 builds of Qwen3.8-Flash-Next that are compatible with Spark, since it also supports NVFP4:
- huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4 (133 GB but the n-gram can be streamed from the SSD (-> 133GB - 30% n-gram (streamed from a SSD) = 93 GB -> fits in a 128 GB unified memory Spark config)
- huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local (this one is new and should definitely fit in a 128 GB unified memory config)