"Superchip"
The NV Spark is a 256-bit APU, running at 273 GB/s. For comparison:
- AMD Strix Halo 256-bit APU, 256 GB/s (came out first, so NV is copying it).
- AMD Radeon 660M-890M 128-bit APU, at up to ~100 GB/s.
- Apple M1-M3 Ultra 1024-bit APU, 819 GB/s. (en.wikipedia.org/wiki/Apple_M3).
- RTX 5070 Ti desktop GPU, 896 GB/s.
Let's continue, why it's not a superchip: For the same file-size and hence memory requirement, a dense AI LLM model outperforms a MoE AI LLM model (see 122B MoE vs 27B dense: huggingface.co/Qwen/Qwen3.5-122B-A10B#benchmark-results), while being 4-5 times smaller.
There needs to be more ~120B MoE models for these up to 128 GB RAM APUs to be usable, as Qwen3.6-27B dense model outperforms them. Basically, only more than 120B MoE models outperform dense Qwen3.6-27B and they require more than 128 GB RAM: Look where 27B dense model sits vs the larger MoE models:
artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cnvidia-nemotron-3-super-120b-a12b%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cqwen3-6-35b-a3b%2Cdeepseek-v4-flash-high%2Cqwen3-6-27b%2Cdeepseek-v4-flash-non-reasoning%2Chy3%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cdeepseek-v4-flash%2Cmistral-medium-3-5%2Cgpt-oss-120b%2Cmotif-0714&intelligence=artificial-analysis-intelligence-index
(Even a 120B model, at a 8-bit, requires 120 GB, so to fit these up to ~300B models, a 4- to 5-bit quant is used.)
What is really needed is a 192 or 256 GB RAM APU. AMD knows this and this is why they announced Gorgon Halo (a Strix Halo refresh, but with up to 192 GB RAM (160 GB allocatable on Windows, and maybe more on Linux, just like with Strix Halo)):
notebookcheck.net/AMD-Ryzen-AI-Max-400-lineup-now-official-with-up-to-192-GB-RAM.1301777.0.html
Deutsch
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar