netrunner
@plotarmordev • 8,161 subscribers
Cofounder & engineer. Running local AI, building things with it, and sharing what I learn.
Shorts
Videos
🆕 The newly announced Xiaomi’s AI Cube and triple custom chips (O3 + O100 + D100) are interesting for what they say about China’s hardware stack, here are the points that stand out to me: - O100 is a 6nm, 3D wafer-on-wafer near memory design claiming 1.22 TB/s ~ about 4.5× GB10. If that bandwidth scales beyond its internal DRAM. - That to me, also raises an interesting question for the next Mac Studio: what happens to Apple’s memory bandwidth and pricing now that it can diversify more of its memory supply from China? - Still, the advertised 330 tok/s was on MiMo-3B. Xiaomi hasn’t shown 1.22 TB/s across the full 160GB pool, or how the three chip system handles 120B weights. - China designing 6nm chips and advanced packaging is BIG. Not full-stack parity yet obviously fab yields, tooling and software still matter, but the design gap is closing faster than many expected.
netrunner27,392 views • 2 months ago
No more content to load