FUNDA(@FundaAI):Clarifying the rumor that Gemini 4 was trained on NVIDIA GB: it wasn't. Still TPU. Google's stack is JAX/XLA/Pathways compiled to TPU topology — a frontier run doesn't move to CUDA mid-flight, and a restart isn't evidence the silicon failed. 2/ The better story isn't who's winning. NVIDIA is pushing hard, Google is pushing hard — and both roadmaps land on the same place: optics. 3/ NVIDIA: Rubin Ultra NVL576 brings optics into the scale-up NVLink domain for the first time. 576 GPUs coherent simply cannot be held in copper. https://t.co/X2C506Ksi3 4/ Google: TPU has run OCS since v4, and 9t is rumored to lift the torus from 3D to 6D, optical content 4x. https://t.co/sPQJaKcN0Y 5/ And it's not just more links. Coherent Lite pushes per-lane 800G → 2.4T, while scale-in moves optics inside the tray — a bandwidth opportunity that could be 10x scale-up. https://t.co/rwc043pfEg https://t.co/WryGn1VEU8 6/ Two very different architectures, same conclusion: past a certain scale, the interconnect has to go optical. That's the trade.

2026.09.21 16:17

Clarifying the rumor that Gemini 4 was trained on NVIDIA GB: it wasn't. Still TPU. Google's stack is JAX/XLA/Pathways compiled to TPU topology — a frontier run doesn't move to CUDA mid-flight, and a restart isn't evidence the silicon failed. 2/ The better story isn't who's winning. NVIDIA is pushing hard, Google is pushing hard — and both roadmaps land on the same place: optics. 3/ NVIDIA: Rubin Ultra NVL576 brings optics into the scale-up NVLink domain for the first time. 576 GPUs coherent simply cannot be held in copper. 4/ Google: TPU has run OCS since v4, and 9t is rumored to lift the torus from 3D to 6D, optical content 4x. 5/ And it's not just more links. Coherent Lite pushes per-lane 800G → 2.4T, while scale-in moves optics inside the tray — a bandwidth opportunity that could be 10x scale-up. 6/ Two very different architectures, same conclusion: past a certain scale, the interconnect has to go optical. That's the trade.

FUNDA@FundaAI

2026.09.18 15:02

Just a month ago, the rumors were all about how Gemini was doomed. We spent a lot of time explaining why this wasn’t actually a bad thing for Gemini. To understand why, you need to understand how labs train models these days. Frontier labs are becoming more efficient and more focused. Of course, some people want to pursue a different path through Neo Labs, or focus more on research that could win a Nobel Prize. But after they leave, the labs actually become more efficient and more focused. We’ve already seen positive results at both Gemini and Qwen as they’ve narrowed their focus and improved efficiency.

커뮤니티로 전달