Carmack: Jetson Thor's 128GB at 273GB/s is over-provisioned for real-time robotics
ID_AA_Carmack · x · 2026-09-11
John Carmack argues NVIDIA's Jetson Thor is mis-provisioned for the real-time robotics market: 128GB of memory but only 273 GB/s of bandwidth means expensive, over-provisioned memory. A model evaluating at tens of fps can use at most 10GB of weights. More memory only pays off with a wide mixture-of-experts model or a large strategic planning model running in small time slices over longer periods. For cost-optimized systems, much less memory should suffice — though more memory always makes development easier.
More from Embodied
- ego2wrist: Faking robot wrist-camera views from egocentric video, then testing if they work — chris_j_paxton · 2026-09-11
- Analog Devices to acquire edge-AI chipmaker Alif Semiconductor for $1.35B — arifj · 2026-09-11
- Using Monocular SLAM + Ego Video to Build Usable RL Environments for Embodied AI — chris_j_paxton · 2026-09-11
- 95% local AI + 5% GPT-6: a full basketball tracking pipeline breakdown — TheMoonMidas · 2026-09-11
- Someone Gave Astra a Credit Card; It Ordered Parts and Built a DJ Controller Assembly Animation — TheMoonMidas · 2026-09-11
- Dev has AI agent design a tool from a photo, research materials, and 3D print it end-to-end — banteg · 2026-09-11