Nvidia claims 30x cost-efficiency gain with Vera Rubin as HBM grows just 3%
创业邦 · wechat · 2026-11-07
Industry news roundup:
- Nvidia's full-stack AI memory design: amid tight HBM supply, Nvidia proposes co-designing GPU, CPU, network, storage, and software—using NVFP4 to compress weights and KV Cache, tiering inactive context to cheaper storage, and distributing across GPUs. In SemiAnalysis AgentX tests, GB300 NVL72 hit up to 10x H200 per-cost throughput; Vera Rubin NVL72 improves that up to 30x while per-GPU HBM grows only 3%.
- Google commits $17B to its first Wyoming data center, with a 702MW fuel-cell project plus carbon capture.
- Musk says Tesla will phase out Cybertruck production at Giga Texas by mid-next year to free capacity for Cybercab.
- UNI-DOGGIE robot dog from Juxing Legend x Unitree: 43,900 yuan, limited to 2,000 units, 1,000 pre-sold to ADATA.
- DJI launches Matrice 5T, the first light industrial drone with standard LiDAR, mmWave radar, and omnidirectional vision; from 57,488 yuan.
More from Infra
- Batch APIs: the 50% LLM discount only pays off after model choice and prompt caching — arpit_bhayani · 2026-11-13
- RTX Pro 6000 bricks mid-Qwen run, owner jokes it "died of a broken heart" over Jensen news — wine_dark · 2026-11-13
- Bittensor's Chutes Handles 256 Concurrent Requests on a Single RTX 5090 at ~18k tok/s — markjeffrey · 2026-11-13
- Google expands TPU v8 lineup: 8tx cuts latency 2x, 8i Max doubles silicon per rack — Beth_Kindig · 2026-11-13
- Horace He to share two mental models for reasoning about inference at SGLang Summit — cHHillee · 2026-11-13
- Used A100 prices fall another 20-30% as sellers rush ahead of a collapse — Bascilian · 2026-11-13