The Compute Threshold for 2.8T Models
basedjensen · x · 2026-07-17
This post emphasizes that the usability of massive models is primarily limited by compute and VRAM.
- Taking a 2.8T parameter model as an example, loading it directly into VRAM in FP16 would theoretically require around 5.6TB VRAM, roughly equivalent to 44 Mac Studios or 15 Blackwell GPUs.
- The original text adds that such models are actually trained and served in FP4; even when uploaded to Hugging Face, the size remains 1.4TB.
- Conclusion: Even if the model's capabilities are strong, only a few organizations can host it locally; compute remains a "soft threshold" between individuals and institutions.
More from Infra
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11