Trending storyDeveloping
Cedana estimates GPU needs for large-scale inference
1 report1 sourceupdated yesterday
What happened
AI digest
On October 7, 2026, Hacker News featured a Cedana writeup on converting million, billion and trillion token workloads into GPU requirements. Its worked example: serving 1 trillion tokens in 30 days on Llama 3.3 70B. Cedana's calculator estimates about 367 H100 GPUs, with a range of 211 to 853. The post frames inference demand by total tokens processed, deadline, model and GPU type. All figures are calculator estimates, not results from a real deployment.
Written by AI from the coverage · updated 3 hours ago
Coverage
Follow the reports to see the story from different sides.
Oct 7
- Hacker News front pageHow many GPUs is 1M/B/T tokens?
Cedana 的计算器估算，Llama 3.3 70B 在一个月（30 天）内处理 1 trillion tokens 约需 367 张 H100，估算范围为 211–853 张。
Heat over time
Not enough continuous observations to draw a trend yet.