one-gpu-capacity / README.md
thanatchon36's picture
Publish one-gpu-capacity results
7975e66 verified
|
Raw History Blame Contribute Delete
4.67 kB
---
license: cc-by-4.0
pretty_name: One GPU Capacity
language:
- en
- th
tags:
- llm-serving
- capacity-planning
- gpu
- queueing
- simulation
- tokenization
size_categories:
- 10K<n<100K
configs:
- config_name: headline
data_files: data/headline.csv
- config_name: grid
data_files: data/grid.csv
- config_name: costs
data_files: data/costs.csv
- config_name: sensitivity
data_files: data/sensitivity.csv
- config_name: thai_tokens
data_files: data/thai_tokens.csv
- config_name: gpus
data_files: data/gpus.csv
- config_name: models
data_files: data/models.csv
- config_name: rent_prices
data_files: data/rent_prices.csv
---
# One GPU Capacity: how many users can one GPU serve?
Results and catalogues behind the preprint *How Many Users Can One GPU Serve? A Capacity Law for LLM Inference
from Consumer Cards to Datacenter Accelerators* (Thanatchon Chongmankhong, November 2026).
- 📖 Paper online: <https://thanatchon36.github.io/one-gpu-capacity/paper/>
- 💻 Code and replication: <https://github.com/thanatchon36/one-gpu-capacity>
- 🧮 Calibrated model card: [`thanatchon36/one-gpu-capacity`](https://huggingface.co/thanatchon36/one-gpu-capacity)
![users per GPU](fig01_headline.png)
Capacity is the number of **concurrently active users**, each alternating between thinking and waiting, for whom at
least 95% of requests meet a time-to-first-token (TTFT) and a time-per-output-token (TPOT) target.
> **No GPU was measured.** Every capacity here is a calibrated model or simulator prediction, validated against
> public benchmarks (90% of 48 MLPerf systems within 0.67–1.5×). Purchase prices are assumptions; rental prices are
> a snapshot taken on 2026-11-14.
## Users per GPU (Llama-3.1-8B, 4-bit weights, chat, TTFT ≤ 2 s, TPOT ≤ 100 ms)
| GPU | users |
|---|---:|
| RTX 3060 12GB | 0 |
| RTX 4060 Ti 16GB | 0 |
| RTX 5060 Ti 16GB | 0 |
| RTX 3090 24GB | 33 |
| RTX 4090 24GB | 172 |
| RTX 5090 32GB | 226 |
| RTX A6000 48GB | 152 |
| RTX 6000 Ada 48GB | 335 |
| RTX PRO 6000 Blackwell 96GB | 334 |
| T4 16GB | 25 |
| L4 24GB | 91 |
| A10/A10G 24GB | 112 |
| L40S 48GB | 335 |
| A100 SXM 80GB | 340 |
| H100 SXM 80GB | 688 |
| H200 SXM 141GB | 706 |
| B200 180GB | 1,010 |
## Tables
| config | file | contents |
|---|---|---|
| `headline` | `data/headline.csv` | Users per GPU for every GPU × model × precision, chat workload, standard SLO (simulator mean of three seeds, product-form and closed-form estimates, 90% band, binding regime). |
| `grid` | `data/grid.csv` | Closed-form and product-form capacity over GPU × model × precision × workload × language × SLO (7,650 cells). |
| `costs` | `data/costs.csv` | Cost per user per month (buy: 4-year TCO at assumed prices; rent: median on-demand price) and cost per million output tokens. |
| `sensitivity` | `data/sensitivity.csv` | Headline capacities under alternative settings (SLO, workload, calibration, sequence cap, Thai). |
| `thai_tokens` | `data/thai_tokens.csv` | Per-sentence token counts for Thai and English on 4,800 parallel pairs across five tokenizers (counts only; no sentence text). |
| `gpus` | `data/gpus.csv` | GPU catalogue with the provenance of every value. |
| `models` | `data/models.csv` | Model architectures. |
| `rent_prices` | `data/rent_prices.csv` | Median on-demand rental price per GPU-hour (SkyPilot catalogue snapshot, 2026-11-14). |
```python
from datasets import load_dataset
grid = load_dataset("thanatchon36/one-gpu-capacity", "grid", split="train")
```
`json/` holds the remaining result files as produced by the pipeline: `results.json` (the numbers ledger every
figure and in-text number of the paper is read from), calibration, validation V1–V5, hypotheses H1–H3, Sobol
indices, pooling and Thai summaries, and `MANIFEST.json` (source, commit and SHA-256 of each third-party raw file).
## What is not here
Third-party raw data (Azure LLM inference traces, Vidur profiles, LLM-Inference-Bench, MLPerf results, llama.cpp
benchmark table, tokenizer vocab files, the SCB-MT parallel corpus) are not redistributed. `scripts/fetch_data.sh`
in the code repository downloads them under their own licences, and `scripts/verify_manifest.py` checks the hashes.
## Citation
```bibtex
@misc{chongmankhong2026onegpu,
title = {How Many Users Can One GPU Serve? A Capacity Law for LLM Inference from Consumer Cards to Datacenter Accelerators},
author = {Chongmankhong, Thanatchon},
year = {2026},
note = {Preprint},
doi = {10.5281/zenodo.24243858},
url = {https://github.com/thanatchon36/one-gpu-capacity}
}
```
License: CC BY 4.0 (derived results and catalogues). Code in the GitHub repository is MIT.