Datasets:
gpu stringlengths 2 10 | label stringlengths 7 27 | tier stringclasses 5
values | model stringclasses 2
values | wbits float64 4.5 16 | N float64 0 1.01k | price float64 300 40k | buy_month float64 50.9 1.27k | buy_month_low float64 49.7 1.04k | buy_month_high float64 52.1 1.51k | rent_hr float64 0.1 7.34 ⌀ | rent_n int64 0 8 | rent_month float64 69.4 5.36k ⌀ | usd_user_buy float64 0.65 10.9 ⌀ | usd_user_rent float64 1.29 30 ⌀ | users_per_kusd_buy float64 0 42.9 | usd_per_mtok_buy float64 0.07 1.1 ⌀ | usd_per_mtok_rent float64 0.14 3.04 ⌀ | tok_s float64 0 3.89k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
rtx3060 | RTX 3060 12GB | budget | llama31-8b | 4.5 | 0 | 300 | 50.882283 | 49.707234 | 52.057333 | 0.095 | 1 | 69.35 | null | null | 0 | null | null | 0 |
rtx4060ti | RTX 4060 Ti 16GB | budget | llama31-8b | 4.5 | 0 | 450 | 54.075281 | 52.900232 | 55.25033 | null | 0 | null | null | null | 0 | null | null | 0 |
rtx5060ti | RTX 5060 Ti 16GB | budget | llama31-8b | 4.5 | 0 | 450 | 55.071731 | 54.57821 | 57.42183 | null | 0 | null | null | null | 0 | null | null | 0 |
rtx3090 | RTX 3090 24GB | enthusiast | llama31-8b | 4.5 | 33 | 800 | 84.487472 | 80.962324 | 89.187669 | 0.295625 | 2 | 215.80625 | 2.560226 | 6.539583 | 10.849735 | 0.272281 | 0.695488 | 118.072553 |
rtx3090 | RTX 3090 24GB | enthusiast | llama31-8b | 16 | 35.333333 | 800 | 84.487472 | 80.962324 | 89.187669 | 0.295625 | 2 | 215.80625 | 2.391155 | 6.107724 | 11.616888 | 0.242548 | 0.61954 | 132.546805 |
rtx4090 | RTX 4090 24GB | enthusiast | llama31-8b | 4.5 | 172.333333 | 1,800 | 114.631458 | 109.90776 | 131.082148 | 0.518125 | 2 | 378.23125 | 0.665173 | 2.194765 | 41.760239 | 0.076505 | 0.25243 | 570.151957 |
rtx4090 | RTX 4090 24GB | enthusiast | llama31-8b | 16 | 82.666667 | 1,800 | 114.631458 | 109.90776 | 131.082148 | 0.518125 | 2 | 378.23125 | 1.386671 | 4.575378 | 20.031991 | 0.131856 | 0.435065 | 330.809764 |
rtx5090 | RTX 5090 32GB | enthusiast | llama31-8b | 4.5 | 226 | 2,500 | 146.436194 | 134.6622 | 169.93718 | 0.4 | 1 | 292 | 0.647948 | 1.292035 | 42.870397 | 0.074742 | 0.149039 | 745.51576 |
rtx5090 | RTX 5090 32GB | enthusiast | llama31-8b | 16 | 206.333333 | 2,500 | 146.436194 | 134.6622 | 169.93718 | 0.4 | 1 | 292 | 0.709707 | 1.415186 | 39.139787 | 0.072188 | 0.143946 | 771.895582 |
a6000 | RTX A6000 48GB | workstation | llama31-8b | 4.5 | 152.333333 | 4,500 | 203.415902 | 179.914916 | 215.166395 | 0.49 | 3 | 357.7 | 1.335334 | 2.34814 | 20.802117 | 0.153665 | 0.270215 | 503.713628 |
a6000 | RTX A6000 48GB | workstation | llama31-8b | 16 | 131.666667 | 4,500 | 203.415902 | 179.914916 | 215.166395 | 0.49 | 3 | 357.7 | 1.544931 | 2.716709 | 17.979948 | 0.178509 | 0.313902 | 433.609912 |
a6000 | RTX A6000 48GB | workstation | llama31-70b | 4.5 | 0 | 4,500 | 203.415902 | 179.914916 | 215.166395 | 0.49 | 3 | 357.7 | null | null | 0 | null | null | 0 |
rtx6000ada | RTX 6000 Ada 48GB | workstation | llama31-8b | 4.5 | 334.666667 | 7,000 | 262.168367 | 238.667381 | 273.918861 | 0.71 | 2 | 518.3 | 0.783372 | 1.548705 | 35.45926 | 0.091233 | 0.180366 | 1,093.456872 |
rtx6000ada | RTX 6000 Ada 48GB | workstation | llama31-8b | 16 | 323.333333 | 7,000 | 262.168367 | 238.667381 | 273.918861 | 0.71 | 2 | 518.3 | 0.81083 | 1.60299 | 34.258448 | 0.09483 | 0.187477 | 1,051.978983 |
rtx6000ada | RTX 6000 Ada 48GB | workstation | llama31-70b | 4.5 | 0 | 7,000 | 262.168367 | 238.667381 | 273.918861 | 0.71 | 2 | 518.3 | null | null | 0 | null | null | 0 |
rtxpro6000 | RTX PRO 6000 Blackwell 96GB | workstation | llama31-8b | 4.5 | 333.666667 | 8,565 | 318.876411 | 293.847861 | 340.849833 | 2.24656 | 2 | 1,639.9888 | 0.955674 | 4.915051 | 29.066178 | 0.111132 | 0.571556 | 1,091.833839 |
rtxpro6000 | RTX PRO 6000 Blackwell 96GB | workstation | llama31-8b | 16 | 334 | 8,565 | 318.876411 | 293.847861 | 340.849833 | 2.24656 | 2 | 1,639.9888 | 0.95472 | 4.910146 | 29.095215 | 0.111032 | 0.571042 | 1,092.817523 |
rtxpro6000 | RTX PRO 6000 Blackwell 96GB | workstation | llama31-70b | 4.5 | 0 | 8,565 | 318.876411 | 293.847861 | 340.849833 | 2.24656 | 2 | 1,639.9888 | null | null | 0 | null | null | 0 |
t4 | T4 16GB | dc-inference | llama31-8b | 4.5 | 25 | 1,500 | 168.980563 | 157.23007 | 192.481549 | 0.526 | 3 | 383.98 | 6.759223 | 15.3592 | 4.109611 | 0.764299 | 1.73674 | 84.129506 |
l4 | L4 24GB | dc-inference | llama31-8b | 4.5 | 91.333333 | 2,500 | 192.667553 | 180.91706 | 216.168539 | 0.558205 | 3 | 407.48965 | 2.109499 | 4.461566 | 13.167952 | 0.243154 | 0.514268 | 301.509722 |
l4 | L4 24GB | dc-inference | llama31-8b | 16 | 37.666667 | 2,500 | 192.667553 | 180.91706 | 216.168539 | 0.558205 | 3 | 407.48965 | 5.115068 | 10.818309 | 5.430579 | 0.594752 | 1.257894 | 123.267075 |
a10 | A10/A10G 24GB | dc-inference | llama31-8b | 4.5 | 112.333333 | 3,000 | 211.672202 | 188.171216 | 235.173188 | 1.29 | 3 | 941.7 | 1.884322 | 8.383086 | 14.741522 | 0.216732 | 0.96421 | 371.634226 |
a10 | A10/A10G 24GB | dc-inference | llama31-8b | 16 | 61.666667 | 3,000 | 211.672202 | 188.171216 | 235.173188 | 1.29 | 3 | 941.7 | 3.432522 | 15.270811 | 8.092527 | 0.350612 | 1.559822 | 229.727095 |
l40s | L40S 48GB | dc-inference | llama31-8b | 4.5 | 334.666667 | 8,000 | 347.777533 | 324.276547 | 394.779505 | 0.8994 | 5 | 656.562 | 1.039176 | 1.961839 | 26.730583 | 0.121251 | 0.228907 | 1,091.41988 |
l40s | L40S 48GB | dc-inference | llama31-8b | 16 | 312 | 8,000 | 347.777533 | 324.276547 | 394.779505 | 0.8994 | 5 | 656.562 | 1.114672 | 2.104365 | 24.920145 | 0.130544 | 0.246451 | 1,013.726033 |
l40s | L40S 48GB | dc-inference | llama31-70b | 4.5 | 0 | 8,000 | 347.777533 | 324.276547 | 394.779505 | 0.8994 | 5 | 656.562 | null | null | 0 | null | null | 0 |
a100 | A100 SXM 80GB | flagship | llama31-8b | 4.5 | 340 | 12,000 | 560.364521 | 489.861563 | 677.869451 | 2.79 | 5 | 2,036.7 | 1.648131 | 5.990294 | 16.854109 | 0.19135 | 0.695481 | 1,114.336549 |
a100 | A100 SXM 80GB | flagship | llama31-8b | 16 | 334.666667 | 12,000 | 560.364521 | 489.861563 | 677.869451 | 2.79 | 5 | 2,036.7 | 1.674396 | 6.085757 | 16.58973 | 0.194768 | 0.707903 | 1,094.783207 |
a100 | A100 SXM 80GB | flagship | llama31-70b | 4.5 | 0 | 12,000 | 560.364521 | 489.861563 | 677.869451 | 2.79 | 5 | 2,036.7 | null | null | 0 | null | null | 0 |
h100 | H100 SXM 80GB | flagship | llama31-8b | 4.5 | 688.333333 | 25,000 | 893.77794 | 776.273009 | 1,128.7878 | 3.12 | 8 | 2,277.6 | 1.298467 | 3.308862 | 21.392753 | 0.143974 | 0.366887 | 2,362.218671 |
h100 | H100 SXM 80GB | flagship | llama31-8b | 16 | 691 | 25,000 | 893.77794 | 776.273009 | 1,128.7878 | 3.12 | 8 | 2,277.6 | 1.293456 | 3.296093 | 21.47563 | 0.143366 | 0.365337 | 2,372.237886 |
h100 | H100 SXM 80GB | flagship | llama31-70b | 4.5 | 95.333333 | 25,000 | 893.77794 | 776.273009 | 1,128.7878 | 3.12 | 8 | 2,277.6 | 9.375293 | 23.890909 | 2.96287 | 1.072283 | 2.732481 | 317.172092 |
h200 | H200 SXM 141GB | flagship | llama31-8b | 4.5 | 706.333333 | 32,000 | 1,058.284842 | 964.280898 | 1,246.29273 | 3.99 | 5 | 2,912.7 | 1.49828 | 4.12369 | 18.539782 | 0.163327 | 0.449523 | 2,465.573983 |
h200 | H200 SXM 141GB | flagship | llama31-8b | 16 | 706 | 32,000 | 1,058.284842 | 964.280898 | 1,246.29273 | 3.99 | 5 | 2,912.7 | 1.498987 | 4.125637 | 18.531033 | 0.163388 | 0.449689 | 2,464.666072 |
h200 | H200 SXM 141GB | flagship | llama31-70b | 4.5 | 97 | 32,000 | 1,058.284842 | 964.280898 | 1,246.29273 | 3.99 | 5 | 2,912.7 | 10.910153 | 30.027835 | 2.546048 | 1.104946 | 3.041125 | 364.448453 |
b200 | B200 180GB | flagship | llama31-8b | 4.5 | 1,010.333333 | 40,000 | 1,274.19333 | 1,039.18347 | 1,509.203191 | 7.34 | 6 | 5,358.2 | 1.261161 | 5.303398 | 22.025555 | 0.124745 | 0.524574 | 3,886.750276 |
b200 | B200 180GB | flagship | llama31-8b | 16 | 1,010.333333 | 40,000 | 1,274.19333 | 1,039.18347 | 1,509.203191 | 7.34 | 6 | 5,358.2 | 1.261161 | 5.303398 | 22.025555 | 0.124727 | 0.524497 | 3,887.323622 |
b200 | B200 180GB | flagship | llama31-70b | 4.5 | 273 | 40,000 | 1,274.19333 | 1,039.18347 | 1,509.203191 | 7.34 | 6 | 5,358.2 | 4.667375 | 19.627106 | 5.951478 | 0.537047 | 2.258373 | 902.813129 |
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
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). |
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
@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.
- Downloads last month
- 69
