跳到正文
Hacker News (official RSS)(cedana.com)·· 2026-10-07品質評分66

GPU 估算工具:1M/B/T tokens 需要多少 GPU

How many GPUs is 1M/B/T tokens?

重點導讀

GPU 估算工具 讓使用者能快速計算在 1 億/兆/兆 tokens 的工作負載下,對應 Llama 3.3 70B 等模型所需的 GPU 數量。

  • 核心計算:1 兆 tokens/月(30 天)約 367 臺 H100 GPU,基於 1,000 tokens/s 的 H100 速度與 50% 利用率。

AI 伴隨問答

自備 Key 深度分析

針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:

正文

» GPU count

367

H100

GPUs / base

211

Low

853

High

46

8x

H100

nodes

1 trillion tokens in one month (30 days) on Llama 3.3 70B needs approximately 367 H100 GPUs. Confidence is moderate.

1101001k10k

Single GPUs

Nodes with 8 GPUs

Base estimatePlausible rangeLog axis. Select a row to see its calculation.

One model copy needs 3 V100 GPUs (77 GB against 32 GB). Treat the V100 result as theoretical.

» Calculation

How the calculator gets the H100 number

Each step changes one quantity. Read from left to right.

  1. 0

    1

    Start

    1 trillion

    tokens in one month (30 days)

  2. 0

    2

    ÷ 2,592,000 seconds

    385,802

    tokens per second, average

  3. 0

    3

    × 0.475 for the token mix

    183,256

    output-equivalent tokens per second

  4. 0

    4

    ÷ 50% utilization

    366,512

    tokens per second of installed capacity

  5. 0

    5

    ÷ 1,000 tok/s for each H100

    367

    H100 GPUs

gpus = tokens ÷ seconds × [output share + input share × input cost] ÷ utilization ÷ tok/s for each gpu

» Sensitivity

Which assumption moves the H100 result most

Each bar shows the GPU count when one assumption moves from its best case to its worst case. The other assumptions stay at the base value. The longest bar is the assumption to measure first.

H100 speed for this modelmeasurement uncertainty

244 at 1,500 tok/s

814 at 450 tok/s

Average utilizationyour operating choice

229 at 80%

611 at 30%

Cost of an input tokenmeasurement uncertainty

251 at 0.1 of output

482 at 0.5 of output

Measurement uncertaintyYour operating choiceVertical line: base, 367 H100 GPUs before rounding

» Per-GPU detail

GPUtok/s per GPUTokens per GPU per dayGPUs per copyLowBaseHigh8-GPU nodes
A10045040.93M14368151,966102
V10014413.10M31,3382,5477,122319
H1001,00090.95M121136785346
B2004,000363.79M1479225212

» Model presets and evidence

Reference speed is tokens per second for one GPU, as low, base, and high. "Measured" values come from public benchmarks across loose and strict latency targets. "Estimated" values have no direct benchmark.

ModelTotal / activeWeightsReference speedB200 ÷ H100BasisSource
Llama 3.1 8B8B / 8B8 GBH100: 3,000, 6,000, 12,5002.3, 4, 6Estimated. One measured point for the high value.morph »
Llama 3.3 70B70B / 70B70 GBH100: 450, 1,000, 1,5002.3, 4, 6Measured on H100 and B200.inferencex »
cerebrium »
gpt-oss 120B117B / 5.1B65 GBH100: 740, 1,400, 2,6006, 9, 12Measured on H100 and B200. Weight size is not verified.inferencex »
DeepSeek V4 Flash284B / 13B160 GBB200: 1,500, 4,000, 9,0005, 9, 14Estimated. No benchmark found. Scaled from V4 Pro and gpt-oss by active parameters.datacamp »
MiniMax M3 428B428B / n/a428 GBH100: 157, 240, 5373.5, 4.5, 5.3Measured on H100 and B200. Weight size assumes 8-bit.inferencex »
DeepSeek R1 / V3671B / 37B671 GBH100: 23, 75, 26612, 16, 21Measured on H100 and B200. Weight size assumes 8-bit.inferencex »
GLM-5.3753B / 40B755 GBB200: 1,400, 3,000, 7,0004, 8, 16Measured on B200 only. Sources disagree: GLM-5 at FP4 gives 1,417 to 2,935, GLM-5.3 gives 4,329 to 10,243. H100 ratio is estimated.inferencex 5.3 »
inferencex 5 fp4 »

size »
DeepSeek V4 Pro1600B / 49B865 GBB200: 468, 1,000, 2,8006, 14, 21Measured on B200 only. H100 ratio is estimated from DeepSeek R1.inferencex »
size »

» Other assumptions

AssumptionValueBasisSource
A100 ÷ H1000.35, 0.45, 0.60H100 measured at 1.8x to 2.9x the A100 on 70B modelshyperstack »
perplexity »
V100 ÷ H1000.10, 0.18, 0.25Specifications only: 125 against 312 TFLOPS, 900 against 2,039 GB/s. No LLM serving benchmark.spheron »
Input cost0.1, 0.3, 0.5Derived from 128:128 and 2024:128 runs on 4x H100 (result 0.33)e2e networks »
GPU memory32, 80, 80, 180 GBV100, A100, H100, B200. The fit check adds 10% to the weight size.inferencex »

Values read on 2 October 2026.

» Limits of this model

  • The result is a planning estimate. Expect an error of 2x in each direction for measured presets, and more for estimated presets. Measure your model on your hardware before you buy.
  • Speed changes with the latency target. The low and high values of each preset come from strict and loose latency targets.
  • For large mixture-of-experts models, the B200 lead over the H100 is 4x to 21x. FP4 support and memory size cause this. The ratio is specific to each model.
  • A100 and V100 results for the large models are theoretical. One model copy needs more GPUs than one node contains.
  • The V100 value has no measured LLM serving data. It comes from hardware specifications.
  • The range combines independent errors as a root sum of squares in log space. It is not a statistical confidence interval.
  • The calculator does not include prompt caching, speculative decoding, reasoning-token overhead, failures, or regional duplication.

讀者討論

理性交流 · 即時討論