āĻĒā§āϰāϝ⧁āĻ•ā§āϤāĻŋāĻ—āϤ āĻ—āĻžāχāĻĄ

LLM āĻŸā§āϰ⧇āύāĻŋāĻ‚ FLOPs āĻ…āύ⧁āĻŽāĻžāύ āĻ•āϰāĻž

For a dense language model, C ≈ 6ND estimates training arithmetic from parameter count N and processed training tokens D.

āĻāχ āύāĻŋāĻ°ā§āĻĻ⧇āĻļāĻŋāĻ•āĻž āĻļ⧁āύ⧁āύāĻāφāχ-āĻŦāĻ°ā§āĻŖāĻŋāϤ ¡ ~3 min

āĻāχ āĻĒ⧃āĻˇā§āĻ āĻžāϝāĻŧ3 āĻŽāĻŋāύāĻŋāϟ āĻĒāĻĄāĻŧāĻž āĻšāϝāĻŧ⧇āϛ⧇
  1. āĻ“āĻ­āĻžāϰāĻ­āĻŋāω
  2. āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ
  3. āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ
  4. The Future of Estimating LLM Training FLOPs
  5. āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇
  6. āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ
  7. āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

āĻ“āĻ­āĻžāϰāĻ­āĻŋāω

Converting that operation count into time also requires realistic accelerator throughput and utilization. The result is a planning approximation, not a promise about runtime, memory, cost, or model quality.

āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ

Training compute counts arithmetic operations, while FLOP/s measures how quickly a system performs them. Keep those units separate. For a conventional dense Transformer, a common first estimate is C ≈ 6ND, where N is the relevant parameter count and D is the total number of tokens processed during training. Repeated passes through the same tokens count again. State how embeddings and other parameters are counted so comparisons use the same convention. The factor six approximates the main parameter-matrix work: roughly 2N operations per token in the forward pass and 4N in the backward pass. It is not an exact count of every operation. Attention, long sequences, architecture differences, and implementation choices can require a more detailed estimate. Applying the dense formula to a mixture-of-experts model using all stored parameters can be misleading because only some experts are active for each token. Consider a constructed example with N = 10⁚ and D = 2 × 10š⁰. Multiplying gives C ≈ 1.2 × 10²⁰ FLOPs. Suppose eight accelerators each have a relevant peak of 100 × 10š² FLOP/s, with an assumed model FLOPs utilization of 0.5. Their effective model throughput is 4 × 10š⁴ FLOP/s. Dividing compute by throughput gives 300,000 seconds, about 83.3 hours. Eight devices running that long represent about 667 accelerator-hours. Measure a representative training pilot before committing to a schedule. Use the intended sequence length, batch size, precision, software, and device arrangement. Record both tokens per second and what the timing includes. Add explicit allowances for evaluation, checkpoints, interruptions, and experimentation when those activities fall outside the measurement. The arithmetic estimate alone does not show whether the model fits in memory.

āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ

A hypothetical dense model with one billion parameters processes twenty billion tokens. The 6ND estimate is 1.2 × 10²⁰ floating-point operations.

A team assumes eight accelerators, each rated at 100 TFLOP/s for the relevant precision, and 50% model FLOPs utilization. Effective model throughput is 400 TFLOP/s, giving about 83.3 hours for the example run.

A researcher processes a ten-billion-token corpus twice. D is twenty billion processed tokens, even though the unique corpus contains ten billion tokens.

A training pilot reaches only half the estimated tokens per second. The team revises the schedule using observed throughput instead of treating the peak chip rating as sustained performance.

The Future of Estimating LLM Training FLOPs

Training systems will continue changing their numerical formats, kernels, parallel execution, and memory strategies. Those changes can alter the useful throughput achieved for an otherwise similar model. Maintain a small estimation sheet with the parameter convention, token budget, hardware assumptions, measured pilot rate, and excluded activities. Update the sheet when the configuration changes instead of reusing a utilization percentage from an unrelated benchmark. Compare the estimate with the completed run to improve future planning, and retain a range when throughput or interruption rates remain uncertain.

āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇

āϕ⧇āύ āĻāϟāĻž āϗ⧁āϰ⧁āĻ¤ā§āĻŦāĻĒā§‚āĻ°ā§āĻŖ

  • āĻ–āϰāϚ āĻ“ āĻŦāĻžāĻœā§‡āϟ. āφāĻ°ā§āĻ•āĻŋāĻŸā§‡āĻ•āϚāĻžāϰ⧇āϰ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤāϗ⧁āϞāĻŋ āĻŦāĻ›āϰ⧇āϰ āĻĒāϰ āĻŦāĻ›āϰ āϧāϰ⧇ āĻ•āĻ°ā§āĻŽāĻ•ā§āώāĻŽāϤāĻž āĻāĻŦāĻ‚ āĻ…āĻĒāĻžāϰ⧇āϟāĻŋāĻ‚ āĻ–āϰāϚ āϚāĻžāϞāĻžāϝāĻŧāĨ¤
  • āϏ⧁āĻ¸ā§āĻĒāĻˇā§āϟ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤ. āĻ•āĻžāϰāĻŋāĻ—āϰāĻŋ āĻļāĻŋāĻ•ā§āώāĻž āĻĻāϞāϗ⧁āϞ⧋āϕ⧇ āϏāĻ āĻŋāĻ• āĻ¸ā§āĻŸā§āϝāĻžāĻ• āĻŦ⧇āϛ⧇ āύāĻŋāϤ⧇ āϏāĻžāĻšāĻžāĻ¯ā§āϝ āĻ•āϰ⧇, āĻļ⧁āϧ⧁ āύāϤ⧁āύāϟāĻŋ āύāϝāĻŧāĨ¤
  • āĻŽāĻžāύ āύāĻŋāϝāĻŧāĻ¨ā§āĻ¤ā§āϰāĻŖ. āĻ­āĻžāϞ⧋ āχāĻžā§āϜāĻŋāύāĻŋāϝāĻŧāĻžāϰāĻŋāĻ‚ āĻĒāĻ›āĻ¨ā§āĻĻ āĻ‰ā§ŽāĻĒāĻžāĻĻāύ⧇ āύāĻŋāĻ°ā§āĻ­āϰāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āϘāϟāύāĻž āĻ•āĻŽāĻŋāϝāĻŧ⧇ āĻĻ⧇āϝāĻŧāĨ¤

āϜāĻ¨ā§āϝ āϏāϤāĻ°ā§āĻ•

  • āĻāĻ•āϟāĻŋ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ• āĻ…āĻĒā§āϟāĻŋāĻŽāĻžāχāϜ āĻ•āϰāĻž āĻŦ⧃āĻšāĻ¤ā§āϤāϰ āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽ āĻĻ⧁āĻ°ā§āĻŦāϞāϤāĻž āφāĻĄāĻŧāĻžāϞ āĻ•āϰāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤
  • āĻ…āĻŦāĻ•āĻžāĻ āĻžāĻŽā§‹ āĻāĻŦāĻ‚ āϰāĻ•ā§āώāĻŖāĻžāĻŦ⧇āĻ•ā§āώāϪ⧇āϰ āĻ–āϰāϚ āĻĒā§āϰāĻžāϝāĻŧāχ āĻ…āĻŦāĻŽā§‚āĻ˛ā§āϝāĻžāϝāĻŧāύ āĻ•āϰāĻž āĻšāϝāĻŧāĨ¤
  • āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽāϗ⧁āϞāĻŋ āφāϰāĻ“ āϜāϟāĻŋāϞ āĻšāĻ“āϝāĻŧāĻžāϰ āϏāĻžāĻĨ⧇ āϏāĻžāĻĨ⧇ āϏ⧁āϰāĻ•ā§āώāĻž āĻāĻŦāĻ‚ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āĻĢāĻžāρāĻ• āĻŦāĻžāĻĄāĻŧāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻĒāĻĻāĻ•ā§āώ⧇āĻĒ

  1. āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ⧇āϰ āφāϗ⧇ āĻŦāĻŋāϞāĻŽā§āĻŦ, āϗ⧁āĻŖāĻŽāĻžāύ āĻāĻŦāĻ‚ āĻ–āϰāĻšā§‡āϰ āϞāĻ•ā§āĻˇā§āϝ āύāĻŋāĻ°ā§āϧāĻžāϰāĻŖ āĻ•āϰ⧁āύāĨ¤
  2. āĻŦāĻžāĻ¸ā§āϤāĻŦāϏāĻŽā§āĻŽāϤ āϞ⧋āĻĄ āĻāĻŦāĻ‚ āĻĄā§‡āϟāĻž āĻ…āĻŦāĻ¸ā§āĻĨāĻžāϰ āĻ…āϧ⧀āύ⧇ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ•āĨ¤
  3. āĻ¤ā§āϰ⧁āϟāĻŋ, āĻĒā§āϰāĻŦāĻžāĻš, āĻāĻŦāĻ‚ āĻŦā§āϝāĻŦāĻšāĻžāϰāĻ•āĻžāϰ⧀āϰ āĻĒā§āϰāĻ­āĻžāĻŦ⧇āϰ āϜāĻ¨ā§āϝ āϝāĻ¨ā§āĻ¤ā§āϰ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāĨ¤
  4. āĻ¸ā§āϕ⧇āϞ āĻ•āϰāĻžāϰ āφāϗ⧇ āϰ⧋āϞāĻŦā§āϝāĻžāĻ• āĻāĻŦāĻ‚ āϘāϟāύāĻžāϰ āĻĒā§āϰāϤāĻŋāĻ•ā§āϰāĻŋāϝāĻŧāĻž āĻĒāĻžāĻĨ āĻĒā§āϰāĻ¸ā§āϤ⧁āϤ āĻ•āϰ⧁āύāĨ¤

āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ

āĻĢā§āϰāĻŋ āύāĻŋāωāϜāϞ⧇āϟāĻžāϰ

āĻĒā§āϰāϤāĻŋāĻĻāĻŋāύ⧇āϰ āĻāφāχ āĻŦā§āϰāĻŋāĻĢāĻŋāĻ‚ āĻĒāĻžāύ

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ āϏāĻ•āĻžāϞ⧇ āϤāĻŋāύāϟāĻŋ āϝāĻžāϚāĻžāχāĻ•ā§ƒāϤ āĻāφāχ āĻ¸ā§āĻŸā§‹āϰāĻŋ, āϏāĻšāϜ āχāĻ‚āϰ⧇āϜāĻŋāϤ⧇ āϞ⧇āĻ–āĻžāĨ¤ āϚāĻŋāϰāĻ•āĻžāϞ⧇āϰ āϜāĻ¨ā§āϝ āĻĢā§āϰāĻŋ, āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχāĨ¤

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ⧇ āĻāĻ•āϟāĻŋ āχāĻŽā§‡āϞāĨ¤ āĻāĻ• āĻ•ā§āϞāĻŋāϕ⧇āχ āφāύāϏāĻžāĻŦāĻ¸ā§āĻ•ā§āϰāĻžāχāĻŦ āĻ•āϰ⧁āύāĨ¤ āφāĻŽāϰāĻž āĻ•āĻ–āύ⧋āχ āφāĻĒāύāĻžāϰ āĻ āĻŋāĻ•āĻžāύāĻž āĻŦāĻŋāĻ•ā§āϰāĻŋ āĻŦāĻž āĻļā§‡ā§ŸāĻžāϰ āĻ•āϰāĻŋ āύāĻžāĨ¤

āύāĻŋāĻœā§‡āϕ⧇ āĻĒāϰ⧀āĻ•ā§āώāĻž āĻ•āϰ⧁āύ

Take the Estimating LLM Training FLOPs quiz

āĻĒā§āϰāϤāĻŋāϟāĻŋ āωāĻ¤ā§āϤāϰ⧇āϰ āϤāĻžā§ŽāĻ•ā§āώāĻŖāĻŋāĻ• āĻĢāĻŋāĻĄāĻŦā§āϝāĻžāĻ•, āĻāĻŦāĻ‚ āϕ⧋āĻ°ā§āϏ āĻĒāĻžāϏ āĻ•āϰāĻžāϰ āĻĒāϰ āϝāĻžāϚāĻžāχāϝ⧋āĻ—ā§āϝ āφāχāĻĄāĻŋ āϏāĻš āĻāĻ•āϟāĻŋ āĻļā§‡ā§ŸāĻžāϰāϝ⧋āĻ—ā§āϝ āϏāĻžāĻ°ā§āϟāĻŋāĻĢāĻŋāϕ⧇āϟāĨ¤

āϕ⧁āχāϜ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ

āĻĢā§āϰāĻŋ āĻāφāχ āĻļāĻŋāĻ•ā§āώāĻžāϕ⧇ āϏāĻŽāĻ°ā§āĻĨāύ āĻ•āϰ⧁āύāĨ¤ AI Understanding āĻāĻ•āϟāĻŋ 501(c)(3) āĻ…āϞāĻžāĻ­āϜāύāĻ• āĻĒā§āϰāϤāĻŋāĻˇā§āĻ āĻžāύ — āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχ, āϕ⧋āύ⧋ āĻĒ⧇āĻ“ā§ŸāĻžāϞ āύ⧇āχ, āĻ•āĻ–āύ⧋āχ āύāĻžāĨ¤ āĻāĻ•āϟāĻŋ āĻĻāĻžāύ āĻ•āϰ⧁āύ

āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

What is Estimating LLM Training FLOPs?

For a dense language model, C ≈ 6ND estimates training arithmetic from parameter count N and processed training tokens D. Converting that operation count into time also requires realistic accelerator throughput and utilization. The result is a planning approximation, not a promise about runtime, memory, cost, or model quality.

In the dense-model approximation C ≈ 6ND, which quantities do N and D represent?

N represents the parameter count under the chosen convention; D counts tokens processed during training.

A ten-billion-token corpus is processed twice. Which D belongs in the 6ND estimate?

D counts processed tokens, so two passes over ten billion tokens contribute twenty billion tokens.

Which distinction between FLOPs and FLOP/s is needed when estimating a training schedule?

Divide total operations by operations per second to obtain seconds.

Eight accelerators each have a relevant peak of 100 TFLOP/s and assumed model FLOPs utilization of 50%. What effective model throughput is used?

8 × 100 × 0.5 = 400 TFLOP/s. A peak rating alone would omit the utilization assumption.

The worked run takes about 83.3 elapsed hours on eight accelerators. Approximately how many accelerator-hours does it consume?

Multiply elapsed hours by device count: 83.3 × 8 ≈ 667 accelerator-hours.

āĻļāĻŋāĻ–āϤ⧇ āĻĨāĻžāϕ⧁āύ

āϏāĻŽā§āĻĒāĻ°ā§āĻ•āĻŋāϤ āĻ—āĻžāχāĻĄ

āĻāχ āĻŦāĻŋāώāϝāĻŧ⧇āϰ āϜāĻ¨ā§āϝ āĻŦāĻžāĻ›āĻžāχ āĻ•āϰāĻž āφāϰāĻ“ āĻ—āĻžāχāĻĄ

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āφāĻĒāĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻ—āĻžāχāĻĄ

LLM API āĻ–āϰāϚ āĻāĻŦāĻ‚ āĻŸā§‹āϕ⧇āύ āĻŦāĻžāĻœā§‡āϟ āĻ…āύ⧁āĻŽāĻžāύ āĻ•āϰāĻž

āĻŸā§‡āĻ•āύāĻŋāĻ•ā§āϝāĻžāϞ

āĻāĻ•āϟāĻŋ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ AI āĻļāĻ‚āϏāĻžāĻĒāĻ¤ā§āϰ āĻĒāĻžāύāĻāĻ•āϟāĻŋ āϏāĻ‚āĻ•ā§āώāĻŋāĻĒā§āϤ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āϕ⧋āĻ°ā§āϏ āĻāĻŦāĻ‚ āĻĒāϰ⧀āĻ•ā§āώāĻž. āϕ⧋āύ⧋ āĻ–āϰāϚ āύ⧇āχ, āϕ⧋āύ⧋ āĻ•āĻžāĻ°ā§āĻĄ āύ⧇āχāĨ¤ āĻĒā§āϰāĻžāϝāĻŧ 2.5 āϘāĻ¨ā§āϟāĻžāĨ¤āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ →