Here is a step-by-step estimate for the total training FLOPs of GPT-3.

Step 1: Identify the key variables

To estimate training compute for a dense transformer, we need two numbers:

• $N$ (Parameters): GPT-3 has 175 billion (1.75 × 10¹¹) parameters.
• $D$ (Tokens): GPT-3 was trained on roughly 300 billion (3 × 10¹¹) tokens. (This is the widely cited figure from the original paper for the final training corpus).

Step 2: Use the standard training FLOPs formula

For decoder-only transformers trained with backpropagation, the total floating-point operations (FLOPs) can be approximated as:

$$\text{FLOPs} \approx 6 \times N \times D$$

Why 6?

• Forward pass: Processing one token requires roughly 2 FLOPs per parameter (1 multiply + 1 accumulate for the matrix operations).
• Backward pass: Computing gradients requires roughly 4 FLOPs per parameter (approximately 2× the forward cost to calculate gradients with respect to both inputs and weights).
• Total: $2 + 4 = 6$ FLOPs per parameter per token.

(Note: This formula dominates the cost; the quadratic attention cost is negligible compared to the linear parameter cost for a model of this size).

Step 3: Calculate

$$\text{FLOPs} \approx 6 \times (1.75 \times 10^{11}) \times (3 \times 10^{11})$$

$$= 6 \times 1.75 \times 3 \times 10^{22}$$

$$= 6 \times 5.25 \times 10^{22}$$

$$= 31.5 \times 10^{22}$$

$$= \mathbf{3.15 \times 10^{23} \text{ FLOPs}}$$

Step 4: Account for uncertainty and overhead

The raw calculation gives ~3.15 × 10²³ FLOPs. However, we should adjust for real-world factors:

• Dataset size variance: If GPT-3 was trained on closer to 400–500B tokens (including potential epochs or different data mixes), the number could be 4–5 × 10²³ FLOPs.
• Implementation overhead: Data loading, checkpointing, communication in distributed training, and the Adam optimizer updates add 10–20% overhead.

Final Estimate

Approximately $3 \times 10^{23}$ to $5 \times 10^{23}$ FLOPs (roughly 300 to 500 sextillion FLOPs).

For context, this is equivalent to about 3,000 to 5,000 petaflop-seconds (or PFLOP-seconds), which aligns with the estimated thousands of petaflop-days often cited for GPT-3's training run on V100 GPUs.

Below, I'll provide a step-by-step estimation of the total floating-point operations (FLOPs) required to train GPT-3, based on my built-in knowledge of the model (from its original 2020 paper and related analyses). I'll avoid any external searches or real-time data lookups, relying solely on established facts and standard approximation methods for large language model (LLM) training. This is an estimate, as exact figures can vary slightly based on implementation details (e.g., optimizer choices, mixed-precision training, or hardware efficiencies), but it should be reasonably accurate.

Step 1: Recall Key Specifications of GPT-3

• Number of parameters (N): GPT-3 is a transformer-based model with 175 billion parameters. This is the total number of trainable weights in the model (often denoted as 175 × 10^9 or 1.75 × 10^11).
• Training dataset size (T): GPT-3 was trained on a massive dataset filtered from sources like Common Crawl, WebText2, Books1, Books2, and Wikipedia. Based on the original paper, this amounts to approximately 300 billion tokens (3 × 10^11 tokens). Tokens here refer to subword units (e.g., via Byte-Pair Encoding), which is the standard way to measure LLM training data volume.
  • Note: This is a one-pass training setup (the model sees the data once), without heavy data augmentation or multiple epochs, which is typical for such large-scale pre-training.

These figures are foundational and well-documented in the literature, so they're a solid starting point for estimation.

Step 2: Understand the FLOPs Calculation for Transformer Training

To estimate training compute, we use a well-established rule of thumb for transformer models like GPT-3. The total FLOPs is approximated as:

Total FLOPs ≈ 6 × N × T

• Why this formula?
  • In transformer architectures (which GPT-3 uses), training involves processing the input sequence token by token.
  • For each token:
    • Forward pass: Involves matrix multiplications and activations across all layers, costing roughly 2 FLOPs per parameter (one for the multiply, one for the add in GEMM operations, scaled by the model's dense structure).
    • Backward pass (gradient computation): This is roughly twice as expensive as the forward pass due to backpropagation through the chain rule, adding about 4 FLOPs per parameter.
    • Total per token: ~6 FLOPs per parameter.
  • This 6× factor is a standard approximation derived from detailed analyses (e.g., by Kaplan et al. in their scaling laws work and confirmed in the GPT-3 paper). It assumes full-precision floating-point ops and ignores minor overheads like activations or embeddings (which are small relative to the dominant matrix multiplies).
  • The formula scales linearly with N (more parameters mean more ops per token) and T (more data means more tokens processed).
  • Caveats: Real-world training might be 10-20% more or less efficient due to optimizations (e.g., FP16 precision reduces FLOPs but increases memory use, or sparsity isn't used in base GPT-3). But for estimation, 6×N×T is widely accepted and matches reported figures for models like GPT-3.

This doesn't include pre-training setup costs (e.g., data curation) or fine-tuning, just the core pre-training phase, which dominates the total compute.

Step 3: Plug in the Numbers

Now, let's compute it explicitly:

• N = 175 × 10^9 = 1.75 × 10^11
• T = 300 × 10^9 = 3 × 10^11
• First, compute N × T = (1.75 × 10^11) × (3 × 10^11) = 5.25 × 10^22
• Then, multiply by 6: 6 × 5.25 × 10^22 = 3.15 × 10^23

So, the estimated total FLOPs is 3.15 × 10^23 (or 315 septillion FLOPs, in plain terms).

Step 4: Validate the Estimate with Context

• This aligns closely with the figure cited in the GPT-3 paper and follow-up analyses (around 3.14 × 10^23 FLOPs), confirming the approximation's accuracy.
• For perspective:
  • Training GPT-3 took about 1 month on a cluster of thousands of GPUs (e.g., NVIDIA A100s or V100s), each capable of ~10^14-10^15 FLOPs per second. At scale, this adds up to the estimated total.
  • Smaller models provide a sanity check: For example, GPT-2 (1.5 billion params, ~40 billion tokens) uses ~6 × 1.5e9 × 4e10 ≈ 3.6 × 10^20 FLOPs, which is ~1,000× less than GPT-3, matching the scaling.
• Uncertainties: If the token count was slightly higher (e.g., 410 billion in some filtered subsets), it could push the estimate to ~4 × 10^23. But 300 billion is the conservative, paper-stated value.

Step 5: Implications and Limitations

This massive compute requirement highlights why training models like GPT-3 is resource-intensive—it's equivalent to billions of human-years of calculation if done serially. In practice, it's distributed across supercomputers. If you're interested in breakdowns for inference FLOPs, comparisons to other models (e.g., GPT-4 estimates), or ways to optimize, let me know!

This estimation is based purely on my training data up to my last update, without any online queries. If new details have emerged since then, they wouldn't affect this historical estimate.