license: apache-2.0
pretty_name: Measured GPU microbenchmarks for transformer training
tags:
- benchmark
- gpu
- performance
- roofline
- transformers
- training
configs:
- config_name: summary
data_files: summary.csv
- config_name: matmul
data_files: data/*/matmul.csv
- config_name: bandwidth
data_files: data/*/bandwidth.csv
- config_name: train_step
data_files: data/*/train_step.csv
Measured GPU microbenchmarks for transformer training
Measured numbers for the GPUs most people can actually get on Colab and Kaggle: matmul throughput across sizes, memory bandwidth, kernel launch overhead, and GPT-2-small (124M) training-step time, each recorded with the GPU clock, power and temperature during the run.
Summary
| GPU | Platform | fp16 burst TFLOP/s | fp16 model-shape TFLOP/s | fp32 TFLOP/s | Copy GB/s | Launch µs | Step ms | Tokens/s | Noise % | Clock MHz (step) | Power W (step) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Tesla T4 | colab | 40.18 | 23.09 | 4.16 | 232.9 | 8.15 | 516.55 | 15866 | 3.98 | 788 | 66.7 |
| NVIDIA L4 | colab | 75.1 | 49.06 | 15.9 | 229.9 | 8.67 | 283.11 | 28938 | 1.44 | 991 | 72.0 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | colab | 427.06 | 276.96 | 77.1 | 1459.1 | 3.32 | 45.0 | 182054 | 0.04 | 2338 | 504.2 |
Baseline step: GPT-2-small (124M), context 1024, batch 8, fp16 autocast, AdamW. "fp16 model shapes" is the median throughput over the four matmul shapes that step actually runs.
Method
- Device time from CUDA events; median of 20–50 iterations after warm-up.
- Before each training-step measurement, 10 s of matmuls bring the GPU to its steady-state clock.
- Clock, power and temperature are sampled every 0.5 s during the timed run (reading them afterwards is wrong: an idle GPU boosts instantly).
- fp32 means real fp32: TF32 is disabled. bf16 is excluded on GPUs without native bf16.
- FLOPs for the training step use 6 × params × tokens plus attention, a rule of thumb.
Files
data/<gpu>_<platform>/: matmul.csv, bandwidth.csv, train_step.csv, hardware.json
(the fact sheet), and per-GPU figures. summary.csv has one row per GPU. figures/ has
cross-GPU charts.
Caveats
- One chip per GPU type and one session each. Chips of the same model can differ (two T4s here ran ~10% apart in clock at the same power).
- Power-capped GPUs (T4, L4) vary 3–4% run to run with temperature; compare runs measured back to back, as ratios.
- Small buffers in the bandwidth sweep can fit in L2 cache and read far above DRAM bandwidth.
- Launch overhead is timed from Python, so it depends on the host CPU as well as the GPU.
- A second T4 session measured 34.7 µs launch overhead, likely host-CPU contention; the table uses the first session's 8.15 µs (Kaggle T4: 7.85 µs).
Contribute a GPU
Run the notebook from the accompanying repository on any NVIDIA GPU and open a pull request on this dataset with your results folder.