FLOPS means floating-point operations per second; a TFLOPS is a trillion (10¹²) per second. An H100 does about 989 TFLOPS (BF16), an L4 about 121.
Training a model takes roughly 6 × parameters × tokens operations, so TFLOPS let you estimate training time.
Example
The training cost estimator uses exactly this rule to estimate hours and taka.
বাংলায়: FLOPS / TFLOPS — প্রতি সেকেন্ডে কতগুলো দশমিক-সংখ্যার হিসাব — GPU-র গতির মাপ।