Glossary
What is FLOPS?: rawcompute.in Glossary
FLOPS (Floating-Point Operations Per Second) is the fundamental unit for measuring computational performance of processors, GPUs, and accelerators.
FLOPS stands for Floating-Point Operations Per Second and measures how many arithmetic operations (addition, subtraction, multiplication, division) a processor can perform on floating-point numbers each second. FLOPS scales across orders of magnitude: GFLOPS (10^9), TFLOPS (10^12), PFLOPS (10^15), and EFLOPS (10^18). Modern AI training clusters are measured in PFLOPS. A cluster of 256 NVIDIA H100 GPUs delivers roughly 500 PFLOPS of FP16 compute.
FLOPS is a theoretical peak metric. Actual achieved performance depends on how well the workload maps to the hardware. Model FLOPs Utilisation (MFU) measures the percentage of theoretical peak that a training job achieves. Well-optimised LLM training typically achieves 40-60% MFU, meaning that a cluster with 500 PFLOPS theoretical might deliver 200-300 PFLOPS of effective compute. Memory bandwidth, communication overhead, and pipeline bubbles are the main sources of efficiency loss.
Why it matters when buying hardware
Use FLOPS to estimate the cluster size needed for your training job. If a model requires 10^23 FLOPs to train and you target a 30-day timeline with 50% MFU, you can calculate the required sustained PFLOPS and work backward to the number of GPUs. This kind of capacity planning is essential before making a multi-crore GPU cluster investment. Rawcompute.in can help you run these calculations and size your cluster appropriately for your training objectives and budget.