What are TFLOPS?

TFLOPS stands for Tera Floating-Point Operations Per Second. One TFLOP = one trillion (10¹²) floating-point math operations every second.

Think of it as the GPU's horsepower rating. A car's horsepower tells you how much mechanical work the engine can do. TFLOPS tells you how much mathematical work the GPU can do.

A floating-point operation is any basic math on decimal numbers: addition, subtraction, multiplication, or division. When a GPU has 990 TFLOPS at FP16, it can perform 990 trillion of these operations every second at half-precision (16-bit) floating point.

The same GPU has different TFLOPS ratings depending on the precision (number format) used. Lower precision = smaller numbers = more operations per cycle.

GIS uses FP16 TFLOPS as the standard metric because it's the most common precision for AI training and is reported by all major GPU vendors.

TFLOPS is a theoretical peak. Real-world performance is always lower because:

• Memory bandwidth bottleneck — the GPU can compute faster than memory can feed it data. This is called being "memory-bound."
• Utilization — not all cores are busy all the time. Real workloads rarely achieve 100% utilization.
• Software overhead — kernel launches, synchronization, data transfers all take time.

This is why memory_bandwidth_tbps matters alongside TFLOPS. A GPU with high TFLOPS but low bandwidth will be bottlenecked. The H100's 3.35 TB/s bandwidth is designed to keep its 990 TFLOPS fed.

TFLOPS appears in two places in a GIS document:

{
  "gpu": {
    "tflops_fp16": 989.5
  },
  "normalized": {
    "cost_per_tflop_hour": 0.00252
  }
}

gpu.tflops_fp16 is the raw compute power. normalized.cost_per_tflop_hour divides the hourly cost by TFLOPS to give you a price-performance ratio.

Lower cost_per_tflop_hour = more compute per dollar. This metric lets you compare a $2.49/hr H100 (990 TFLOPS) against a $1.10/hr A100 (312 TFLOPS) on equal footing.