What Is a Teraflop (TFLOPS), and How Is It Different From TOPS?

TFLOPS stands for tera floating-point operations per second: one TFLOPS, or one teraflop, is a trillion (10¹²) calculations on floating-point numbers every second. FLOPS is the base measure, and the prefixes scale it: gigaflops for billions, teraflops for trillions, petaflops and exaflops for the largest supercomputers. A single teraflop is one trillion operations per second.

Floating-point numbers carry decimals and a wide range, which graphics, simulation and model training need. Integer formats such as INT8 are smaller and simpler, so hardware can run many more of them per second. That is why one chip can show very different figures: our NVIDIA RTX 2000 Ada card is listed at 12.0 TFLOPS FP32 and 192 AI TOPS, the second measured at a much lower precision.

Spec-sheet TFLOPS are theoretical peaks, worked out from the hardware rather than measured on a real program. The usual formula is the number of cores, times the clock speed, times the floating-point operations each core finishes per clock. Most GPU cores complete a fused multiply-add each cycle, which counts as two operations.

The RTX 2000 Ada shows how it works: 2,816 CUDA cores at a boost clock of about 2.1 GHz, times two operations, gives roughly 12 TFLOPS at FP32. Real applications reach only part of that peak, because they also wait on memory, data transfers and code that cannot run in parallel. TFLOPS tells you the ceiling, not the speed of your workload.

A TFLOPS figure means nothing without its precision. Halving the number of bits usually doubles the operations the same hardware can complete, so vendors often quote the lowest precision their hardware supports.

Two more details inflate headline numbers. Many AI figures assume structured sparsity, where half the model weights are zero, which doubles the quoted value. And Tensor Core figures apply only to matrix maths, not to the rest of a program.

TOPS uses the same tera prefix but counts operations of any kind, and in practice almost always means INT8 integer maths for AI inference. That makes TOPS the natural figure for NPUs and AI accelerators, which run quantized models, while TFLOPS remains the figure for GPUs doing graphics and general compute. Converting a model from FP32 to INT8 is what lets inference hardware use its TOPS rating, a step explained in our guide to INT8 quantization.

Knowing which precision a figure uses stops a spec sheet from misleading you. Our GPU vs CPU FAQ explains why GPUs lead on parallel maths, and the split between general and AI hardware inside a GPU is covered in CUDA cores vs Tensor Cores.

Edge hardware shows how varied the quoting has become. Low-profile workstation GPUs such as the RTX 2000 Ada lead with an FP32 TFLOPS figure, as desktop graphics cards do. The NVIDIA L4 inference card is rated at around 30 TFLOPS at FP32, with much higher figures quoted for FP8 and INT8 with sparsity. Jetson Orin modules are rated in INT8 TOPS, while NVIDIA rates the newer Jetson Thor modules in FP4 TFLOPS, up to 2,070 on the T5000.

None of these numbers is wrong, but each answers a different question. When two datasheets disagree, find the precision and whether sparsity is included before drawing any conclusion.

For an industrial or edge system, TFLOPS is a starting point for shortlisting GPUs, not the final answer. Memory size decides whether a model fits at all, memory bandwidth often limits real speed, and power and cooling decide whether the card can hold its clocks inside a sealed enclosure. A low-profile 70 W card with fewer TFLOPS can deliver more useful work in a fanless chassis than a larger card that overheats.

Our industrial GPU computers put that performance in a compact system, sized to the workload rather than the largest number on the datasheet.

We specify GPU, NPU and accelerator hardware from the workload, then fit and validate it in each system as standard. Send us your models, camera count and power budget, and we will size the hardware to match.