AI glossary · Basic · Efficiency & infra
FLOPS
FLOPS is the number of floating-point operations (operations on numbers with decimal points) a computer can perform per second, and it is the standard unit for describing the computing speed of AI chips.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Infrastructure & using models
- Fields
- Efficiency & infra
- Added
- 2026-10-12
How it works
Most of the computation in an AI model is multiplication and addition, and counting these one by one gives the number of operations. Chip specs state how many can be done per second as FLOPS, and because the numbers are large they add TFLOPS (trillions) or PFLOPS (quadrillions). The value varies greatly by precision, so spec sheets list a different number for each of FP32, BF16 and FP8. The lower the precision, the more operations the same chip can handle.
FLOPS vs. FLOPs
A common convention is that a capital S at the end means 'operations per second (speed)', while a lowercase s means 'total number of operations (amount)'. When a paper says 'training used X FLOPs of compute', it means the total. A common rough estimate for getting a feel is 'total training compute ≈ number of parameters × number of training tokens × about 6'. Dividing this total by the chip's actual per-second throughput gives a rough training time.
FLOPS vs. FLOPs
FLOPSOperations per second
- Chip speed
- Listed in GPU spec sheets
- Varies by precision
FLOPsTotal number of operations
- Compute a task requires
- Used for training compute
- Estimated from parameter and token counts
Where you'll see it
It appears mainly in performance tables in GPU and TPU spec sheets, cloud instance descriptions, and papers and articles on scaling laws. Some AI regulation discussions use total training compute as a threshold. It is also used in model comparisons when talking about 'compute per inference'.
Good to know
Spec sheet numbers are theoretical peaks, and actual training usually achieves far less because of time spent waiting on memory movement and communication. Some spec sheets also list values that assume 'sparsity', so check under what conditions a number was measured. Numbers at different precisions should not be compared side by side. LLM inference is often bottlenecked by memory bandwidth rather than compute, so FLOPS alone is a poor guide to inference speed.
Example
An engineer choosing a GPU looks at the BF16 TFLOPS value in the spec sheet to estimate training speed, but if the same line is marked 'with sparsity', they treat roughly half of that as the real baseline. When choosing an inference server, they look at VRAM capacity and memory bandwidth before FLOPS.