FLOPS
Also called: TFLOPS, PFLOPS, floating-point operations per second
The number of floating-point operations a chip performs per second. It is the unit of computing power for GPUs and AI chips.
FLOPS stands for “floating-point operations per second”, the number of operations on floating-point numbers carried out each second. It is used for GPUs, processors and AI accelerators. The multiples follow the metric system: one TFLOPS equals a trillion operations per second, one PFLOPS a thousand TFLOPS. Not to be confused with hashrate, which measures a miner's speed in hashes per second.
The advertised FLOPS are peak values, calculated from the number of compute units and the clock speed. In real use, only part of this maximum is reached, because the chip is often waiting for data coming from memory. This is especially true for text generation by a language model, which is limited mainly by memory bandwidth rather than raw computing power.
On a product page, a FLOPS figure only makes sense together with its precision: FP64, FP32, FP16, FP8 or FP4. The shorter the format, the higher the figure, and the “with sparsity” mention doubles it again. To compare two GPUs, use the same precision, then look at the amount of memory and its bandwidth. For training, BF16 or FP8 FLOPS are the most telling; to run a model locally, memory often matters more.
Our page comparing ASICs and GPUs helps you understand which machine to choose depending on whether you want to mine or work with AI.