[KongchangAI]
Concept每秒万亿次浮点运算

TFLOPS

A metric of hardware computing performance denoting trillions of floating-point operations per second. It is commonly used to assess the compute capability of GPUs and other accelerators, though actual effective throughput is affected by factors such as memory bandwidth.

Timeline (last 90 days)

Nov 14

The TFLOPS calculation formula is 2×M×N×K / elapsed time (seconds) / 1e12, where the factor 2 represents one multiply-add operation.

Unverified50%
Nov 1

The trainer automatically detects the hardware configuration (NVIDIA, AMD or CPU) and runs matrix multiplication benchmarks to measure the hardware's raw TFLOPS compute power.

Unverified50%
Nov 1

TFLOPS is a common metric for measuring hardware computational throughput; the nominal TFLOPS of consumer-grade GPUs usually refers to FP32 or FP16 peak, but the effective compute in actual training is often significantly below peak.

Unverified50%
Oct 25

TFLOPS is a theoretical peak calculated based on core count and clock frequency, assuming all execution units work at full load at all times.

Unverified50%

All Facts (4)

Source Articles