Computing power (computer, chip, TOPS, DMIPS, MACs, etc.) related abbreviations and definitions

1. MAC

• MAC: memory access cost, memory access cost

MAC describes the amount of parameters of a neural network, i.e. the memory space required to store that model. For example, a certain model requires 256,000 floating-point parameter definitions, converted to bits and multiplied by 32 to get 8,192,000 bits, and then divided by 8 to convert to Byte, 1024KB, which is 1M, then the size of this model is about 1M.

2. MACs、MACC、MADD、MACCs、MADDs

• MACs = MACC = MADD = MACCs = MADDs: multiply-accumulate operations, multiply-add operations, multiplication and addition operands

MACs, MACC, MADD, MACCs, and MADDs have the same meaning; one MACC instruction is two OPs, that is, 1MACC ≈ 2OP;

3. FLOPs

• FLOPs: floating point operations, floating point operands

Note that s is lowercase, which can be used to measure the complexity of the algorithm/model

4. FLOPS、MFLOPS、GFLOPS、TFLOPS、PFLOPS

• FLOPS: floating point operations per second, floating point operations per second

Note that the uppercase S is an indicator to measure hardware performance

• MFLOPS: mega floating point operations per second, million floating point operations per second, 10 ⁶ FLOPS

• GFLOPS: giga floating point operations per second, one billion floating point operations per second, 10 ⁹ FLOPS

• TFLOPS: tera floating point operations per second, trillion floating point operations per second, 10 ¹² FLOPS

• PFLOPS: peta floating point operations per second, quadrillion floating point operations per second, 10 ¹⁵ FLOPS

5. MIPS、DMIPS

• MIPS: million instructions per second, million instructions per second

• DMIPS: dhrystone million instructions executed per second, dhrystone executes one million instructions per second

where dhrystone is an integer arithmetic test program

6. OPS、MOPS、GOPS、TOPS、POPS

• OPS: operations per second, the number of operations per second

Note that the uppercase S is an indicator to measure hardware performance. The specific evaluation of computing power needs to be combined with data precision (such as int8, fp16, fp32). It can be converted with FLOPS only after combining the specific data type precision.

On the basis of determining the size of the MACC array, computing power = frequency * number of MACCs * 2 (a single computing unit can complete multiplication and addition); for example: suppose there are 512MACC computing units, the operating frequency is 1GHz, the data structure and precision of INT8, computing power It is 512 x 2 (2 understands that one MACC is one multiplication and one addition, which is two operations) x 1 GHz = 1000 Billion Operations/Second = 1 TOPS(Tera-Operations/second). FP16 precision is 0.5TOPS, and FP32 precision is 0.25TOPS.

In some cases, TOPS/W is also used as a performance index to evaluate the computing power of the processor. TOPS/W is used to measure how many trillion operations the processor can perform under the condition of 1W power consumption.

• MOPS: mega operations per second, million operations per second, 10 ⁶ OPS
• GOPS: giga operations per second, one billion operations per second, 10 ⁹ OPS
• TOPS: tera operations per second, trillion operations per second, 10 ¹² OPS
• POPS: peta operations per second, quadrillion operations per second, 10 ¹⁵ OPS