What are FLOPS in Computing and How Do They Impact You?

What are FLOPS in computing and how do they impact you?

Today, I’ll be tackling these questions and break down what this metric actually means for you, the end consumer, and when you should or shouldn’t pay mind to it.

Let’s get into it!

What are FLOPs in Computing?

A FLOP is a Floating Point Operation and FLOPS are Floating Point Operations per Second, which are fairly commonplace in computing.

Your CPU, though, for example, isn’t going to report its performance to you in FLOPS— it will report its performance in GHz. Your CPU is still performing Floating Point Operations in every one of those billions of cycles per second, though— as is your GPU.

The easiest way to break down FLOPS is to remember that your CPU and GPU are both effectively performing rapid-fire math at any given point of operation.

The computing power of any given component, then, can often be measured in expected FLOPS, though I’ll get into more detail about how FLOPS are actually measured in the next section.

Not all FLOPS are the same, either. “Floating Point” refers to the level of numerical precision being used in a given workload.

For example, FP16 uses 16 bits for half-precision calculations, while FP32 uses 32 bits for single precision calculations. FP64 also exists for double-precision calculations. Integers, in comparison, have no decimals, while FP16 has 4, and FP32 up to 9 (depending on how big the number is).

You can probably already tell that simpler numbers (such as integers) will calculate faster on most hardware, albeit at lower precision, but let’s not get ahead of ourselves. Here’s what Linus has to say about FLOPS:

The level of floating point precision required for a given workload will depend on the workload in question. But in general, higher levels of floating point precision like FP32 and FP64 are used for scientific purposes, like neural networks or medical research.

Lower levels of Floating Point Precision require less memory and perform faster, but are also less ideal for certain workloads.

Some workloads may even use mixed precision. For example, Nvidia supports mixed precision for training neutral networks on their GPUs, using FP16 most of the time for performance and only using FP32 when absolutely necessary.

How are FLOPS Measured?

FLOPS are most frequently measured in Teraflops (or Gigaflops) when used in marketing.

Image Credit: NVIDIA

You may also see that number provided alongside the level of floating point precision (FP16, FP32, etc.) being used for that Teraflops number since higher precision on the same hardware will lower the FLOPS per second.

Does More FLOPS Equal More Performance?

Well…yes and no, especially if we’re comparing GPU architectures.

Ultimately, a FLOPS number is given as an estimate of raw compute performance.

The problem is, raw compute performance doesn’t tell the full story for any CPU or GPU. These components don’t exist in isolation inside any machine: your RAM and storage will also play a considerable part in your final performance numbers.

How Do FLOPS Impact You?

Truthfully…they probably don’t.

It’s similar to how Gigahertz doesn’t necessarily correlate to real-world performance.

A 3 GHz CPU from 10 years ago is not going to perform anything like a 3 GHz CPU made today, for example. For comparing hardware, FLOPS calculations make the most sense when comparing CPUs or GPUs with the same underlying architecture.

The workloads where you actually should be paying attention to maximum Teraflops aren’t gaming or even editing or rendering.

Raw compute presented in this form is most relevant to data research and analytics, deep learning, and supercomputers.

Gamers and prosumers should be looking more toward actual benchmarks for the applications they’re using rather than the floating point numbers for a given piece of hardware.

Actual hardware testing is still what matters most for determining real-world performance, not hardware specifications presented in isolation.

Over to You

And that’s all!

I hope this article adequately addressed what FLOPS are and how they actually impact you, the end user.

If you have any other questions about PC hardware or how it works, feel free to ask in the comments section below, where me or another CGD team member will be happy to help you!

Alternatively, you can also head to our Forum for longer-form tech discussions with other Enthusiasts and Experts.

Until then or until next time, happy computing!