Blog

FLOPS vs FRAPS: cars and GPUs

What makes the difference where the rubber meets the road? FLOPS—Floating-point Operations Per Second is a measurement of a processor’s arithmetic processing speed and has been used since the early 1960s to evaluate and compare CPUs. FRAPS—FRAmes Per Second— is a benchmarking scheme for measuring how fast a graphics processor can update the screen. It is more commonly expressed as ...

Robert Dow

What makes the difference where the rubber meets the road?

FLOPS—Floating-point Operations Per Second is a measurement of a processor’s arithmetic processing speed and has been used since the early 1960s to evaluate and compare CPUs. FRAPS—FRAmes Per Second— is a benchmarking scheme for measuring how fast a graphics processor can update the screen. It is more commonly expressed as frames per second (fps). The use of FLOPS to characterize the performance of computers persists, even though it’s not a very accurate description. 

While speaking with one of our investor clients the other day, I found myself once again trying to explain why the FLOPS rating of a processor, especially a GPU, is not an absolute indicator of performance. I hit upon a metaphor with the car. 

FLOPS is comparable to horsepower (HP) in a car, and FRAPS (or fps, if you prefer) is what 0-to-60 time is to a car. So if two cars had a 400 HP engine, and one could do 0-to-60 in 5 seconds, while the other could make 0-to-60 in 3.5 seconds, the importance of the horsepower isn’t very relevant. Similarly, if one GPUs with 5,000 FLOPS can get 40 fps, while another (with 5,000 FLOPS) can deliver 60 fps, the FLOPS value becomes irrelevant. 

The differences are things, the stuff, between and within the engine and the end user (where the rubber meets the road). 

Within the engine there is the compression ratio, which is equivalent to the memory bandwidth in a GPU, and both have a great impact on the work output of the engine/device. The efficiency and number of valves in an engine is equivalent to the caches in a GPU. And you could make the analogy of cylinders to shaders. 

Other comparisons are the API. The API is the equivalent of the automobile’s flywheel, while the transmission of the automobile is the equivalent of the software driver of a GPU. To complete the analogy, the display is to the user, what the road is to the tires—where the rubber meets the road. 

The software driver and the architecture are the major (some might say few) places a GPU supplier can differentiate. Until recently, all GPUs were made at the same fab and used the same silicon mix; they are constrained by the same OS, and API, and even bounded by the same power supply and cooling system. And yet, two GPUs from two different companies, or two different generations, could have the same FLOPS capability and one would outperform the other. And, just as ratings can be capricious, your mileage may vary. 

Epilog 

Another comparison that could be considered is miles-per-gallon (MPG) as a compared to fps/Watts—power efficiency. 

It’s controversial to make such analogies, and as my grandmother said about noses and opinions, everybody’s got one, so when a comparison like this is offered, opinions get expressed as to what should added, changed, or removed. There is no direct comparison between a car and an AIB, it’s just an analogy, a metaphor to help someone not familiar with the nuances of an AIB or GPU relate to the importance of the parameters used to express such devices. 

And one more thing: you should never use unleaded gas in your GPU. s 

Trendspotting

The year shapes up