Making a Mountain out of a Mali: What a Billion Chips Really Means

As you might already know, given that we’ve mentioned it just a few times before, Mali is the number 1 shipping GPU in the world. More than 1 Billion Mali GPUs were shipped by our partners in 2016.

A BILLION.

Personally, that number is just a bit too big for me to visualise given that we usually only see hundreds, or possibly thousands, of any given thing in our day to day lives. I mean, I understand that 1Bn equals 10 to the power of 9, or a thousand times a million – but it’s a number so big that it’s really difficult to relate to the world around us. Probably the closest I can get is population. It means we effectively shipped a GPU for every single person in the US, UK, Indonesia, Pakistan and Brazil put together. Now it starts to make sense, but it got me thinking, are there any other ways we can quantify the success of Mali GPUs?

Dense crowd image

Since my time with Arm’s Media Processing Group I’ve learnt a lot about GPUs. The high performance roadmap is particularly interesting to me as we’ve seen such massive gains in the last couple of product generations, the Mali-G71 and Mali-G72, powering some of the latest and greatest flagship devices from massive names like Samsung and Huawei (more on that sooon…) The growth in computational power of these products has been phenomenal, with my latest phone, a Samsung Galaxy S8, outperforming my laptop in terms of graphics.  

So, we know there is a lot of computational power in the latest mobile GPUs, but how can we understand just how much? We could use scores from GFXBench as they’re easily available for almost any mobile device, but it would be difficult, if not impossible, to accumulate scores over thousands of unique Mali based devices. So, through consulting my more numerically blessed colleagues, we found a simpler and more elegant solution. They suggested we simply look at the amount of theoretical compute power in each device’s respective GPU and add it up based on the total number of shipments, giving us the total computational power shipped.

When talking about available compute power, a usual measure is the amount of programmable floating point operations per second, or FLOPs/sec. When we talk about GPUs, we usually have a lot of computation power – so we use giga-FLOPs/sec as our baseline unit. A GFLOP/sec equals one billion floating point operations per second (and just to be clear, we are talking about full precision floating point operations (FP32)).

Doing this maths is tricky so I sat back while the tech guys worked their magic, looking at all the different flavours of Mali GPUs out there, their clock frequencies, their configurations, shipment volume and so on. After much checking and double checking and many cups of coffee, they presented me with the magic number…

In the last year, new Mali GPUs provided a total 46,839,499,948 GFLOPs/sec of additional computing power to the world. Apparently, that equals 46.84 exa-FLOPs/sec (46.84 times 10 to the power of 18 floating point operations per second) – an incredibly large number by anyone’s standards.

Again though, I find myself puzzled by a massive number that I have no way of relating to. This got me thinking about other big numbers and how they’re quantified for us. I’m sure you’ve seen a vast transfer fee for a top-flight footballer and thought ‘200 million is an absolutely insane figure’ (mentioning no names). If you have, then chances are you’ve also seen those articles where the media try to make sense of it. You know the kind of thing, instead of buying this striker you could have X Bentleys, Y Lear Jets or N unicorns. So surely there must be something out there to put this computing power in perspective for an innumerate tech novice like me.

Since joining the tech industry, I have become weirdly fascinated with supercomputers. The fact that we build entire buildings, if not whole complexes and developments, dedicated to hosting a single computer is staggering. But on the other hand, these supercomputers allow us the ability to do all sort of clever simulations and calculations that can make a big difference to our lives, including weather forecasting, genome research and nuclear test simulation to name just a few. Surely, the supercomputers of the world must be in a league way beyond the computational capabilities of mobile GPUs…right?

Server image

The TOP500 list details the 500 most powerful supercomputers in the world. Currently the number 1 spot is held by the Sunway TaihuLight from the National Supercomputing Center in Wuxi, China, with 125 peta-FLOPs/sec computing power. Sounds pretty mindblowing doesn’t it? Well, it’s actually still several orders of magnitude less than the computational power provided by all those Mali GPUs!

In fact, we have to combine the total computing power of every one of the TOP500 supercomputers in the world to end up with 1.132 exa-FLOPs/sec of computational power and finally reach a number in the right order of magnitude!

Hopefully this gives you some insight into the massive amount of Mali-powered goodness there is to go around; in the past 12 months, Mali GPUs shipped 41.4 times the combined compute power of the TOP500 supercomputers in the world.

Finally there’s a number I can begin to understand, even if I’m still in awe of the fact that all this power is in consumer devices such as smartphones, VR devices, DTVs and more. It really puts it all in perspective.

Anonymous
Graphics & Multimedia blog