The last major architectural improvement to the basic GPU design happened when the major GPU suppliers went to unified shaders with the introduction of Direct3D 10, Shader Model 4.0, in 2008. That was ten years ago. The unified shading architecture was introduced with the Nvidia GeForce 8 series, ATI Radeon HD 2000, S3 Chrome 400, Intel GMA X3000 series, Xbox 360's GPU, Qualcomm Adreno 200 series, Mali Midgard, PowerVR SGX GPUs and is used in all subsequent series and designs.
In our 2017 summary report on GPUs, we listed 10 GPU suppliers. In the first half of 2018, we have expanded the list to 11, in 16 categories, shown in the following breakdown.
|GPU topology circa 1H’18|
If we expand the list to include AI/ML and GP SMID accelerators, we can include AMD’s Instinct, Fujitsu’s DLU, and Nvidia’s Volta.
There are two new entrants to the field. Intel is building a discrete chip after years of fielding its good-enough integrated GPU graphics, and Samsung is building its own mobile GPU. Both entrants have to do something different if they were going to be anything other than late to the party. Just another couple of me-too GPU suppliers.
They have. And you will learn more about their cleverness in due time—probably this year. The TAM for GPUs is surprisingly large and diverse, ranging from smartphones and tablets, game consoles and PCs, workstations and TVs, supercomputers and autonomous vehicles, airplane cockpits and medical instrumentation, to test equipment and robots. I put my mind to it, I could probably come up with some more. Expressed in units, the GPU market something north of 2 billion a year. Over a dozen companies offer GPUs of various sizes, price, and performance, and most of them have been in business for over 15 years, so obviously it’s sustaining them, and there is strong demand.
However, over 60 other companies and their investors learned the hard way that designing, manufacturing, and then selling a GPU (assuming you were even clever enough to design one that worked) is far, far from trivial, in fact it’s damn hard. It takes a team of several hundred, and one of the companies in the illustration above has 3,000 engineers involved in the activity. It takes about three to five years to design and produce a GPU, longer obviously if its your first one. It also takes about a billion dollars. So, anyone not willing or able to make that kind of investment in time, people, and dollars would do well to find something else. The two new entrants exemplify the point. Intel has been designing and shipping hundreds of millions of GPUs for about 15 years. Intel and Samsung are the giants of the semiconductor space and are equipped with something no other GPU supplier has—a state of the art fab.
And neither company is stupid, or foolish with their capital. That two such heavyweights have entered the market tells you there must be one hell of a payoff in order for them to make such a commitment.
But size and resolve don’t guarantee success, as I’m sure their competitors will tell you. It’s still very very early days for both companies, and we won’t know the results of their efforts for two years or more. They will stumble as they learn the intricacies of the market, don’t view those training-wheels mistake as a harbinger of their future—remember, these are not light-weights dependent on VC infusions, these companies have staying payer and long horizons.
We say welcome to the party. You will drive the competition to excel even more. However, none of the existing players are lazy or stupid, they will surprise you with their reactions.
These next few years are going to be the most exciting in GPU land for quite some time. I’m going to be chasing some damn amazing pixels.