Markets work when consumers of a widget don’t feel there is significant differentiated value in making their own and would rather get lowest possible cost from experienced widget makers who can amortize their investment over high-volume sales to many customers. But that changes when a large consumer finds they can increase differentiated value if they build their own.
A case in point is a rather large speed bump for what had seemed like boundless potential for GPU growth around deep learning/neural net applications. That speed bump is a Google announcement of their Tensor Processing Unit (TPU) which is a dedicated accelerator for learning. This doesn’t mean the market goes away for GPUs in general or GPUs applied to learning. But when one of the largest companies in the world, who are deploying this function in their datacenters/cloud, announces they have this function covered themselves – that’s a speed bump for GPU makers.
TPUs are designed to accelerate learning in TensorFlow in applications like StreetView, RankBrain and voice recognition for translating voicemails to text, among other applications. Because the design team understands the applications they apparently have been able to reduce precision, allowing more operations to be packed into a given area, leading to greater performance. No magic – they just have done what we all would do if we didn’t have to design a general-purpose solution.
Google also said that they have no intention of selling TPUs to anyone else. Why would they – this gives them a technical advantage. And what used to be a barrier for smaller ventures building their own chips – cost, building expertise and so on – is pocket-change for Google.
I suspect this is another step in a trend, not an isolated case. As Moore’s law slows, accelerators will rise again in popularity; as new needs appear in big data, machine learning, fintech, it’s not unreasonable to expect more in-house development. This won’t just be for Google. Facebook, Amazon, Netflix and other 800 pound gorillas will be looking for every advantage they can find. General market hardware will still have a place but not around the bleeding edge, especially when much of the software in these domains is now open-source.
Again, I’m not saying the merchant semiconductor market goes away. They can still sell to PC makers, automakers, some (though maybe not all?) IoT device makers and other existing consumers; perhaps also to the gorillas in some applications. But when new large dollar volume opportunities go internal in some of the largest tech companies, that’s an important shift.
You could look at this as a bad thing or a good thing. Maybe not so good if you want to stay with the (old) status quo in semiconductor design. But actually pretty good if you’re a design engineer looking for leading-edge opportunities. The jobs aren’t going away – they’re just shifting to different companies.
You can read more HERE.