Heterogeneous Compute

The market for processors is huge. Depending on how you count them, the CPU market was over $80 billion last year and the market for GPUs was about $20 billion (according to Grand View Research). Those figures have been growing strongly for years, and we see no reason why that will change any time soon. That being said, we think the growth curve for this market is going to change as new types of chips and new entrants come into the landscape.

Today, there are three main markets for processors – phones, PCs and data center servers – and then somewhere over the horizon is the market for processors in cars. Each of these markets is in its own state of turmoil. This turmoil is caused in part by a changing supplier base, especially as large customers move towards building their own chips, but the churn is also driven by the change in workloads, which can really be summed up by just calling it AI.

Processors for phones are called Applications Processors, and we covered those a few weeks ago. Here the problem is that all the major handset makers are building their own chips. Driven by the need to keep pace with Apple, these companies are all exploring their own chips, while incumbent vendors like Mediatek and Qualcomm are racing to beef up their offerings. Again, a large part of this change is the growing need to add AI to these chips. In this context, AI really means better image and video processing, but as this is the core competitive domain for phones, it is a pressing problem. At this stage, the field is wide open. It is unclear who will win in the race between internal solutions and the merchant vendors.

For PCs, the change is a bit simpler. Apple has launched its own CPU, the M1, and comparisons with other chips leave the leading suppliers like Intel and AMD looking exposed. We think it is unlikely at this stage that the other PC vendors will take this path, it is just too far from their core competency and hard to do without greater cooperation form an apparently indifferent Microsoft. We also tend to think that alternatives to traditional PC, Chromebooks for example, may be able to claim share gains, and for these ‘net-top’ devices, there is a large selection of processor vendors, many of whom are based in China. Short of a major shift in that direction, the PC CPU market is unlikely to change much for the time being.

By contrast, the data center is set to see major flux. Here we see both forces in effect. The hyperscale data center owners (i.e. the Super Seven Internet companies) are designing their own chips, and there are new entrants to the market. These are coming in the form of Arm-based CPUs. Perhaps an even stronger force is the shift of workloads to “AI”. Google really kicked all of this off five years ago when they launched their first “TPU”, a chip purpose-built to do the linear algebra that we all call “AI”. At the time, Google said the TPU would halve the number of data centers they would need to build to handle the anticipated task of voice-based search. Since then the demand for AI has grown significantly.

This shift cold lead to some serious changes in the data center. First and foremost, as AI workloads shift towards dedicated AI ASICs (purpose-built chips as opposed to general purpose processors) the demand for traditional processors will shrink – that is what Google’s ‘halving’ means. We have not seen signs of that yet. In part because data center demand is growing so quicklyin absolute, compounded by all the time we are spending at home. But another shoe has yet to drop. Today, most AI calculations, especially at everyone other than Google, are still run on traditional processors, especially GPUs. AI ASICs are still an emerging product class, with all the associated growing pains (some of which we discuss here and here). That being said, we are rapidly approaching the tipping point when this will change. The leading chip vendors, especially Intel, have been bulking up their AI offerings through acquisition. Nvidia, who is the most threatened by this through their exposure to GPUs, is busy segmenting its product line and entrenching its distribution advantages. But there are dozens of promising start-ups out there with good (or good enough) alternatives.

Finally, there is the next battleground – the car. As it stands today, this market is wide open. There are some early leaders, but the market itself is yet to come into full focus. It is clear that autonomous vehicles will need some very heavy processing power, but autonomous vehicles are still many years away. Far enough, that no start-up could hope to raise money to build a chip for the market. Far enough, that none of the incumbents can claim any certainty about their future prospects. And even before we get to full autonomy there are likely to be many big jumps in semis content for cars, but whether these will be gated by various stages of semi-autonomy or other requirements is also unclear. We plan to dedicate more time to this market, but for our purposes here suffice it to say that this market offers immense potential and uncertainty.

For many years, the market for all forms of processors was fairly static in terms of competitive dynamics, but increasingly the changing needs of these systems and the entrance of many new vendors/customers means that these markets are set to change considerably in coming years.

And we haven’t even touched on the subject of China….

One response to “Heterogeneous Compute

  1. Pingback: Gaming Internet – Magic Packets | Digits to Dollars·

Leave a Reply