Get email delivery of the Cadence blog featured here
It's 2021 finally. Although 2020 was actually a good year for the semiconductor industry, and for Cadence in particular, for many it was not (such as my daughter, son-in-law, and daughter-in-law, who all work in some way in food and beverage). So let's hope 2021 is an improvement.
Every year, the UN decides the year is the "International Year of X". I bet you didn't know that 2021 is the International Year of Fruits and Vegetables. Not as interesting as a couple of years ago when we had the International Year of the Periodic Table. It is a more fun picture than just a calendar, but I don't think I can "squeeze" a blog post out of it.
It's time for some predictions about 2021. Last year, I wrote a piece The Five Waves: AI, 5G, Cars, Clouds, IoT, so let's use those waves to make some predictions, before adding a couple more.
They are not entirely independent. For example, automotive uses artificial intelligence (AI) for ADAS and autonomous driving. The mobile industry is always saying that autonomous driving will need 5G, but I never hear automotive people say that. For driving, you can't depend on 5G being 100% reliable and always on, and for the parts that need cellular access, such as updating maps or over-the-air software updates, 4G or LTE is fine. But let's go in an order that builds up sensibly.
First, there are going to be a lot more hyperscale data centers. The trend to move more and more into the cloud, despite the occasional company going in the opposite direction, will continue as existing data centers require refurbishment. For example, just before Christmas, Twitter announced that it is going to abandon its own data centers and move to Amazon AWS.
I've already said that I think that this year will be the year of Arm in data centers. AWS's Graviton 2 is already out there. I expect to see more chips from the likes of Microsoft, who have already have at least one Arm data center built for internal use. Now that Amazon and Apple have demonstrated it is possible to build a state-of-the-art processor with better single-thread performance than the incumbents, I would expect that anyone with both its own data centers and chip design capability is at least thinking about this. It is also unclear whether NVIDIA's pending acquisition of Arm is going to change anything.
Netflix has said that much of its workload is running on Graviton 2 instances. It was ahead of the curve and gave up its own data centers a long time ago. Perhaps as significantly, when Twitter announced that they would migrate to AWS, it said it would run on Arm instances.
For EDA, I expect to see more movement to the cloud. Anyone starting an SoC company today would probably start with either pure cloud, or perhaps cloud for those big, peaky workloads like physical design and signoff. Larger companies will need to go to their CFOs for the capital budget to completely rebuild their existing obsolete data centers, and since many CFOs have already moved all their financial systems to the cloud, they have a lot more experience and knowledge than the last time that conversation came up three or four years ago.
I think 5G will continue to be over-hyped, although increasing numbers of handsets will be 5G ready. The iPhone 12 supports not just the low and mid-bands, but in the US (only) it also supports mmWave. The rest of the world has more spectrum in those bands, so mmWave is something for future capacity growth, not a requirement for initial buildout. Since mmWave doesn't go through walls, windows, trees, or hands, it is of limited utility. In some ways, WiF-i 6 may be more important since a huge amount of internet access and even phone calls are made using Wi-Fi, especially since many of us leave our homes a lot less than we used to. Comparatively little access is made outside—when did you last take a phone call walking down the street? Sending text messages at a ballgame maybe...but we'll have to see when we get ballgames again. Uploading IG pictures of your food at restaurants...something else, that at least around here, is not happing much, and restaurants have Wi-Fi most of the time.
The Consumer Technology Association numbers predict that even so, less than half of phones sold in 2021 will be 5G (in the US market). That was from CES last year and it might have been updated since. My prediction would be that most phones sold will be 5G but coverage will be so spotty and congested that it will not be available in most places. And by the way, if you have AT&T, it still insists on calling LTE-A 5G and putting that on your screen. But that's actually against the 3GPP rules, which define the Gs in detail.
Obviously, I'm not going out on a limb to say that the importance of AI will increase. Just in the last few months of 2020, we had the results about protein folding (see my post Google's DeepMind's AlphaFold Solves Protein Folding). Also, the release of GPT-3 (Generalized Pre-Trained Transformer 3) with its 175B training weights. As Wikipedia says "The quality of the text generated by GPT-3 is so high that it is difficult to distinguish from that written by a human."
For the semiconductor industry (apart from NVIDIA, of course), most of the interest is in inference at the edge, which can be done on an SoC. The recent Linley microprocessor conference had many presentations on just this, all of them claiming to be "the best" in some way. It is unclear to me who is going to buy all these chips, since the high-end smartphone manufacturers have all designed their own deep learning inference engines. It reminds me of a few years ago when people thought there was a gold-rush for mobile processors. Qualcomm and Mediatek captured the low end, but all the high-end smartphone manufacturers have designed their own application processors (and they all contain Arm processors under the hood) so there turned out not to be as much of a market as expected.
The second part of my predictions, and some more information about 2021, tomorrow.
Sign up for Sunday Brunch, the weekly Breakfast Bytes email.