Only two quarters ago, AMD’s datacenter business – meaning sales of Epyc CPUs plus Instinct GPU accelerators – broke through $1 billion. And it was a big deal.
If current trends persist, this business will break through $1.5 billion in the first quarter of 2022 and $2 billion in either the third quarter or the fourth quarter of this year. With demand from hyperscalers and cloud builders on the rise for Epyc CPUs, some major exascale-class HPC systems using a mix of Epyc and Instinct compute engines, and enterprises starting to turn to AMD instead of an Intel still struggling to unfold its Xeon SP roadmap, there is every reason to believe that this year will be the best one ever that AMD has had in the datacenter. Just like last year was its best ever. And we could be saying each year is AMD’s best ever in the datacenter for at least a couple of years.
“We have set out a roadmap for, frankly, not just 2022, but beyond, which allows very aggressive growth goals,” Lisa Su, AMD’s chief executive officer, explained in a conference call going over the fourth quarter 2021 financial results for the chip supplier. “We work on a regular basis with our customers and our supply chain partners. I would say we have better visibility than we have ever had from a customer demand standpoint, and so that gives us pretty good confidence in terms of what is needed, but there are always going to be some puts and takes. And so we have enough flexibility to do that. But our goal is to dimension for success. At the end of the day, that’s what we want to do is we want to satisfy customer demand.”
Ironically, given the tightness of the semiconductor foundry capacity and the supply chain for chip making components and packaging and testing, the hyperscalers and cloud builders who were the first buyers of Epyc processors when the line was launched in 2017, have to tell AMD what they need. They can’t be coy and only tell AMD what Epycs or now Instinct compute engines they might want off the AMD roadmap because there is no spare capacity anywhere in any foundries and they might not get them when they need them otherwise. Now, AMD is very much part of the capacity planning thanks to the pandemic-induced semiconductor shortages, and has “visibility now multiple quarters and, in some cases, multiple years out,” as Su put it.
This explains why Su and Team have the confidence to put together expanded wafer supply and chip etching contracts with Taiwan Semiconductor Manufacturing Co, as they did as 2021 was coming to a close. They already know what the customers who consume the majority of their datacenter chips are going to want, and when they are going to want them.
AMD has not experienced this since the peak of the Opteron server era nearly two decades ago.
In the quarter ended in December, AMD’s Compute and Graphics group, which sells PC chips and all GPUs, posted sales of $2.58 billion, up 31.8 percent, with an operating income of $566 million, up 34.8 percent.
The Enterprise, Embedded, and Semi-Custom group, which makes Epyc server chips, game console and other custom chips, as well as embedded versions of Ryzen desktop and Epyc server chips, posted $2.24 billion in sales, up 74.6 percent, with operating income of $762 million, up by a factor of 3.1X over Q4 2020.
Take out other corporate investments and expenses and pay the taxes, and AMD overall had $4.83 billion in sales, up 48.8 percent and had a net income of $974 million, off 45.3 percent but against a Q4 2020 when it booked a $1.3 billion tax benefit. Take that out, and net income more than doubled – just like graphics revenues did, overall datacenter CPU sales did, datacenter GPUs in the Instinct line did, sales to hyperscalers and cloud builders did, sales to enterprises did. It looks like game consoles did, too. So many parts of AMD’s business doubled it is hard to imagine why the company didn’t double its sales, too.
On the call with Wall Street, Su said that the datacenter business, which is not broken out specifically, contributed a “mid-20 percentage of overall revenue” and that AMD expected that datacenter percentage of overall sales to rise as 2022 progressed.
Here’s how the AMD datacenter business stacked up against its two groups:
We estimate that datacenter sales comprised 26 percent of overall revenues, or $1.26 billion in Q4 2021. Of this, Epyc CPU sales came to $1.11 billion, up 103.4 percent year on year, and Instinct GPUs accounted for $148 million in revenues, up 105 percent year on year. We took a wild estimate and figure that in Q3 and Q4 of 2021, the custom “Trento” Epyc 7003 CPUs and “Aldebaran” Instinct MI250X GPU accelerators used in the 1.5 exaflops “Frontier” supercomputer pumped about $223 million into AMD’s coffers, and that this is a little less than half of list price for these motors as best as we can estimate. Those supercomputer deals sure did help the top line, and we wonder if the hyperscalers and cloud builders get similar discounts. (Probably not, especially given the wonkiness and pricing in the CPU and GPU markets right now.)
While AMD says that all of its businesses are expected to grow in 2022, overall revenues are only expected to grow by 31 percent for the full year. AMD had $16.43 billion in sales in 2021, and with 31 percent growth, it would hit $21.5 billion. Now, some fun with math. If the datacenter business can grow at an average of 20 percent sequentially, as it did in 2021, then we can map out future datacenter sales pretty easily and then calculate the percentage of overall sales for each quarter to add up to a total revenue of $21.5 billion for all of 2022.
When you do that, here is what the revenue model we came up looks like:
There are many ways to fit curves to what Su said on the call, but this is a nice linear model that might reflect what customers will do, given shipment times for current and future compute engines and the competitive pressure.
And there will be competitive pressure from Intel on both the CPU front with “Sapphire Rapids” Xeon SPs and “Ponte Vecchio” Xe HPC GPU accelerators. As Intel’s Q4 2021 demonstrated, if you can make any compute engine in volume, even if the competition can kick the tar out of it, you can sell all you can make because companies have to buy compute engines. And demand is keeping prices high, even for relatively uncompetitive parts, as has been the case throughout the coronavirus pandemic. AMD is going to be able to hold its own against Intel with 96-core “Genoa” Epyc 7004 CPUs and Instinct MI200 GPU accelerators – and then some. But Su is not complacent about it, and is diversifying the compute engine lines to chase more opportunities.
“We always expect the competitive environment to be very strong and very aggressive,” Su explained. “And that’s the way we plan our business. That being the case, I think we’re very happy with the growth that we’ve seen in the business sort of last year. As we look forward, we see opportunities in both cloud and enterprise. On the cloud side, we are in ten of the largest hyperscalers in the world. As they get familiar with us over multiple generations, they are expanding the workloads that they are using AMD on. So we see that across internal and external workloads. In the enterprise segment, we doubled year-over-year here in 2021. We continue to add more field support to have more people get familiar with our architecture. We have very strong OEM relationships. So I feel very good about our server trajectory. And yes, it’s very competitive out there. But we think the datacenter business is a secular growth business. And within that, we can grow significantly faster than the market.”
At best, the server market will grow revenues by maybe a few points, so AMD is growing amazingly faster than the market at large. And by the way, at least some of that datacenter revenue is going to be driven by AMD passing through higher semiconductor manufacturing costs to customers and opportunistic pricing as demand chases too little supply.
Thank you for the very knowledgeable article.
“AMD is going to be able to hold its own…” against Intel, with 96-core Genoa? That is a very conservative estimate, considering you won’t see SPR until after Genoa is released.
AMD “Datacenter” = Instinct + Epyc = “INSTEPYC”
It’s not just feeds and speeds that matter, but availability. If AMD only makes X CPUs and Y GPUs for compute, that’s all it has got. Intel gets what is leftover. As its own Q4 shows. The server market had a killer quarter, and that is why Intel did. I strongly suspect some channel stuffing by Intel, but I can’t prove it.
A facet of this transparency is risk/funding, as we are seeing with TSMC.
~”If you want us to go out on an investment limb to assure ur future needs, we need some prepaid orders.”
TSMC have collected many such billions to fund new fabs – even from Intel. Sweet