AMD is committed to gaining a significant share of the market for data centers

There is an equally positive and negative trend driving all computing: innovation requires competition in order to set them in motion, and competition requires innovation in order to fit it.

This is best illustrated by Intel’s growth in the data center market over the past three decades, as well as the ups, downs and revivals of rival AMD in the same market. However, this is not enough for AMD. The company wants to get even a larger “piece of the pie” of data centers than in the era of Opteron processors, and AMD management outlined a plan for this.

We will not go into all the delicate technical details about AMD technologies, but right away we will start by calculating the income that AMD expects to receive in the data center market until 2023. Then we’ll talk about what the company's engineers said about its roadmaps for CPU and GPU, about the competitive positioning that they will have, and on which the chief executive officer Lisa Su relies, believing in achieving her goals.

First of all - people, development and profit - then

Needless to say, Lisa Su is the best thing that has happened to AMD since Jerry Sanders, CEO of the company from 1969 to 2002, founded the chip company with a bunch of engineers from Fairchild Semiconductor, which also spawned Intel among many others. the so-called Fairchildren from Silicon Valley . Su, who led AMD in rather poor condition in 2014, has assembled a very good team that can change the company for the better.

Of all the charts that AMD showed during Wall Street presentations, the most important was the one presented by Devinder Kumar, the company's CFO. Take a look:

This chart is important because it shows the impact of all the design and sales efforts undertaken since 2015, when AMD sold $ 3.99 billion in sales, mainly from low-end chips until 2019. AMD completed the year 2019 with sales of $ 6.73 billion, of which just over 1 billion was generated for CPUs and GPUs sold for data centers. The rest of the company's sales came from CPU and GPUs for manufacturers of PCs and game consoles. The second reason for the importance of this schedule is the demonstration of the long-term growth of the company as a whole until 2023, as well as plans to more than double the share of its revenue from the data center market as part of the rapidly growing total sales.

There are various ways to switch from actual sales in 2019 to expected sales in 2023 for AMD, and neither Devinder nor Su did not specify what exactly they intend to do to achieve their goal. But we have some ideas on this subject.

When we analyzed AMD’s fourth-quarter and full-year 2019 revenues back in January, we decided that based on Intel’s dynamics and difficulties in moving beyond its 14-nanometer Xeon SP processors and the absence of discrete graphics processors for data centers, AMD will be able to increase more than doubled their income from data centers in 2020. New game consoles from Microsoft and Sony will further increase the share of AMD. If we add our 2020 model to what happened in 2019, then in the end we will get about 20 percent average annual growth rate for the entire business and more than 30 percent market share in 2023 for data centers and GPUs. Here's what our model looks like:

For those who prefer visualization - the same data on the chart:

When AMD announced the share of sales of CPU and GPU for data centers larger than 30 percent, we were pessimistic about this. However, a change in the percentage is quite possible. Based on AMD data, its total revenue from exiting 2023 will be approximately $ 14 billion. As already mentioned, we expect that the growth rate will be higher in 2020 and then will decrease over the years, which is quite natural given the activity of Intel.

Now let's look at Intel and Nvidia, which do not stand still and have a dominant share in sales of CPU and GPU for data centers both at the beginning of 2020 and during the last decade. Intel accounts for about 99 percent of server chip shipments and about 90 percent of server revenue; the share of revenues from server CPUs is lower because there are more expensive processors, but the high price that IBM can dictate for its Z and Power motors is actually offset by the increase in sales of cheap AMD X86 server chips and the relative share of Arm server chips sold Marvell, Ampere Computing, Huawei Technology and others.

In the flat market, based on the current difference in TCO, what AMD shows with the "Rome" Epyc 7002 series compared evenwith Intel's freshly updated “Cascade Lake R” Xeon SPs , each $ 1.00 sale of Rome chips reduces the profitability of Intel Xeon SP by $ 1.40 to $ 3.40, based on the price / performance advantage that Forest Norrod, General Manager The manager of Datacenter and Embedded Systems Group at AMD, announced in his presentation below:

The graph compares the 16-core chips that Norrod described as the “heart of the market” and if we average the indicators, we can talk about approximately $ 2.25 lost for Intel for every $ 1.00 AMD earned.

In many cases, this savings will be redistributed as additional capacity, which will further reduce Intel's revenues, and in other cases - especially if the economy cools down due to the outbreak of coronavirus - it simply will not be spent at all. In any case, Intel expects a blow to revenue. Intel is still saving the expansion of the server market, with some spikes, but, as we noted earlier, For the entire 2019, Intel's revenue in the Data Center Group grew by only 2.1% to $ 23.48 billion, while operating profit actually fell by 10.9% to $ 10.23 billion. Sales of server processors in these indicators exceed sales of motherboards, systems, switches and other products. AMD, of course, has an impact on bottom-line Intel and if the cost of servers is reduced, it will also begin to harm its top-line and will be strengthened by increasing the cost of launching its 10 and 7 nanometer CPUs and GPUs. While the market is expanding, everyone can conduct their own business, but it will be very unpleasant if the market tightens - Intel will lose revenue and profits in a price war.

To a certain extent, the same applies to the position of Nvidia in GPU computing, where both AMD and Intel will try to get their share. Nvidia can only lower the price by losing revenue and profits or keep prices high, but still lose revenue and profits. The only thing that has so far saved Nvidia from such an influence is the huge software ecosystem created over a dozen years. Hundreds of applications and a myriad of native code have been ported to Nvidia Tesla GPU accelerators. However, it is worth noting that those who can control their own software stack, such as the Oak Ridge and Lawrence Livermore laboratories of the US Department of Energy, do not really care about this, because they have their own independent stacks and their own applications. If they work together,in order to make the Radeon Compute Environment, or ROCm, a good enough environment, and hyperscalers and cloud builders who need GPU acceleration will help, AMD can build the ecosystem in much less time than Nvidia needed. With the help of Intel and AMD, they can turn the narrow Nvidia trail, which has become a dirt road, into a two-lane highway, turning into a four-lane, with bridges and tunnels, wider and more convenient for everyone.

AMD expects the Radeon Instinct GPU-based computing business to grow much faster than the Epyc CPU-based business, and that’s just because it’s much smaller. We now believe that two-thirds of AMD’s revenue from data centers is generated by Epycs, and one-third from Radeon Instincts. By 2023, revenues from GPUs can account for a much larger portion of revenues in the AMD data center market. If AMD can offer an advantage in the cost of the GPU compared to Nvidia (and the performance and price of the Frontier and El Capitan systems suggests that this will be the case, at least as part of these big deals, because AMD still got some of these GPU contracts ), then Nvidia will definitely feel financial pressure.

Much depends on the prevalence of the GPU. Now, according to our estimates, only 1% to 2% of servers in the world have GPU acceleration. While many have eight or even sixteen GPUs for every two processors, the ratio is 1: 4 or 1: 8. It will take many large hybrid machines to push GPU sales to something even close to CPU sales in the server market. Based on some assumptions with a very wide margin of error, the sales of GPUs for servers in 2023 may be equal to the sales of CPUs for servers, provided that the calculations soar to the load level of HPC and AI. Both will be more widely distributed in the clouds and in enterprises.

What we are talking about (and what AMD representatives did not say) is Radeon Instinct’s business growth of $ 2 billion by 2023, which is about 6 times more than at present, and probably with a pretty decent margin . To achieve this result, AMD will need to adhere to an aggressive pricing policy for Radeon Instinct and win in market volumes.

This seems quite real, given the TAM that AMD sees before itself in 2023:

Achieving approximately $ 2 billion in Radeon Instinct sales is not an exaggeration, as AMD’s market share will reach $ 11 billion by 2023, which is exactly 18 percent of revenue. TAM server CPUs will receive a volume of about 19 billion dollars, and telecommunication / infrastructure space will be another 5 billion dollars and will include both CPU and GPU components. AMD’s share of server CPU revenue is only 2.3 percent. That sounds absolutely achievable - especially if AMD keeps prices low and performance remains high.

Money matters

A little more math. If you look into the future until 2023 and assume that the TAM for data centers is exactly what AMD sees it - AMD's share in this TAM, with at least $ 4.33 billion in expected sales by mathematical calculations in the table above, is 12.4%. Everyone knows that during the time of the notorious Opteron, AMD was able to get a 20% share of server processor shipments, and in some cases reached a peak of 25%. While computing TAM for data centers is growing at 5 percent a year, closer to 2026 AMD will have about 20 percent of the $ 40 billion market, and by 2027 it will have about 25 percent of the $ 42.5 billion market. So, in fact, the company is only one tenth of its way.

2019 was a damn good start. And if a recession is coming (as it was in late 2007, when Intel restarted Xeon on the 64-bit Nehalem platform by 2009), we suspect that this time AMD will win, not Intel. AMD will have a larger market share of server CPUs and GPUs.