The Steady Patience Of AMD In The Datacenter

We have said more than once in recent weeks that recessions accelerate technology transitions but they do not cause them. We now have a textbook experimental case developing between AMD and Intel where it concerns CPUs for servers and where it concerns GPUs and both AMD’s and Intel’s desires to take on the hegemony of Nvidia in GPU compute in the datacenter.

To put it bluntly, there has never been a better time for AMD to try to take on Intel in CPUs and to try to blunt the attack it will bring in GPUs. But the recession, however long-term or short-term it turns out to be, makes it harder in some ways to execute because this is very choppy as well as fast-moving water we are in instead of a roiling ocean at high tide as an economy with competitive forces tends to be. And we are, frankly, surprised that AMD has not been able to take more shipment and revenue share from Intel in general and particularly as companies examined their server spending plans in the first quarter and then forecast out into the future for 2020 and 2021. There is something that we don’t understand that is giving Intel leverage with customers to keep that Data Center Group hitting new highs, as it did in the first quarter of this year and has done repeatedly throughout the second half of 2019, too.

For all we know, the hyperscalers have a lot of code tuned very tightly to the Xeon SP design and the cloud builders worry that their customers, who bring their own applications to the instance slices they sell, have similar issues with their own applications that they might run on the public cloud. We don’t think this is the case, but we don’t know it for sure. We find it hard to believe that server buyers still doubt AMD’s commitment to server CPUs and server GPUs. AMD is in it for the very, very long haul, and has the exascale supercomputing wins to prove it.

None of Intel’s persistent success takes anything away from AMD, which has consistently grown its own datacenter business in the past three years since launching the “Naples” Epyc 7001 processors in June 2017 and following up with the much more impressive “Rome” Epyc 7002 processors in August 2019, which have been tweaked twice to pursue niches since then. AMD has kept the heat on Intel in a way that we have not seen since the “SledgeHammer” Opterons way back in 2003 and 2004. The Rome Epycs can stand toe-to-toe with the “Cascade Lake” Intel Xeon SPs on a per core basis, and when it comes to aggregate throughput, AMD cleans Intel’s clocks because it has more than twice as many cores, a third again as many memory controllers, and 60 percent more PCI-Express I/O lanes that run at twice the speed. And as for bang for the buck, AMD absolutely and clearly wins by a huge margin – at list price, anyway.

All we can figure is this. The 14 nanometer technologies that Intel is using to make the Cascade Lake and Cascade Lake-R refresh chips must be so mature and the yields so high that it can chop the prices by a huge margin, make it up in volume, and still preserve its revenue and most of its operating profit stream. This must also mean that 14 nanometer yields with “Broadwell” Xeon E5/E7 and “Skylake” Xeon SP chips back a few years ago were not as great as many of us had believed for margins to still be holdable at such deep price cuts today. We don’t know how deep those price cuts are, but they have to be pretty damned steep if Intel can compensate for such a huge performance and price/performance differences with AMD Epycs at list price. And if that is the case, then maybe there is an argument for incumbents to hang back and make money when they can using older chip etching processes, which is just the weirdest thought we have ever put in this publication. If this is indeed what is happening, we can assure you that this was not a planned strategy by Intel, but one born of necessity as it has struggled to bring 10 nanometer manufacturing online for years and years. It goes against everything Intel has believed in since its inception.

Clearly Intel has made the absolute best of a very bad situation. It is nothing short of amazing, honestly.

It is unclear in this scenario how deeply AMD can cut prices on Epyc prices to compensate, since the company already lowered its costs by moving to chiplets and came out with aggressive pricing to get the market share that it has attained thus far. It is on much more advanced 7 nanometer processes from Taiwan Semiconductor Manufacturing Corp for the Rome core chiplets, but still using very mature 14 nanometer processes from Globalfoundries for the I/O and memory hub chiplet at the center of the Rome package. That chiplet packaging adds cost, too, but the net cost of a processor socket is lower than it might otherwise be with a monolithic design. Hopefully, there is some scale in that chiplet packaging and it will get easier and cheaper over time for AMD to get big compute in the field for a higher profit margin.

In the quarter ended in March, AMD’s revenues rose by 40.4 percent to $1.79 billion, and net income rose by a factor of more than 10X to $162 million, representing a very healthy 9.1 percent of revenues. The company ended the quarter with $1.39 billion in cash and investments as the March quarter came to a close.

The Compute and Graphics division had sales of $1.44 billion, up 73 percent, and operating income  rocketed up by a factor of 16.4X to $262 million. So clearly the client CPU and GPU business is doing well. Sales of circuits for game consoles have been taking it on the chin, and will have to start ramping soon to meet the holiday shopping season. Depending on the depth of the recession, a new game console may be the one big ticket item for families. . . . It’s hard to say. In the short term, the game console chip business in the Enterprise, Embedded, and Semi-Custom division, where Epyc server CPUs are parked in AMD’s numbers, fell by 21.1 percent to $348 million in sales in Q1 2020, and the division posted an operating loss of $26 million compared to a gain of $68 million a year ago. AMD had a one-time licensing gain of $60 million in that year ago period, so it was not as profitable back then as we thought based on the core chippery and that also means the operating loss to the core business is not as bad as it looks.

Lisa Su, AMD’s chief executive officer, said on a call with Wall Street analysts going over the numbers for the quarter that datacenter sales – meaning Epyc CPUs plus Radeon Instinct GPUs sold into the glass house – drove the “high teens” of the company’s overall sales. So we are going to peg that at 18.5 percent, based on gut and compared to around 15 percent in prior quarters.

Based on thin data, past trends, and hunches, we reckon that the Epyc CPUs drove $252 million in sales in the March quarter, up by 2.2X compared to a year ago and up 9 percent sequentially in a quarter that just plain finished weird for every company in the world. We think that datacenter GPU sales (which includes both Radeon and Radeon Instinct cards made by the Compute and Graphics division) amounted to $78 million in Q1, up a smidgen year-on-year and down a smidgen sequentially from Q4 2019.

In the conference call, Su said that server unit shipments grew by double digits sequentially from Q4 2019 and more than tripled year-on-year, which means AMD is absolutely putting the pedal to the metal on price/performance to take share from Intel.

“We continue gaining momentum across cloud, enterprise, and HPC customers,” explained Su. “We saw particular strength with cloud providers, introducing new instances and accelerating current deployments. Microsoft Azure, Google, and IBM all announced new offerings powered by second-generation Epyc processors, highlighted by Google launching multiple general purpose VMs and Microsoft rolling out an all-AMD virtual desktop offering that also includes Radeon Instinct GPUs. Several cloud providers accelerated their infrastructure deployments to address rising demand from the growing number of users working and schooling from home. For instance, one of our large cloud customers was able to deploy 10,000 second-generation Epyc servers in less than ten days to support the surge in demand for their collaboration services.”

The effect of this heavy sales to hyperscalers and cloud builders is that average selling prices came down, and we think there was some price cutting on top of that even at enterprise customers. The good news is that AMD does not have a very large SMB business for Epyc server chips and none for Radeon Instinct GPU accelerators, so those companies that are most adversely affected by the coronavirus pandemic were not buying these datacenter products, anyway. One other interesting tidbit: Su added that much of the growth for sales of Epyc CPU sales at hyperscalers and cloud builders was for internal use, not for publicly available instances on clouds. In the enterprise, sales have been particularly strong for HPC workloads, according to Su.

Many of us are impatient for an immediate and explosive adoption of Epyc CPUs because that is indicative of intense competition. But this is not Opteron versus Xeon from 2003 through 2008. Intel is not as weak now as it was back then, even though it seems otherwise. And Su is counseling an expectation of steady and sure gains rather than explosive ones.

“The way I look at it, this is very, very similar to the ramp that we saw in the PC business,” Su explained. “The ramp in server is something like steady as she goes. And each quarter, we had platforms. Each quarter, more platforms are qualified. Each quarter, they ramp. It is a little bit different from a pure market phenomenon. And again, I understand that there are market phenomena and then there are growth expectations based on platform launches, as well as software being qualified, and so on and so forth. So as it relates to our expectations, it’s actually going quite well. As it relates to the acceleration of cloud, I think we are pleased with it. We are not ready to upside numbers at this point. I think we already had very aggressive growth assumptions.”

AMD may be patient, but that doesn’t mean the rest of us are. AMD is still on track to launch the “Milan” Epyc 7003 processors, with Zen3 cores, in late 2020, and will bring out new RDNA GPUs – hopefully with Radeon Instinct versions – around the same time, too. We will see what round three of the fight brings soon enough, but we are still in round two for at least two more quarters, and this is a very complex fight with intricate martial arts and weapons and barriers that we cannot always see. This is an epic battle, so be careful to listen for the grasshopper at your feet.

Sign up to our Newsletter

Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.

Subscribe now

7 Comments

  1. Personally, I find it frustrating that equipment vendors like Supermicro are not offering more certified/validated AMD based systems for wider adoption.

    Even today, they announced validated “Red Hat Hyperconverged Infrastructure” all based on Intel CPUs. Further, their “Supermicro Solutions for SAP HANA” solution are all based on Intel CPUs.

    I am not suggesting that Intel system should be ignored, but some AMD based servers would be good.

    • I agree. You could argue that a two-socket Epyc is just as good at HANA as a four socket Xeon SP. This is simply just not something that AMD is focused on at the moment, and there doesn’t appear to be a plan for a four-socket or eight-socket Epyc, from AMD or otherwise, and that means there is a performance ceiling that will make some customers who need to grow HANA nervous. I think there is a credible argument for making bigger Epyc iron, but that will have to come later. AMD needs to pick its battles right now.

      • Well, AMD does have the first VSAN with Dell. It is interesting why they don’t do a 4P beyond the 2P sleds. And Rome is one NUMA node. The 7Fxx series destroys everything and it is being picked up for high perf customers.

  2. The more cores you pack in a socket the less benefit you get from adding more sockets in the same server. It’s better to have 1 (maybe 2) sockets per server but have the software scale across multiple servers in a more flexible way. Having 128 cores using only 2 sockets is better than 112 cores using 4 sockets. Big box brands go Intel because: 1) the cost of hardware is small compared to software+service; 2) no one got fired for choosing Intel; 3) long lead times mean Intel was already chosen; 4) Intel is seen as more sales for such a small market; 5) Intel has deep pockets to (bribe) I mean subsidise manufactures, partners and end-user companies to choose/optimise using Intel. It takes >5 years to establish the performance, reliability and sales credentials before the market shifts on mass. AMD is starting from a low base so it is taking a while. If AMD reaches a market % limit set by Intel internally then Intel will increase the fight back.

  3. Fortunately in all our procurement, we have the luxury of going by price and cost math only, so we suddenly stopped buying Intel! Didn’t have a choice. And you are talking Desktops and servers.

  4. I have long been a supporter of AMD in it’s competition with Intel, particularly in recent years with introduction of EPYC CPU line.

    However I found it bewildering that AMD chose to focus almost exclusively on it’s new chip line initially for Windows PC gaming, when it would have been appropriate and prudent for AMD to also simultaneously offer EPYC based hardware for smaller systems servers, as was mentioned in comment on Supermicro.

    Most of my clients during the past twenty five plus years have been small businesses – byAmerican standards, in parts of the USA but mostly in hospitality industry in the Caribbean and Central America, for which such hypothetical Supermicro or any other server vendor hardware of similar configurations would be more suitable for their Linux based hotel, touring and related services.

    Small business and organizations owners and operators in particular have a much grander concept of Intel based computing compared to that of AMD as generally non-technical citizens, and have commented to me of not being interested in Gamer based computer hardware – as gleaned from public AMD promotion and marketing, no matter how much praised in tech media for it’s superiority of performance and lower costs or Total Return on Investment (ROI), which are normally critical aspects of all business processes considerations in areas of world less affluent and wealthy than USA.

Leave a Reply

Your email address will not be published.


*


This site uses Akismet to reduce spam. Learn how your comment data is processed.