[Editor's note (10/13/13): The article was erroneously published using quarterly revenues instead of yearly. The numbers have now been corrected.]
Last week, Verizon (VZ) and Advanced Micro Devices (AMD) announced their collaboration of building cloud servers for Verizon that offer a unique set of features and introduce potential cost savings for Verizon.
This article will be a deep dive into server revenues, a discussion of AMD's SeaMicro division, and will conclude with a speculative discussion of a possible upcoming, unannounced design win.
Looking Back To Think Forward
Image Courtesy of Google:
Microservers, along with traditional servers, have potential to become a more meaningful source of revenue for AMD as cloud infrastructure is developed in order to support the explosive growth of smart devices and the need for increased compute capacity. If history tells us anything, it's that we underestimate the need for compute capacity. In 1943, the President of IBM predicted 5 computers ought to be enough for the world. In 1977, the founder of DEC stated there isn't really a need for computers in homes (Source: TechHive).
As we explore new ways to use microchips, we will have to also build out the infrastructure to store and process the data they transmit. As devices become more mobile, the lack of processing power in those devices is supplemented by offloading some of the processing requirements to the cloud.
As for an illustrative example, think of Facebook. Users upload media files, make posts, and comment on pictures. Users rarely delete the data that is stored on Facebook, meaning the amount of data processed and stored by Facebook grows every time a comment is made or a picture is uploaded, and as more members join, the rate of growth accelerates.
Quantifying Server Revenues
According to iSuppli, microserver growth is on track to grow from .2% of the overall market in 2011 to 10% of the market, constituting 1.2M shipments, by 2016. Given current server revenues of around $50B annually, and accounting for growth in server revenues between now and 2016, 10% of the overall server market would include somewhere between $4B to $6B of overall server revenue by 2016.
According to IDC, in Q2, total server shipments totaled ~$11.9B in revenue.
Looking specifically at the current x86 server market share, IDC lists total revenue for servers at $8.7B with 1.9M shipments. I typically see values of Intel's (INTC) x86 server market share between 90% to 95%, leaving AMD with ~5%. AMD operates with lower margins and typically lower ASPs when compared against Intel. So given a 5% market share and lower ASPs than Intel, I think a good assumption for AMD's server revenue is somewhere in the neighborhood of +/- $250M.
This $250M figure seems reasonable given an overall Computing Solutions revenue of $841M during Q2 and a PC market share of ~15% of PC vendor shipments. However there is one caveat to this number that will be discussed later.
AMD's SeaMicro Acquisition
The acquisition of SeaMicro gave AMD an immediate boost in its ability to better provide data center solutions. The caveat of this purchase, however, was that SeaMicro servers actually used Intel chips. So of the $250M I estimated above in server revenues, part of that money actually comes from Intel chips sales, I believe.
However, in an article on VentureBeat, AMD's Mr. Andrew Feldman pointed out that SeaMicro will now be offering the SM15000 server with Opteron solutions as well. VentureBeat described this as a "blow" to Intel, signifying that AMD's SeaMicro division will no longer be relying solely on a competitor's silicon to manufacture servers, leading to better margins for AMD as it is now able to place its own silicon inside these servers. Note that AMD will still be offering Intel based SM15000 series servers -- AMD based options will be in addition to the Intel-based ones.
One specific advantage of SeaMicro is that the acquisition gave AMD the ability to sell directly to manufacturers. When you look at the Vendor revenues, you'll notice that Intel is not on that list, and neither is AMD or SeaMicro. But, SeaMicro does give AMD an outlet to supply directly to corporations, bypassing the middleman, and passing savings to customers. An interesting article regarding the Verizon/AMD server deal, posted on Wired, details the shift of big data users -- think Google or Facebook buying straight from Asian suppliers to procure cheaper, custom-designed servers. The SeaMicro acquisition gives AMD this same ability, as evidenced by the press releases detailing the collaboration between AMD and Verizon. And according to Mr. Feldman, Verizon is SeaMicro's largest win ever.
AMD's Positioning For Explosive Growth
There are two tidbits of information that I would like to point out. The first is a quote taken from the IDC server article linked to prior:
Density Optimized servers achieved the highest growth of any segment in the server market. The datacenter build-outs by service providers are driving growth in the industry and represent a strategic opportunity for OEMs, while at the same time IDC is seeing new participants enter the market targeting the hyperscale datacenter segment.
The second comes again from IDC data, but is regarding cheaper HPC systems:
Worldwide factory revenue for the high performance computing (HPC) technical server market jumped 7.9% to nearly $2.6 billion in the second quarter of 2013 (2Q13), up from $2.4 billion in the same period of 2012, according to the newly released International Data Corporation (IDC) Worldwide High-Performance Technical Server QView.
HPC systems are sometimes better serviced by processors specifically designed to handle work loads that are slightly different than traditional workloads. The term FLOPs (floating point operations) is used to measure the power of many of these systems. By implementing co-processors to supplement the traditional CPUs, manufacturers can boost the number of FLOPs of compute power. Although AMD typically lags Intel in traditional workloads, the performance delta in floating point workloads is not nearly as drastic.
Furthermore, AMD will be releasing the "Berlin" CPUs next year based on AMD's upcoming "Steamroller" architecture. Steamroller should deliver decent performance gains when compared against AMD's "Bulldozer" architecture. In my opinion, Bulldozer was a major misstep for AMD. When Bulldozer was first released, AMD's new architecture had a hard time competing against the architecture it was designed to replace, causing AMD to take a step back around the time Intel released Sandy Bridge. If Steamroller can deliver the expected gains over Bulldozer based cores, this will be a big win for AMD. While I do not feel it is likely, if Steamroller only provides a modest performance boost over Bulldozer/Piledriver, this would not bode well for AMD.
According to the IDC,
"The non-x86 server market is in the middle of a major transformation, as hardware platforms associated with Unix continue to look for a bottom in their secular decline, mainframes find new 3rd platform workloads, and the market prepares for new RISC-based system-on-a-chip architectures to enter," said Kuba Stolarski, Research Manager, Enterprise Servers at IDC. "As the lines between x86 and non-x86 workloads continue to blur, customers will find compelling cases outside of x86 for hyperscale and scale-up solutions for their analytics, cloud, mobile, and social platforms."
In short, Steamroller should fix, at least partially, the misstep in Bulldozer, AMD's smaller x86 cores are competitive, and shifting workloads in servers will allow AMD to better leverage the fabric IP acquired with the purchase of SeaMicro. Evolving workloads in the server space, along with HPC applications, will also allow AMD to leverage graphics IP. Take Google Glass for example. Graphics IP is much better suited to parse a visual data feed streaming from the hardware to the cloud for processing. AMD has an extremely competitive IP portfolio to analyze this type of data. AMD will offer both ARM and x86 chips to suit the changing workloads that IDC mentions. Lastly, the SeaMicro acquisition gives AMD the ability to work with customers directly in order to create unique servers designed to handle a company's specific workloads.
Making Cents of Performance and Positioning
Regarding microservers, assuming AMD is able to obtain a similar marketshare as the company carries in traditional servers, this represents 5% of my estimation of a $5B market by 2016, based on data provided by iSuppli. However, the SeaMicro acquisition, along with AMD's relatively more competitive smaller cores and introduction of ARM architecture, should enable AMD to capture a higher marketshare in microservers. Five percent of the TAM for microservers represents $250M a year in incremental revenue by 2016; 10% represents $500M.
The growing HPC market, based on the quotes above, also represents a opportunity for AMD. These types of applications will allow AMD to utilize both CPUs and GPUs to create high performance systems designed for specific tasks. IDC pegs the HPC market at ~$11B a year. AMD's SeaMicro acquisition, graphics IP enabling more efficient processing of parallel workloads, and an updated Steamroller architecture should allow AMD to capture more marketshare in this specialized segment. However, there is not enough data resolution to make an accurate prediction at this point. The most useful indicator to keep an eye on I feel will be an updated Top500.org list. AMD has lost market share in recent times on this list, despite during the earnings call reporting record revenues in workstation and server GPUs. A significant increase in the percentage of computers using AMD hardware could signify an increase in the strength of AMD's IP.
Another wildcard is cloud-based gaming. Both AMD and Nvidia have launched GPUs aimed at this new, growing segment. The most interesting case of this currently, in my opinion, are the servers being designed by Microsoft (MSFT) to support the Xbox One. Given AMD's involvement with Microsoft and the launch of the Xbox One in November, these servers would have to be built now if Microsoft plans on having them available at launch. Given AMD powers the Xbox One, I do not think it is out of the question to think AMD powers the servers as well. XBLAFans pointed out the following tweet from Microsoft's Mr. Phil Spencer:
If Microsoft goes with AMD to power these servers, this could be a lasting source of revenue as Microsoft could continually build these servers out to ensure the XBox One is supplemented by cloud compute, providing more powerful and unique game experiences.
Lastly, the bigger, more exciting market I feel is the mundane enterprise server market. A ~5% market share increase in x86 servers would yield a yearly revenue increase of ~$1.5B/year.
According to AMD's most recent earnings call, a positive surprise in AMD's Computing Solutions group was partially driven by increased demand for the company's Opteron 6300 series chip. These are likely the same class of chips places in SeaMicro's SM15000 servers. AMD having a venue in SeaMicro to be able to more directly supply customers, as well as an updated big core architecture, can help fuel growth above a 5% market share. A quote from InformationWeek serves to both give testament to SeaMicro's IP, and offers a potential hint at the Verizon deal serving to impact Q3 earnings:
"The technology we developed turns the cloud paradigm upside down by creating a service that an enterprise can configure and control as if the equipment were in its own data center," Feldman said. In an interview last week, John Considine, CTO of the Verizon Terremark cloud unit, said it was the more-dense hardware that will allow Verizon to launch its Cloud Compute and Cloud Storage services in the fourth quarter.
However, due to the build time associated with creating the Verizon servers, it's entirely possible that the revenue increase was being baked into Q2, and the Verizon deal accounted for the increased demand in the Opteron 6300 series mentioned earlier. If the Opteron 6300 series chips were not used, another possibility is Verizon opted for the lower power 4300 series Opterons. If this is the case, take note if AMD mentions increased revenues partially driven by the company's Opteron 4300 series.
AMD's Involvement With Azure And Microsoft
This section is highly speculative.
Recently, an announcement was made by Microsoft regarding the company's Enterprise Cloud service.
Finding information specifically regarding the mix of silicon used in various applications is spotty at best. Information shows that both AMD and Intel silicon are used in the Azure service. Intel's chips seem to be used quite a bit for HPC applications. The chips used to build the HPC servers were Sandy Bridge based Xeons, and the collective of these chips earned Microsoft's Azure HPC service a spot on the Top500.org list.
But it appears the AMD silicon used in these clouds servers is aging. Based on the charts above, "AMD64 Family 16 Model 4" refers to a Phenom II X2 CPU. The slide copied above was posted in 2010. I agree, it seems odd seeing a Phenom part in server applications, and I will admit this article is the deepest dive I have done into AMD's server prospects. As a second data point, benchmarks point to an Opteron 4171 HE still being used in May of this year.
This represents two independent data points showing Microsoft utilizes the aging K10 architecture in the Azure servers. As a third data point, CloudHarmony shows Azure servers running on the Opteron 4171.
Another piece to this speculative puzzle is Microsoft's involvement in Heterogenous computing. In 2005, Mr. Herb Sutter authored a paper, pointing how processing limitations of processors were being reached. In 2011, Mr. Herb Sutter gave a presentation regarding heterogeneous computing and applications at AMD's Fusion Developer Summit.
Later, in 2012, Microsoft and AMD announced a collaboration between the companies regarding Windows Server 2012. In this news release, AMD discloses that Microsoft makes "extensive" use of Opteron in Azure servers, and that the collaboration between Microsoft and AMD has "deep roots."
Given the age of the AMD chips that appear to be used in Microsoft's Azure servers, the higher demand for Opteron 6300 chips could also signify Microsoft is in an upgrade cycle of replacing aging silicon. Worst-case scenario for AMD would be if Microsoft decided to go with Intel CPUs, but given Microsoft and AMD's involvement together, I do not feel this will be the case. And based on keynotes given regarding Microsoft's view of the future of computing, the companies' priorities seem aligned, meaning I believe we will see further collaboration between AMD and the software giant moving forward. This earnings call, pay attention to the Computing Solutions revenue, and any statements in the CFO commentary. If Mr. Kumar mentions anything regarding Opteron 4300 series chips, this could also signify Microsoft undertaking an upgrade, given the Opteron 4171 HE and 4300 series chips utilize the C32 socket, offering an easier and less expensive upgrade path.
Regarding the less traditional server markets, AMD is positioned to take advantage. There is growth in HPC and cloud gaming represents a new market. Microsoft's Azure servers could offer another big design win for AMD in the coming months, provided Microsoft does not switch to Intel. And AMD will be offering both x86 and ARM based server solutions, and will also offer iGPUs with these chips to increase compute power for servers using single-chip solutions.
Two specific, but unconfirmed, potential sources of revenue for AMD in the near term include Microsoft either building out Xbox One server infrastructure or Microsoft upgrading their Azure servers. Hints or confirmation for these include any clues during the upcoming conference call, or Microsoft disclosing specifics after the launch of the Xbox One. One potential source of news could be announcements made at APU 2013. Microsoft will be delivering a presentation on running GPU accelerated applications in the Cloud.
To be more competitive in traditional, big x86 core server markets, AMD needs to ensure Steamroller delivers on performance. Kaveri is set to launch at the tail end of this year. Take note of the benchmarks as they appear. Semiaccurate tested consumer grade Piledriver cores (the upgrade after Bulldozer), and Anandtech tested the Opteron version in February. Based on information I have read, I believe we will see the consumer class Kaveri chip prior to the Opteron version, but as you can see between similarities in the Semiaccurate and Anandtech reviews, performance gains reflected in consumer class chips are mirrored in their server counterparts.
Looking at this from an optimistic perspective, one could see the glass even greater than half full if measured from the edges of the meniscus and not the bottom of the dip. Bulldozer set the performance bar lower than it should have been set, and AMD is starting from a very small market share. If you are Intel and in the dominant position in servers, you are at the mercy of the overall growth of the market and must maintain your behemoth market share not to lose revenues. If you are AMD, gaining a small percentage of the overall market share translates to a large percentage increase in related revenues due to the "law of small numbers."
The cumulative of an ~5% in traditional servers and 10% growth of the server market represents around an approximate $2B revenue increase by 2016. How does AMD capture traditional x86 market share? By offering a higher performance part in Steamroller, having access more directly to customers via SeaMicro, and competitive IP in SeaMicro allowing AMD to build competitive server solutions. The success of AMD will largely depend on Steamroller performance, and we should get a preview of this potential by the end of Q4.
Additional disclosure: I own both shares and calls in AMD. I actively trade both my AMD and INTC positions, and may add to or liquidate positions at any time.