Oct 17 2020 - AMD To Acquire Xilinx: The Imitation Games Continue

AMD rumored to acquire Xilinx for B, following on the heels of Nvidia acquiring Arm.

In my Nvidia-Arm analysis, I remarked Nvidia is imitating Intel to at least some degree, with the reverse also being true.

A Xilinx acquisition by AMD would imitate Intel’s 2015 Altera acquisition. Potential synergies include AI, DPU and 5G. Data center is also a small, but growing segment.

AMD's data center group is currently led by the former VP of Intel's Programmable Solutions Group.

I called Nvidia-Intel the 'semi battle of the decade'. It seems AMD does not want to be left behind.

Investment Thesis

On the heels of Nvidia (NVDA) acquiring Arm, reports have emerged that Advanced Micro Devices (AMD) would be close to acquiring Xilinx (XLNX) for B. The acquisition would be highly reminiscent of Intel's (INTC) acquisition of Altera in 2015.

Looking at the FPGA market itself, it seems a mixed bag. Both Intel and Xilinx have seen mixed results in recent years. While leading products such as Stratix and Versal are likely growth drivers, uptake of FPGAs in the data center and for AI likely remains relatively minor. Nevertheless, the promise of the democratization of FPGAs persists for both companies.

Looking at it from a broader view, competition between Intel and Nvidia has been intensifying as the future of compute becomes more heterogenous. Acquisitions such as Mellanox, Altera and Arm also indicate consolidation in the industry. Nvidia shifting to CPUs and Intel developing a GPU stack further indicates imitation in the industry as many of the leading silicon providers pursue the same markets to drive further growth.

So, with AMD's strengthened position in recent years, it seems it does not want to be left behind. In terms of execution, AMD's data center group is led by the former leader of Intel's FPGA group.

AMD's and Nvidia's market valuations are both indicative of high expectations to continue their growth, so such acquisitions seem in line with expectations in the market.

What happened previously

Towards the middle of last decade when smartphones were maturing, the industry was sort of facing the question of what would come next. Much of what was heard around that time was about the emergence of IoT, with the expectation of 50 billion devices by 2020, for example smart watches. While, indeed, there have been many investments and growth in IoT, there has been no such explosion (yet). Instead, the biggest driver of demand became the data center and cloud, and that was were Intel shifted its resources primarily towards, in conjunction with the emergence of AI (while also growing its IoT business at a steady clip, to be sure).

But before AI fueled billions in additional silicon demand, there had been expectations of FPGAs becoming much more predominant in the data center, with some estimates of 30% uptake by 2020. With so many 'threats' of third-party silicon potentially cannibalizing CPU demand in the data center, Intel eventually would invest in the whole portfolio of heterogenous compute.

While FPGAs didn’t quite pan out (yet), since GPUs were much more apt for AI training than CPUs given their inherently more parallel nature, Nvidia did emerge as a new player in the data center. To further bolster its presence as a strong, and potentially independent end-to-end data center provider, Nvidia further acquired Mellanox and recently Arm.

So, in recent analysis of this acquisition, I remarked that Nvidia seemed to be imitating Intel more than Intel was imitating Nvidia, given its pivot towards CPUs: Nvidia Vs. Intel: The Semi Battle Of The Decade. To that end, I indicated that Nvidia still lacked memory and storage, so I suggested it might acquire Micron (MU) next. Adding FPGAs might have been another possibility, but it seems another company has other plans about that.

AMD: The dark horse

Since my analysis of Nvidia vs. Intel omitted AMD, some mentioned AMD in the comments. This is perhaps justified given its increasingly more competitive position. AMD has been conquering the desktop enthusiast segment with no response from Intel, moved to TSMC 7nm well before Nvidia, and the expectations of the upcoming Big Navi are also quite high. AMD has also increasingly differentiated its investments for the data center, for example by developing the upcoming CNDA2 architecture.

AMD’s elevated ambitions were also seen in several other points this year. Its long-term model clearly indicated that it wanted to increase its gross margins. This was recently seen in the price positioning of Zen 3. AMD argues it has a leading product proposition, and charges accordingly. No longer is AMD a budget CPU vendor. While outsiders will often talk about market share, AMD’s strategy is not to grow share necessarily at whatever cost, but to grow both top and bottom lines.

Since many contend AMD has competitive products both in CPUs and GPUs, perhaps the real 'semi battle' would really be Nvidia vs. Intel vs. AMD of perhaps three almost equally competitive companies with a broad product portfolio.

However, as I detailed, for example, in this article, AMD has been lacking in the 'product portfolio' aspect: AMD Is Not Half An Intel. In that regard, an acquisition might change that.

AI

The AI revolution has been entertaining because every silicon vendor deemed its silicon apt for AI workloads.

• Intel argued CPUs only appeared to be worse at AI because its software was not yet highly optimized in the early days, and pointed to its AVX-512 vector units in its CPUs, and also at inference workloads which often differ materially in requirements from training workloads.
• Dozens of start-ups suddenly also found themselves vying for a new, emerging market, which renewed venture capital interest in silicon development.
• Mobile vendors, too, were quick to add IP for AI in their silicon, and Qualcomm (QCOM) is also entering inference in the data center: Qualcomm Cloud AI 100: Impressive Specs, Competition To Nvidia, Intel.
• And FPGA vendors, likewise, argued their silicon provided key benefits such as on-the-fly programmability and competitive processing performance because of the DSPs they contain.

5G

There has also been another notable trend, 5G. With 5G comes the emergence of software-defined networking, which changes the role traditional CPUs as well as FPGAs play in the network.

To that end, Xilinx has also been adapting, and calls its latest 7nm FPGAs not FPGAs, but ACAPs: adaptable compute platform. Xilinx is developing a range of 7nm Versal silicon with dedicated IP geared towards networking, AI, edge, etc.

SmartNIC

Nvidia recently announced its DPU roadmap, leveraging assets from Mellanox. Intel's answer seems to be its SmartNICs. As Intel describes:

"A SmartNIC is a Network Interface Controller that accelerates networking functions and is adaptable to existing and emerging use cases including security, virtualization, storage, load balancing and data path optimization. SmartNICs extend Foundational NICs and provide more advanced functionality by including a programmable engine such as CPU and/or FPGA to accelerate tasks that the server CPU would normally handle."

Since they are FPGA-based, AMD would be able to compete in this segment of the data center if it acquires Xilinx.

Heterogenous compute

The trends above are not new. They have been and are still unfolding. Early last decade or so, they were called 'More than Moore'. More recently, Intel (Raja Koduri) has called it the architecture era. While Moore’s Law is definitely not over, the industry has recognized the role that more specialized and heterogenous silicon can play: in fact, leveraging Moore’s Law to include dedicated IP for specific workloads (such as AI, vision, etc.). In most SoCs, the actual CPU cores only take up fraction of total silicon area.

Xilinx acquisition

Simply put, competition is increasing, and so has consolidation. This has resulted in fewer players, but those want to play an increasingly end-to-end role as a silicon provider for their customers, whether in the data center or at the edge. I earlier pointed at Nvidia and Intel as the two giants that were increasingly competing against each other in CPUs, GPUs, data center, self-driving silicon, IoT, etc.

By acquiring Xilinx, reminiscent of Intel’s Altera acquisition half a decade ago, AMD would make a point that it doesn’t intend to be left behind as just a CPU and GPU provider for desktops and consoles. Even more so, AMD by market share would actually be bigger than Intel in FPGAs.

Analysis

The general concerns about the FPGA market lately has been that despite all promises, they have not or do not seem to come to fruition.

Altera’s revenue has been basically flat over the last half a decade. Xilinx has seen some growth, but also declines. While FPGAs are clearly growing in the data center, it is still a fairly small market. FPGAs are used in a variety of markets, but growth in one segment often seems to be offset by declines in others. (Most recently, Xilinx saw declines in wireless, despite the build-out of 5G, while Intel saw growth in that segment.)

Summed up, while some reports may mention AI and the data center as applications for FPGAs, those still do not seem to be a major revenue driver for that market (yet), even though it may be a growing segment. For example, Intel had been vocal about plans for increased integration between the Altera FPGAs and its Xeon CPUs, but it seems to have largely shelved those plans, in favor of an interconnect approach via its CXL open standard.

On the other hand, some have also remarked that just as AMD’s financial profile is improving, AMD would be taking on a lot of debt. Nevertheless, while one can disagree with the market, the market is valuing AMD at ~0B despite revenue of less than B, so an acquisition would be in line with the general trends in semiconductors and likely the expectations of the market of AMD to be a third competitor against Intel and Nvidia.

Dan McNamara

Lastly, something I haven't seen mentioned in most reporting, it might be noted that AMD already has some FPGA experience in-house, as its data center group (Server Business Unit) is led by former SVP of Intel's Programmable Solutions Group, Dan McNamara.

Shortly before his transition to AMD, he had in fact been promoted to SVP of a newly created Network and Custom Logic Group:

"And opportunity because when these customers discuss their needs — be it how to better leverage data, how to modernize their infrastructure for 5G or how to accelerate artificial intelligence (AI) and analytics workloads — they realize the massive prospects in front of them.

To help our customers capitalize on the opportunities ahead, Intel has created a new organization that combines our network infrastructure organization with our programmable solutions organization under my leadership. This new organization is called the Network and Custom Logic Group.

Both original organizations executed on record design wins and revenues in 2018. Their merger allows Intel to bring maximum value to our customers by delivering unprecedented and seamless access to Intel’s broad portfolio of products, from Intel® Xeon® processors SoC, FPGA, eASIC, full-custom ASIC, software, IP, and systems and solutions across the cloud, enterprise, network, embedded and IoT markets. To that end, FPGA and custom silicon will continue to be important horizontal technologies. And this is just the beginning of a continuum of Custom Logic Portfolio of FPGA, eASIC, and ASIC to support our customers’ unique needs throughout their life cycles. No other company in the world can offer that."

By acquiring Xilinx, AMD would be able to pursue a similar opportunity as Intel in the 5G network infrastructure segment, which it announced early this year it would enter.

Software

One piece often mentioned is software: CUDA, x86, Arm. To that end, AMD does not really have a software ecosystem of itself. For example, Intel management will most often use the term 'Intel Architecture' instead of x86. Intel has also been developing oneAPI as an open industry standard.

Nevertheless, given the mention of oneAPI, the extent of software as a differentiator can nevertheless be questioned. For example, oneAPI recently received support for AMD GPUs. So, while oneAPI seems clearly positioned as Intel's alternative to CUDA, others might also benefit from this model.

Xilinx’s leading position in FPGAs in any case takes away concerns about software on that side; both Intel and Xilinx have their own FPGA ecosystem.

Summary: Imitation Games

• Intel becoming a GPU vendor
• Nvidia becoming a CPU vendor
• Intel and Nvidia Mellanox acquisition/data center networking
• Intel and Nvidia data-centric, IoT-centric, AI-centric, self-driving-centric, software-enabled strategies
• AMD becoming a FPGA vendor (acquiring Xilinx)

Takeaway

Synergies between CPUs and FPGAs in the data center arguably have not been fulfilled at anywhere near the pace once expected, for example by Intel likely. Nevertheless, the promise of the democratization of FPGAs through higher-level programming tools is still there. For example through Intel’s OpenVINO and oneAPI tools. Xilinx's ACAPs are also an indication of this.

FPGAs can also be leveraged for AI workloads, which both Xilinx and Intel have been pursuing, via their ACAPs and Stratix 10 NX with AI Tensor Blocks, respectively. Furthermore, (even with the emergence of software-defined networking) wireless will likely also remain a core market for FPGAs, a segment that AMD this year announced it is also pursuing with its Epyc CPUs with Nokia.

To date, AMD has been watching Intel and Nvidia both growing and acquiring companies to become more relevant in the data center and beyond, increasing their competition. With AMD this year laying out a long-term financial model for increased revenue and margins, and given its own increased investment in the data center and emergence as a more competitive company since its Zen CPUs, it seems AMD, too, is starting to look beyond the CPU and GPU to further growth opportunities.

To that end, adding FPGAs would add a new computing platform to AMD's portfolio, with some synergies in AI, 5G and data center. Xilinx itself was already on a path to increase its relevance by defining the new 'ACAP' category as FPGAs with additional IP, connected by a network-on-chip (NoC).

In a previous analysis, I characterized Nvidia and Intel as the semi battle of the decade, but it seems AMD does not want to be left behind. Recent trends such as increased pricing of its Zen 3 CPUs further indicate that AMD wants to receive its fair profits from the (competitive) silicon it sells.

AMD, just like Nvidia, has received more-than-premium stock valuations, so these companies expanding their markets to continue their growth seems in line with expectations as well as industry trends in the last half a decade such as AI, 5G, edge/IoT, data center and consolidation.

Concerning Intel, while FPGA leadership on technical aspects can be discussed, in terms of market share/revenue, Xilinx is still well above Intel, so this acquisition would definitely add to their multi-decade competitive history.