The latest potential semiconductor industry merger is between AMD and Xilinx, a deal which has been reported, over the past couple of weeks, to be imminent, but has not been confirmed by either side. If an acquisition materializes, the price is expected to be about $30bn, and it would open the way for AMD to be a more significant player in the 5G landscape.
The main thing that AMD and Xilinx have in common is head-to-head competition with Intel. AMD, of course, has been battling the world’s biggest chipmaker throughout its 50-year life, and built a lot of its early success on its deal to license the x86 architecture to provide IBM with an second source processor provider for the PC. As that market slowed, AMD went through a painful process of evolving its products and business model to come out from under Intel’s shadow and forge its own identity. One aspect of its transformation, which has started to deliver considerable success in the past few years, was to target the souped-up cloud infrastructure that would support virtualized 5G networks – a prime target for Intel too.
Meanwhile, Xilinx came up against Intel when the larger company acquired Xilinx’s main competitor, Altera, in late 2015. Xilinx and Altera were the biggest suppliers of FPGAs (field programmable gate arrays) a category of chip that the former had invented as an interim solution between dedicated ASICs (high performance but also high cost and inflexible), and general purpose processors (very flexible and cost-effective, but not optimized for specific tasks).
The purchase of Altera was a sign that Intel recognized that its traditional strategy – assuming that every task in the cloud world would be able to run on general purpose processors – would not fly in the era of hyperscale and of very demanding workloads linked to 5G or AI/ML. At least in the near term, the most processor-intensive tasks, such as 5G dynamic spectrum sharing, would need to be offloaded from the CPU onto accelerators, based on more specialized chips, that could be optimized for particular processes. There are many approaches to acceleration, and many tasks to accelerate, but FPGAs, flexible ASICs and GPUs are all playing their part as the 5G and cloud platforms emerge.
The big chip providers increasingly want to assemble portfolios which include at least two or three, if not all, the main acceleration enablers. The trend has thrown a spotlight on FPGAs, which have also taken a far bigger role in 5G than in 4G because they can be reprogrammed while in the field, and so the rapid changes in standards and implementations in the early years of a new platform can be more easily accommodated than with inflexible ASICs, while maintaining performance levels beyond those of a general purpose CPU for certain tasks.
In addition, flexibility is becoming critical throughout the technology chain and lifecycle, with equipment and device makers increasingly wanting the ability to customize more extensively than in the past (see separate item on ARM).
All this makes FPGAs attractive, and with Altera gone, Xilinx was the next target. We had speculated that Qualcomm might make a move, particularly when the two companies cooperated to develop Qualcomm’s ARM-based server processor family. That was put on the back burner, but the companies remain close partners. Of course, Qualcomm and AMD are no strangers either – Qualcomm acquired the latter’s graphics technologies in 2009, making them the basis of its Adreno family, and five years ago was said to be interested in acquiring the rest of AMD too, though that deal did not materialize.
A purchase of Xilinx may not either, but it would strengthen AMD’s position against Intel and Nvidia in demanding infrastructure markets like 5G and cloud data centers, and would also extend its addressable market and diversify its revenues, since there are very few overlapping products. However, AMD would have to fund the purchase by taking on significant debt, or issuing additional stock, and its CEO, Lisa Su, tends to be conservative about such tactics.
Xilinx’s CEO, Victor Peng, on the other hand, though he has not indicated the company is for sale, might be more open to the idea, since he was previously a senior AMD executive. And Peng has been open about the way that Xilinx has been hit by the US/China trade war, with the prospect of being restricted in supplying Huawei and other key customers; and this uncertainty has been worsened by the impact of Covid-19 on the semiconductor supply chain and the delays to some customers’ deployment plans.
Among the Xilinx crown jewels for AMD would be its newest architecture, Versal ACAP (adaptive compute acceleration platform), which expands on traditional FPGAs to provide a far broader platform; and its wired and wireless group (WWG), which accounts for 32% of Xilinx revenue but is an area of weakness for AMD. Last month, Xilinx added PCIe cards to support virtualized 5G baseband units, in open RANs, to its range, seizing a leadership position in an emerging sector.
And earlier this year it signed a landmark deal with Samsung for 5G chips. Samsung will use Versal ACAP for complex, real time signal processing for 5G base stations, including sophisticated beamforming, especially in millimeter wave spectrum.
Versal ACAP features chips which, rather like SoCs, combine various elements – ‘adaptable hardware’ (programmable logic), scalar engines (CPUs), intelligent ‘engines’ (AI or DSP accelerators) and hard IP blocks for connectivity and security – to support specific high performance functions such as 5G or AI workloads.
This shows how the key roles of Xilinx’s offerings will still be to offload and accelerate compute-intensive tasks from the CPU, but they will rely on a variety of components and software to do that. Among the key tasks which Versal is designed to accelerate, according to Xilinx, are beamforming for 5G wireless communications, network packet communications, and smart controllers for large solid-state storage systems in data centers. Such demanding processes led Xilinx to create a platform with multiple different engines, each addressing a different type of requirement, from raw performance to power efficiency.
The Samsung news came hard on the heels of the announcement, in April, of Versal Premium, a high end member of the family targeted directly at telecoms networks. This is the third of six groups of products within the Versal portfolio to be announced in detail. Each one has a different combination of all the ACAP elements, according to its target applications. We have previously seen the general purpose Versal Prime and the Versal AI Core family for AI acceleration. Versal AI Edge and AI RF series – supporting RF and edge compute applications with AI engines – will follow later this year, while high bandwidth memory (HBM) series, an extension of Premium, will debut next year.
Kirk Saban, VP of product and platform marketing at Xilinx, said the increase in traffic and device connections driven by 5G would put the core network under huge stress, with operators needing to provision for up to 100 times more traffic capacity within the same space and power constraints. Traditional CPUs, GPUs and FPGAs would not address that challenge alone, hence the need for combinations of different elements.
In Premium, the focus is on adding hardened IP blocks to support networking processes. These include 400Gbps high speed cryptography engines for network security applications, multiple 600Gbps Ethernet cores and 112Gbps PAM4 high speed serial transceivers. The hardened IP saves a lot of space and power. In the previous generation of 16-nanometer Xilinx FPGAs, 22 chips would have been needed to program the capabilities Versal Prime offers from a single 7nm chip.
Such products would certainly extend AMD’s reach in the telecoms market. This could help it continue the momentum of a solid three-year run of growth, mostly at Intel’s expense. In 2018, AMD’s shares traded at around $10, but now they sit at around $85 – a market cap of some $85bn, compared to Xilinx’s $28.8bn.
Around 2015, after years of struggle, AMD finally achieved a competitive architecture, with its Zen CPUs and Vega GPUs. Then it released its Epyc line of CPUs, aimed squarely at Intel’s Xeons in the data center.