AMD wants Xilinx: the battle with Nvidia and Intel for data centers and supercomputers
And again there is a mega takeover in the tech industry. AMD wants to take over Xilinx for no less than 35 billion dollars, which is about 30 billion euros. Xilinx won’t sound like a familiar name to the average consumer, and some user may not be fully familiar with the company either. Why is AMD willing to pay such a large amount for a company? What development underlies the takeover?
AMD is not known as a company that regularly makes acquisitions, but of course for that one big acquisition in 2006, when the chip company acquired video card manufacturer ATi for $ 5.4 billion. That was a large amount for a tech acquisition at the time. That acquisition not only enabled AMD to release its own video cards, but was especially important for the production of APUs, processors with an integrated GPU. The ATi acquisition led to a large debt burden, which was exacerbated in the following years by large losses and fierce competition with Intel. Incidentally, an antitrust settlement with the same Intel, for an amount of $ 1.25 billion , gave AMD a lot of air in 2009. Financially, AMD had few opportunities for major acquisitions.
Under the current CEO, Lisa Su, who took office in 2014, the situation has improved considerably. Under her, AMD has started to focus more on what she thinks the company is good at: designing CPUs for desktops and servers in particular. That resulted in the introduction of the Ryzen and EPYC generations respectively in 2017, putting AMD back on the map. Long-term debt has been significantly reduced and the share price has risen enormously since 2016 . That opened up opportunities for acquisitions and AMD is financing the acquisition with shares.
We are of course mainly familiar with the Ryzen processors for desktops and laptops, but in the meantime AMD has done good business with EPYC. While Intel had the server market in its grip for years with its Xeon processors, AMD now has a formidable alternative offering with EPYC. Thanks to their high memory bandwidth, the large number of cores, and relatively low price, they are particularly popular for high-performance computing, or HPC, such as artificial intelligence and big data.
| AMD EPYC Codenames | |||
|---|---|---|---|
| Gene. | Year | Name | core max. |
| 1st | 2017 | Naples | 32xZen 1 |
| 2nd | 2019 | Rome | 64x Zen 2 |
| 3rd | 2020 | Milan | 64x Zen 3 |
| 4th | 2022 | Genoa | ? xZen 4 |
The current Top500 of supercomputers lists eleven supercomputers with AMD. That’s not much, but in 2018 only two systems were based on AMD. Planning and building supercomputers takes time, and we’ve seen a lot of announcements for supercomputers with AMD hardware in recent times. These include the University of Stuttgart ‘s Hawk supercomputer, the Flemish Hortense, the El Capitan exascale supercomputer from the Lawrence Livermore National Laboratory and recently the European LUMI, which will be located in Finland.
From Zilog to Xilinx
AMD’s plans to acquire Xilinx have everything to do with the expected further growth of the hpc market, according to Lisa Su’s words during the announcement of the acquisition plans. Xilinx is an American company founded in 1984 by former Zilog employees who believed that there should be low-cost chips that customers could program in hardware to optimize them for specific tasks. This field-programmable gate array, or FPGAs, had to differentiate itself from generic chips. In the 1990s, the market for FPGAs grew strongly, especially for telecommunications and networks.
In addition, FPGAs were increasingly used for industrial, embedded, automotive, and broadcast applications. In recent years, FPGAs have gained popularity for a variety of computing tasks, such as with cloud providers Amazon and Microsoft. For example, Amazon offers Amazon EC2 F1 instances with Xilinx FPGAs to enable developers to deploy certain compute tasks in research, networking, security, video editing, big data, and analytics. Microsoft has now equipped every Azure server with FPGAs as an accelerator. The chips can not only handle a specific task faster but also be more energy efficient. In addition, flexibility is an advantage: the same chips can be set up in such a way that they can fulfill different functions.
The rise of adaptive computing
According to Lisa Su, the market for high-performance computing is growing rapidly to cope with the increasing computational workload and FPGAs would play a greater role in this. Research from GrandView supports her expectation: the estimated total market for FPGAs was about $9 billion in 2019 and that value is expected to grow at an annual rate of 9.7 percent from 2020 to 2027. Su speaks of ‘adaptive computing’, where hardware is adapted down to the chip level to the computing tasks for which it is used. This is one of the ways the computing market is trying to look beyond Moore’s Law .
Su refers to ACAP, or Adaptive Compute Acceleration Platform. Xilinx markets its Versal offering under this name. These are chips produced by TSMC at 7nm, including ARM Cortex cores, programmable network-on-a-chips, DSP engines, AI engines, PCIe interfaces and CCIX interconnects. For example, Samsung and Ericsson will use ACAP for 5G networks.
The chips work in combination with Xilinx’s Vitis software platform, which should make programming the chips easy. By combining the various components on a compact chip, Xilinx now has the necessary experience with packaging, in particular with the vertical stacking of various components or chiplets, knowledge from which AMD can benefit. In the future, AMD could expand upcoming EPYC processors with programmable components that users can optimize for their computing needs.
Currently, data centers are still responsible for a modest part of Xilinx’s turnover: 11 percent. The company derives nearly half of its revenue from the aerospace, defense and industrial systems industries. About 29 percent of the turnover comes from the network market and 16 percent from the automotive and broadcast sector. According to AMD, the merger with Xilinx would create a company with a turnover of $ 11.6 billion. Xilinx reported a turnover of $ 3.16 billion in its recent annual figures, AMD had a turnover of $ 6.7 billion last year, but expects its annual turnover to increase by 41 percent this year. AMD had about 11,000 employees at the end of 2019, Xilinx has about 5,000. Together, the companies will have 13,000 engineers , according to AMDand R&D investment totals $2.7 billion.
Xilinx’s main competitor is Altera, which also produces FPGAs and thus serves 32 percent of the market, according to AMD, while Xilinx’s market share is said to be 54 percent. Altera was acquired by Intel in 2015 for $16.7 billion, still Intel’s largest acquisition. Whether that takeover has been so successful so far is a matter of debate. Quarterly revenue for Intel’s Programmable Solutions Group, which includes Altera, fell 19 percent from last year to $411 million. In any case, Intel will also have competition from AMD in this area.
Nvidia and Intel
And then there’s Nvidia, which acquired Mellanox last year for $6.9 billion. Mellanox makes adapters, switches and chips for InfiniBand and Ethernet networking equipment used in supercomputers and data centers. Xilinx, which was also reportedly in the running to acquire Mellanox, then bought Solarflare , which also develops high-bandwidth, low-latency network interconnects. AMD thus also gets this technology in its hands.
The plans for a takeover that hit like a bomb this year were those of Nvidia, which wants to acquire Arm for 40 billion dollars . We know Arm as the designer of energy-efficient chips for smartphones, but the SOCs are increasingly being used for more powerful devices and also for supercomputers. In fact, the most powerful supercomputer today is Fujitsu’s Fugaku supercomputer with 7.3 million ARM cores . Nvidia therefore wants to take over Arm mainly because of the growth opportunities in the market for artificial intelligence and other HPC applications.
If AMD and Nvidia’s acquisition plans receive regulatory approval, the cards for chips for data centers and supercomputers will be reshuffled. Intel in particular can prepare for fierce competition. As we all use more applications involving complex computation, the battle for the chip market will increasingly be fought out in the ever-expanding computing centers and computing clusters. AMD wants to strengthen its position considerably and broaden its portfolio considerably. Before that happens, we will be a year further: the acquisition could not be completed until the end of 2021 at the earliest.
You can read more information about FPGAs in this background article: Programming at the transistor level – FPGAs on the rise, also for the hobbyist