Acceleration Wars: Intel, Nvidia, Xilinx And Altera

| About: Intel Corporation (INTC)
This article is now exclusive for PRO subscribers.

Summary

Cloud computing inefficiencies can be expensive.

Intel’s failed attempt to acquire Altera shows how important this rivalry is.

Intel could license technology rather than make an acquisition.

Cloud computing has its own imperatives. Speed and power consumption are two of the main ones, once a cloud service is otherwise doing what it was designed to do. While software is important, speed and power consumption are highly dependent on the silicon processing the code. Making the cloud respond quickly to user requests is not a trivial task.

In the case of search services, the number of web pages is vast, so quickly getting a user to the best set of pages based on the user's search terms is crucial. Currently the two most used search engines are Microsoft Bing and Google Search. On April 23 Altera (NASDAQ:ALTR), a maker of a type of semiconductor chip known as PLDs or FPGAs, announced that its chips were being used to accelerate Bing searches. Earlier in April the possibility of Intel (NASDAQ:INTC) buying Altera had been widely discussed. The reason is that Intel is worried that it might otherwise lose the acceleration wars. Altera rejected Intel's $54 per share buyout offer, according to anonymous sources, though of course rumors have not entirely ceased to circulate.

In addition to Altera, rival Xilinx (NASDAQ:XLNX) and graphics specialist Nvidia (NASDAQ:NVDA) are major players in the acceleration space. Like Nvidia, AMD (NASDAQ:AMD) makes dedicated GPUs, but it does not appear to be a significant player in the space at this time.

Accelerators are nothing new to computing. They usually work alongside the general purpose computing chips, CPUs, that most people understand are in their PCs and other computing devices. General purpose CPUs, including ARM-based processors, do quite a few things well and quickly. They now often contain some accelerators on-chip to do common, specialized tasks like video conversions and graphics processing.

A GPU is a chip normally devoted to graphics processing but can act as an accelerator in the cloud for a class of highly parallel problems. Many practical and science problems can be broken into similar parts that can be done all at once, in parallel, to get a desired result.

A PLD is a chip that can be programmed to be dedicated to a single calculation (which may itself be complex) over and over again. Essentially all of the silicon is programmed at the hardware level to do one job. It typically does not need to consult an operating system to do its job. Raw data comes in at one end, and processed data goes out at the other end. For certain tasks PLDs are much faster and more power efficient than CPUs and GPUs.

Nvidia, Xilinx, Altera and possibly AMD will all likely benefit from the ramp in sales of accelerators for cloud, datacenter and HPC (high-performance computing) during the next decade. How much is that a threat to Intel, if it cannot develop or acquire its own accelerators? Table 1, showing 2014 revenue, profits, and y/y growth, highlights the competitors' situation:

2014 revenue,
$ billions
2013 revenue,
$ billions
y/y revenue
growth
2014 net income
$ millions
Intel 55.9 52.7 6% 11,704
Nvidia 4.7 4.1 15% 631
Xilinx 2.4 2.4 0% 648
Altera 1.9 1.7 12% 473
AMD 5.5 5.3 4% -403

The main takeaway from Table 1 is that Intel dwarfs each of its acceleration rivals, and in fact dwarfs their combined revenue and combined net income.

The acceleration play is not necessary to Intel's survival. It is an expansion play, with the lure of high margin products. While success in selling acceleration products would be a big plus for any of the other companies listed, for Intel it would be slightly more than background noise.

Given Intel's poor track record outside of general purpose CPUs, it may not be in the interest of Intel investors to pursue this course. Intel spent vast amounts of money trying to develop graphics technology, but in the end decided to not produce a discrete GPU and to license its graphics technology from Nvidia. Intel has failed twice at taking a significant share in the smartphone market.

Nvidia has benefited the most, so far, from accelerating datacenters and HPC. Nvidia does not break out how much it makes from this specialty use of its GPUs, but certainly has made it clear that the GPU coprocessors used in datacenters are high-margin products that contribute significantly to revenue and profits. AMD, the only clear rival in the GPU space, has not done a good job penetrating that market, which is one reason it consistently has lower profits than Nvidia despite generally having higher revenues.

Xilinx and Altera are traditional rivals in the PLD/FPGA market which covers a wide range of industrial products, but not much in the consumer space. They are going to fight over the acceleration market as well as struggle to take as much of it from Nvidia as they can.

The nature of the problems to be solved will be the main criteria for deciding how much of the acceleration market remains with Nvidia GPUs and how much of it is won by Xilinx and Altera PLDs. Obviously most operations having to do with graphics or video will be very efficient on GPUs. Highly parallel tasks, where GPUs already have the correct instruction sets, also will likely remain on GPUs. But many tasks could be more efficiently run on PLDs than on GPUs.

In response to an analyst question on April 23, Altera stated that customers themselves are saying PLD solutions are more efficient than GPUs, particularly when it came to power consumption. Both Altera and Xilinx have stated that what had delayed PLD solutions in datacenters was a shortage of programmers experienced with PLDs. The release of compilers by both companies that allow programs written in standard computer languages like C, C++ and OpenCL to be run on PLDs was the key factor in the recent rush to deploy the new technology. Altera stated that one factor that was helping it win customers was compiler templates for acceleration blocks for certain common tasks. In addition to datacenters, Xilinx reported it is selling acceleration PLDs for software defined networks and industrial IoT (Internet of Things).

Unfortunately neither the suppliers nor the consumers of accelerator chips have been sharing the level of revenue figures with investors or analysts. In the case of Altera the accelerators would be part of the networking and computer segment, which was 17% of revenue, or $74 million, in the first quarter. However, this trend is clearly at an early stage.

I can't put numbers on it at this point, but the likely winners going forward, compared to any base that existed in 2014, are Altera and Xilinx. There will still be plenty of acceleration business for Nvidia given the vast amount of gaming and watching of video that takes place on the Internet.

Before rushing out and buying Altera and Xilinx you should consider the rest of the respective businesses. Both have not done as well as expected lately due to a delay in the rollout of LTE cell phone service in China. Both are hoping for the rollout to accelerate this year, but the timing is up to the Chinese telecoms.

Intel could acquire Altera or Xilinx or a smaller PLD player. Or Intel could lure away engineers from them and start its own PLD division. That would likely take years of effort, and it would leave competition in place that's far more experienced and unlikely to be caught up with. So I believe the most likely outcome will be that Intel will simply concentrate on general-purpose CPUs for servers, which is a space that it dominates and has shown to be highly profitable. If Intel does pursue PLDs, they will likely license the technology.

Disclosure: The author is long AMD, XLNX.

The author wrote this article themselves, and it expresses their own opinions. The author is not receiving compensation for it (other than from Seeking Alpha). The author has no business relationship with any company whose stock is mentioned in this article.