Sellings: How does ARM's rise and post-Altera "survive"?

Intel and Altera have been on and off several times since the news of their acquisition. The stock I bought went up $5, and I rewarded myself with a McDonald's meal, but that information might have been more valuable to pre-acquisition Altera.

On January 16, 2008, I was at a press conference with Intel's strategic marketing people to discuss the embedded market. It was more like a cross-examination, with Intel asking open-ended questions about trends and the competitive landscape, but not about acquisitions because Intel didn't share information about itself at the time. I casually mentioned the concept of "SoC reconfigurable", where placing an FPGA near a processor can speed up subsequent I/O. After hearing this, most people looked at each other and took notes, but got no response, let alone detailed architectural questions.

At the time, I confirmed that Intel was already thinking about it, had even begun to work on it, and perhaps my problems had sown the seeds. Two and a half years later, at IDF 2010, they introduced the Stellarton, a 45nm Atome600 processor packaged with an Altera FPGA.

On March 1, 2011, Cyrus released the Zynq-7000 series. They used an actual embedded kernel, ARM-A9, and tightly integrated two 28nm programmable logics (processors), including PL and PS between the enhanced cache ports. Artix-7 or Kintex-7 devices, logic units 28K to 444K. The smallest version is 225 pins, 13X13mm BGA, and the largest is 900 pins in a 31X31mm package with I/O coverage equivalent to dual-core speed.

Silings won this round, no doubt, with a combination of IP (ARM plus their own), manufacturing (TSMC), architecture and packaging, and the software we mentioned in Vivado. At the time, Intel had not yet discovered integrated and small packages, or even quark cores. Altera has thoughtfully responded that ARM-based SOCs already exist, but are still lagging behind Zynq in market share, partly because of the capabilities of manufacturers such as AFR's Zedboard series and Adapteva Parallela, and partly because of manufacturing process issues.

To make up for the process shortfall, Altera and Intel seized on the February 2013 14-nm FinFET process node. FPGAs are a great way to demonstrate that advanced process nodes make sense. In March 2014, this agreement was extended to cover multiple wire drawing system packages. A few months later, Intel began talking about Xeon parts, such as Knights, with FPGAs on chips to speed up landing, but not Altera. The goal is a new "workload optimized" server processor.

Readers know that we sometimes write about "workload optimization" processors. At this point, the question becomes why buy the cow when the milk has already been produced? A year ago I wrote that Intel would license technology, but not acquire it. So why did Intel sell Altera for $16.7 billion at 8.6 times its estimated price? I was puzzled. My prediction was that we would buy at twice the pre-estimate.

This keeps Altera's technology from falling into the hands of other companies, such as AMD, AppliedMicro, Broadcom, Cavium, and Qualcomm, all of which use 64-bit ARMv8 cores for low-power, high-density SOCs, or are already in the process of adopting them. Meanwhile, an anxious Intel pawed in when Broadcom was sold to Avagra for $37 billion at 4.4 times its estimated valuation. If Intel wants to keep Arm at the door of the data center, this is the best time to do it. (However, there may be another problem here. Altera has a range of military customers, about 22 percent of them -- including industrial, military and automotive. Expected: Intel's spin-off unit will need Altera's former defense business regulator to approve the deal.)

Meanwhile, Sellings pressed their advantage. In November 2013, their first batch of UltraScale products was ready under TSMC's 20nm process. In February 2015, they announced that Ultrascale+ would be manufactured on TSMC's 16FF+ process line, and began talking about upgrading the Zynq as well. Just an upgrade or an underestimate, because the Zynq Ultrascale + MPSoC is completely redesigned to contain not just one, but three ARM IP cores -- the application processing unit is equipped with dual Corcomb-R5 cores, and the graphics processor unit is equipped with Mali-400MP.

Sellings: Living in the era of ARM and post-Altera

"It's not a fully embedded SoC." At least in terms of solid configuration. The smallest version is 484 pin in a 19X19mm package, while the largest is 1924 pin in a 45x45mm package. Their product listings and selection guides tell the whole story. In fact, as you can see from the table above, most of it is for infrastructure.

On the one hand, we have new, upgraded Intel-Altera, perhaps with a small defense product that can be used for this application, for server and cloud services.

On the other hand, we have the hot ARM architecture chip company, which has great potential to support server and client, mobile or embedded SoC.

Re/Code quotes Patrick Morehead on FPGA:

ASIC prototypes and workloads are optimized when the company works for the server CPUs. One thing he forgot was that when building equipment is in low volume, demand cannot judge the cost of ASIC customization. This is why Silings has a huge market in defense and video broadcasting circles -- annoying processing requirements that are suitable for FPGA accelerators to handle, but not much system building. Typical radar and image processing architectures have high data rates, which can be handled by FPGA at the front end, and low data rates by common CPU and software.

Of course, this requires two kinds of chips. A regular Cyrus Virtex Ultrascale + FPGA can of course be connected to a Broadcom SoC with an ARMv8 core. (Interestingly, Avago-Broadcom acts as the server platform.) But its two chips and cache consistency are still problematic in some applications. For other reasons, if nothing wrong, inheritance wins.

Interestingly, Morehead pitched Sellings as many suitors as he could: IBM and Qualcomm. In the longer term, IBM, which was forced to replace OpenPower, is unlikely to buy Sellings's struggling business. Qualcomm is rich and willful enough to buy Cyrus and buy its customers.

Another opinion, mainly from Steve Casselman and analyst Gus Richard, suggested that Silings should switch to the defensive and buy AMD. This will allow Silings to get into the X86 business, removing some of the uncertain ARM services, and what's more exciting is that GlobalFoundries can get involved. Casselman also noted that Sellings has a leading edge in real-time processing and locally reconfigurable technology.

There is also a cultural factor. Intel has learned from the Wind River acquisition, leaving them as a separate subsidiary to the original brand. Altera could be going the same way, because if Intel doesn't bring them in, there will be a lot of attrition at the end of the two-year agreement. Cyrus may not have the strong culture to satisfy other SOC companies.

As you may know, Sellings has a prototype topology based on FPGA and has a complete embedded programmable SoC. An embedded platform is an external card. Laitis still spends most of his energy on logic to strengthen the business. Microsemi has fared well in some lucrative markets, notably in the controlled temperature part of aerospace, but has not been able to compete with Silings. Depending on what Intel decides to do, Altera may not be able to get into the embedded SoC business.

FPGA still has the problem of programmable in data center business. Intel is trying to carve a way out of OpenCL because FPGAs can speed up special instructions in multi-core Xeon solutions. With Sellings embracing the SDACCEL environment at the end of 2014, design and code services are going through a huge transformation.

If you boil this acquisition down to the Intel-Xilinx race, the scene is messy. On the one hand, if Altera were to stand alone, Intel would have a tough time in the embedded space where Cellings is. On the other hand, Intel simply can't match their size in servers -- they have an advantage in infrastructure that they want to defend aggressively. I don't think Sellings will lose here, but most of the media is predicting it will.

When all the members of the ARM architecture agree -AMD, AppliedMicro, Avalgo, Cavium, Qualcomm, and Cyrus are at the top of the list - it's similar to the battle between Apple and Android in the smartphone space. To be precise, ARM is like a bank. No single investor is dominant, but if many investors band together, they might catch up with or surpass Intel in the server space. However, ARM must find the right information. Power consumption and mold size do not highlight their position in the mobile field, customization is the best way.

I hope the low volume performance will continue to work. Sellings will win the competition through Zynq Ultrascale + MPSoC in some areas such as large broadcast centers, or the Internet of Things hybrid cloud, and other areas that look more like embedded. The number of wins for these types of SOC is going to increase dramatically, and Silings is going to grow dramatically -- at the top of the table are the traditional FPGA and embedded SOC businesses. At the same time, Qualcomm will go into experimental mode to try to find the right partner in conjunction with the ARM SoC-based network infrastructure. A combination of Intel and Altera will win in the high-performance computing environment business, perhaps running IBM Watson in the cloud, so how will Altera's legacy business change?

The interaction between Altera and Xilinx will disappear, and it will be very different from the current situation. At that time, there will be no such thing as "FPGAs" -- just a lot of different programmable technologies, and they will have to keep evolving.

Everything that Sellings got went well, and the real battleground was not the foundry -- which is important, but not the whole answer. Once chip size, power consumption, cost, yield, and reliability take hold, it boils down to competition for stability, integration, programmable tools and software support. Sellings has the capability to support embedded and data center applications. It doesn't make much sense to have only one foundry or one data center.

Just as Avalgo's takeover of Broadcom changed the financial landscape, so too could things change dramatically if Silings suddenly found itself a seller or a buyer. From what you see, what are the opportunities here?