Reading Intel analysis lately has been a lot like reading fantasy baseball analysis. Intel should buy Altera. Intel should waive Atom. Intel should fab for Apple. All of those have a near-zero probability of happening IMHO, and yet pundits continue to pitch their version of alternate reality, dealing away product lines and strategies left and right.
Let’s debunk this fantasmagorical deal-making – and no, I do not own any investment position in Intel or the other companies I’m about to mention.
I’ve been to a baseball game, one World’s Fair, a picnic, and a rodeo, and the stupidest thing I’ve ever heard come over a set of earphones is: Intel buying Altera. The alleged Xeon-with-FPGA-glued-on isn’t the first Intel attempt at such a contraption, as many outlets reported. There was “Stellarton”, an Atom core with an Altera FPGA alongside via PCIe in a single package. Intel released no details on what is in this Xeon E5 thing with their tease, and it isn’t eminently clear it is Altera on top – it could be Achronix or Tabula, when you look at their tools and experience with Quick Path Interconnect and 40 and 100G Ethernet.
Whoever it is, and I find it odd they didn’t name Altera if it is them, Intel is probably frantically trying to iron out details of those programming tools and support channels before a release into the wild. This is a data center play for workload-optimized servers, by providing FPGA-based QPI integration and hardware acceleration of algorithms such as pattern matching and protocol parsing. The results from a Microsoft experiment on Bing accelerated with Altera technology are promising.
No, if Intel had wanted to buy Altera, they wouldn’t have signed a 12 year fab technology deal, and a non-exclusive one at that – they would have already taken them out. One would need somewhere around a 1.3x revenue multiple, maybe as high as 2x, to acquire a semiconductor target. For Altera, that’s $2.2B to $3.4B in cash or stock. Intel certainly could do that with $30B lying around.
However, shelling out that much just to compete in the workload-optimized server segment doesn’t make a lot of sense. It is far cheaper just to license FPGA technology and get the same effect. Plus, an acquisition of Altera would almost guarantee both Achronix and Tabula to immediately hit the exit button from the Intel foundry biz, perhaps with penalties or legal action. While that might not have a big revenue impact, the optics of competing with customers gets magnified quickly.
The final nail in this bad idea coffin: Intel has no intentions of taking on Xilinx on their turf, risking a pummeling to go with the one they are taking in mobile. It is a perfectly acceptable fight for Altera to wage; they have just as much of a claim to FPGA heritage as Xilinx does, and they are in the same league. I’m sure Xilinx would love to matchup with a much bigger Intel in a short series, but I doubt they’ll be given that chance.
Speaking of the mobile pummeling … I tell ya, embedded gets no respect, no respect at all. Some pundits don’t understand that Atom is about more than mobile and tablets – a lot more. I agree that Intel may finally come to their senses and step back from the smartphone before anyone else gets hurt. I also agree that the 2-in-1 is “Haswell” territory.
There is a whole world of billions of parts requiring much lower power – cars, TVs, set-top boxes, and the fog computing needed to drive the IoT being just some examples. That lower power core is “Silvermont”. I think analysts got confused when Intel made a big splash trying to pass off “Silvermont” inside “Bay Trail” as the tablet solution, forgetting that “Silvermont” also powers embedded Atom families that far outsell tablets. CEO Brian Krzanich, from the 1Q14 earnings conference call:
The Internet of Things Group’s Atom product nearly doubled over Q1 of last year.
You know that didn’t come from tablets, right? Intel is not going to pull a growing embedded business out of the game, no matter what name they decide to label the org chart with next. Atom is the flag-bearer for new embedded design wins. The Edison maker module has been retooled to carry an Atom variant, and the automotive platforms Intel released a few weeks ago also carry an Atom part.
Here’s another thing most pundits completely overlook. “Haswell”-based parts designed for 2-in-1s have a best-if-used-by date, typically 18 to 24 months before they go EOL and are replaced with the next version. Embedded designs have much longer lifecycles – a part on Intel’s embedded roadmap is secure for 7 years. While Chandler can cherry-pick “Haswell”-based Core variants for the embedded roadmap, it has to be with the full blessing of Santa Clara and special handling to keep those versions around. A big chunk of the embedded roadmap today is Atom.
The other overlooked aspect of “Silvermont” is the central role it plays in manycore. The Xeon Phi coprocessor, in the latest version known as “Knights Landing”, will pack 72 “Airmont” cores (the 14nm version of “Silvermont”). This is a beast, designed to compete with the NVIDIA Tesla in the high performance computing market, but also as a coprocessor for server applications. Welcome to the lawless post-Moore world; there is no way we ever see 72 “Haswell” cores on any one piece of silicon.
And that brings us to Apple. I can assure you that when it comes to understanding a sales pipeline, capacity planning, and high level relationships, Intel is world class. If there had been a snowball’s chance in Chandler that Intel could have won Apple’s ARM-based A7 business, or its successors, Fab 42 would be open, or at least there would be tooling rolling in for a 2015 start. The decision to mothball was not taken lightly, and there is zero indication of the tarp coming off.
The hot rumor right now is Samsung and Globalfoundries are tooling up Fab 8 for a test run on Apple A9 parts in 14nm. TSMC is still out there, looking to 16nm FinFET for the A9. Intel “still has a chance” according to sources, which is the right thing to say until somebody actually gets 16nm or 14nm working in the majors. Of this list, it has the best chance of coming true, but it is going to take a meltdown by someone else for it to happen.
It all makes for great fantasy tech-ball. Agree, or disagree? What would you deal if you were Intel?