What is Intel Up To?
Intel has been very coy about their intentions for the deal with Achronix and whether this signals a huge turning point for the company. There are three explanations or theories that could be inferred based on what is currently known.
One theory, espoused by a number of analysts and journalists, is that Intel is interested in getting into the foundry business. While it’s possible that Intel is testing the foundry waters, this seems somewhat unlikely for both technical and business reasons.
Intel’s process technology is highly specialized for CPUs; while their SOC process is more suitable to the diversity of ASIC designs, it probably retains a high degree of specialization. For instance, Intel extensively uses design constraints to improve yields and overall manufacturing costs (e.g. by avoiding immersion lithography till 32nm). However, these constraints could be problematic to some foundry customers, especially those that are used to (or rely on) a much more permissive set of design rules. Additionally, Intel does not have the third party ecosystem that TSMC does – this encompasses design tools from Cadence, Synopsys and third party IP blocks. A customer would not want to retrain their entire design team to new tools just to use a different foundry.
From a business standpoint, the theory is even less plausible. The foundry business is relatively competitive and has low gross margins by Intel’s standards. TSMC, which is the dominant player and the most profitable, hit a high point of 50% gross margins recently; Intel’s most recent margins were 65%. Moreover, the foundry business is about to become a three way bloodbath. TSMC is the entrenched goliath, with a highly lucrative business. Global Foundries is entering the market with the proposition of tapping into IBM’s R&D and the backing of oil rich Abu Dhabi. More recently, Samsung, the second largest semiconductor company and the largest DRAM manufacturer, has announced that they intend to double their share of the foundry market annually and eventually rival TSMC. All three have the resources to spend heavily on new fabs and technology and will end up competing for business – leading to a great market for fabless companies and interesting times for the foundries. No sane company would voluntary jump into such a snake pit. Moreover, some of the large, cutting edge foundry customers are direct competitors to Intel (e.g Nvidia). Last, shifting the manufacturing group from internal to external customers is a huge change. It would be a fundamental reorganization around services, rather than product development and Intel would also have to constantly grapple with conflicts of interest (of which it has many due to sheer size and breadth).
A second theory is that Intel views manufacturing partnerships (generally) as an accelerated acquisition option – where the goal is to invest in innovative new companies and expand. Conceptually, Intel would view the manufacturing partnership like an accelerated call option. If they eventually want to acquire Achronix (or another similar partner), much of the technical integration work is already complete and the time to break-even would be accelerated. This in turn means that Intel can justify paying slightly higher prices for a ‘partially integrated’ company while still achieving the same (or higher) return on investment from such a deal. For a company like Intel with $4.3B in net income, it’s hard to get excited about a mere $50M-$100M/year in revenue. However, they could have similar partnerships with a several companies, and if any of them begin to hit sufficient revenue levels – Intel would snap them up. If this is the case, then subsequent partnerships should materialize in other interesting areas that are compatible with CMOS logic. Under this theory, the specific partnership with Achronix would represent Intel’s long term interest in the FPGA world, which seems possible.
A third theory is that Intel is partnering with Achronix to specifically complement and integrate with their existing product portfolio and roadmap. There are two obvious areas for such collaboration.
The first opportunity is a coprocessor for Intel’s high-end CPUs (and perhaps the successors to Larrabee). FPGAs have unique compute capabilities that can easily exceed the performance and efficiency of CPUs or GPUs for the right workloads. In HPC, this approach has been advocated by companies such as SGI or Maxeler. There was even a joint presentation from Maxeler and Schlumberger at Hot Chips 2010 about FPGAs for oil and gas exploration. Certainly, the NSA would love to have high performance FPGAs tightly coupled to CPUs – even more than the popcount instruction. In the commercial server world, Netezza has been using FPGAs to improve query performance for data warehousing for 10 years. FPGAs are even more popular in embedded communications, where they can act as DSPs for wireless basestations, networking line cards for wireline applications and many other roles. The fact that Achronix will have access to QPI strongly suggests a future as a coprocessor for high-end applications, whether they are in the server or embedded market.
The second pairing would be with Intel’s Atom-based SOCs, where FPGAs could be used by third parties to produce specialized variations. In some ways, the advantages are conceptually similar to the goals for Intel’s port of Atom to a TSMC process – system or SOC vendors can design their unique IP to work with an x86 CPU. However, the TSMC/Atom partnership does not appear to be bearing any fruit, so this could represent a second attempt by Intel to tap into the broader SOC market and enable third parties. In fact, Intel already announced a product (Stellarton) that packages Atom together with an Altera FPGA. Moving from package to chip level integration may be a logical next step, but obviously the FPGA would have to be designed on Intel’s process. It would also be a way to compete with embedded ARM cores (as soft IP blocks) in FPGAs, in essence offering an x86 core as a hard IP block. Of the two areas, this seems the most critical (long term) for Intel. However, QPI is no help here since it is not used by low power x86 designs and Achronix’s high performance and power targets may not fit well with the Atom target markets. Moreover, for broad use the software environment used for FPGAs is critically important and Altera and Xilinx have a pretty big advantage over Achronix – so in many respects they would be more logical partners. So while FPGA integration with Atom is probably very interesting for Intel, it does not appear to be related specifically to the Achronix deal.