Intel, Achronix and Manufacturing
Intel is a semiconductor manufacturing company that happens to design x86 microprocessors (quite good ones, incidentally). It is first and foremost an Integrated Device Manufacturer (IDM) – a company which designs, manufactures and sells its own chips; the nature of these chips is somewhat secondary. This fact isn’t obvious to many people outside the industry, because Intel is most famous for their Pentium microprocessors and has spent tremendous money and effort raising awareness through marketing campaigns. Intel is much akin to iceberg; manufacturing is the invisible mass that is the bulk of the company, with roughly half the head count and the first 4 CEOs (Noyce, Moore, Grove and Barrett).
Intel’s high-volume logic manufacturing prowess is unmatched in the industry. IBM has equally talented research and development, but silicon is just one dimension of their overall system-centric approach (e.g. packaging, interconnects) and they are a low volume and high cost player. As a pure-play foundry, TSMC has great volumes through their fabless customers, but less impressive R&D and lower performance silicon. The nascent Global Foundries has the potential to marry IBM’s superb R&D expertise with high volumes, but it is far too early to tell. The bottom line is that Intel perceives manufacturing as a key differentiator and core competency and has the results to prove it:
- Intel is 6-18 months ahead of competitors to new manufacturing nodes (e.g. 32nm in late 4Q09, Global Foundries and IBM 32nm expected in 2Q11).
- Intel was the first to commercially implement strained silicon.
- Intel was the first to commercially implement high-k gate dielectrics and metal gates and a full generation (3+ years) ahead of everyone else (45nm vs. 32nm).
Last Monday, Intel announced that they had voluntarily entered into a manufacturing partnership with a third party, one of the first times in the company’s history. This broke an implicit assumption held by many observers, and prompted quite a few questions, which we will endeavor to answer.
- How could Intel open up their crown jewels to a third party?
- Does Intel want to get into the foundry business?
- Why is Intel doing this?
To answer these questions, it is essential to first understand exactly what was announced. Intel’s partner, Achronix is an FPGA start up (under $10M in revenue last year). Achronix uses novel asynchronous logic techniques and specializes in producing high speed FPGAs. Their current products are fabbed on TSMC’s 65nm process and reach 1.5GHz, roughly 3X faster than FPGA leaders Altera and Xilinx. Achronix’s second generation Speedster 22i FPGA will be fabbed on Intel’s 22nm process technology, with engineering samples expected by 4Q11 – simultaneous to Intel’s initial ramp of 22nm. The capacity and revenue for Intel will be miniscule (<1%), although they have also said that other deals may follow. Reportedly, Achronix will also have access to Intel’s cache coherent QuickPath Interconnect (QPI). This strongly suggests that Achronix will be using Intel’s high performance 22nm process technology (P1270), since it is unclear if Intel has designed QPI to work on their low power process and SOCs.
There was a bit of amusing marketing that Achronix will have the only FPGAs made in the US – which is true, but probably irrelevant. Actel (recently acquired by Microsemi) dominates the military and aerospace FPGA markets. This is primarily due to their anti-fuse and flash based technology, which is more secure and radiation resistant than the more popular RAM based designs (such as Achronix, Altera and Xilinx). Many of Actel’s products are made by UMC, so clearly the Taiwanese are considered to be reasonable suppliers. Perhaps a more valid issue is that Intel’s fabs are located such that they are much less prone to natural disasters – but that has little to do with the United States itself and more to do with local geography.
Technically speaking, this partnership is a good fit. FPGAs are highly regular designs and amongst the first tape-outs at most foundries. This regularity is essential, since Intel’s co-optimization of design and process has highly restrictive design rules for manufacturability and yield. Achronix’s competitors are both using TSMC for new products, and the density advantage of a 22nm process could help level the playing field (and ease concerns over manufacturing related delays). The density is also essential for low power designs, since using more logic (e.g. area or density or both) is more power efficient than high frequency for FPGAs.
FPGAs are particularly interesting due to the economics of chip design. ASIC design costs are increasing due to complexity, which means that volumes (or prices) must also rise. For some lower volume applications, or those needing reconfigurability, FPGAs are a great alternative to ASICs. In many respects, they are a compromise between the programmability of a CPU and the performance of an ASIC. In fact, as FPGAs have grown in size and complexity, some include embedded processor cores, typically ARM or PPC variants.
Given this background, we can try to infer Intel’s motivations behind the announcement and what this implies for the future.