21. Dezember 2018, 10:51 Uhr | Frank Riemenschneider
Four years after the big announcement, Intel's foundry business has apparently come to an end.
With its Custom Foundry program, Intel wanted to make its leading edge chip production attractive to other manufacturers as well. At the IEDM conference in San Francisco, Intel has now indicated that it will close the business, even though official confirmation is still pending.
The reasons for the decision can currently only be speculated about. Industry insiders speak of too complicated tools and design rules with the result that the process from chip design to silicon takes too long. In addition, there are the eternal delays in the introduction of the 10 nm process, which was advertised years ago and was originally also intended to be made available to business partners. TSMC is now in mass production with 7nm chips for Apple and has caught up with Intel's former 2-year lead over Samsung and the Foundries.
From the outset, industry experts opposed the opening of Intel's state-of-the-art production facilities to the fabless community, as it would distract from Intel's core competency of microprocessor manufacturing. It's not a new realization that a functioning ecosystem is critical to success in the foundry business, it takes time, money and close technical cooperation with customers to build it, three things that Intel obviously underestimated.
A single major customer for 14 nm was Altera - the FPGA manufacturer that Intel later bought up and now manufactures for itself. TSMC, which had previously manufactured chips for Altera, was deeply touched at the time and learned its lessons. It was TSMC's founder and chip legend Dr. Morris Chang himself who made the announcement about Altera's loss. It was mentioned that TSMC saw this as a learning experience and would ensure that the loss of a close partner like Altera would never happen again.
There was indeed a very close relationship between Altera and TSMC until Xilinx came to TSMC at 28 nm, as Xilinx's previous contract manufacturer, UMC, could not follow Moore's law at the same speed as its great foundry rival, TSMC. TSMC treated Xilinx just as well as its long-time customer Altera from the start, which was detrimental to the Altera relationship.
Rumor has it that Altera edited the first version of Intel's 14nm design rules, which is absolutely uncommon in the foundry business. After many delays, Intel set up its own implementation team for the first 14nm Altera tapeout, which undoubtedly resulted in a very competitive FPGA chip. Without the prolonged delays, Xilinx would certainly have had major problems because the Intel 14nm FPGA significantly outperformed TSMC's 16nm Xilinx chip in both density and performance of the transistors. TSMC's first FinFET chip was basically only a 20nm chip where the planar transistors were replaced by FinFETs, but the geometries such as gate length etc. remained unchanged. Intel's 14 nm process was a real shrinkage compared to the 20/22 nm generation.
Where Intel's custom foundry business would be today if there hadn't been the ongoing problems and delays is speculative. One thing is for sure, many broken promises like the availability of the 10 nm process (the mass production of the own CPUs will probably start in Q2 or Q3 2019) have apparently led to the fact that the foundry business has come to a standstill and will soon be officially closed.
After all, there is good news for Intel's bread-and-butter business, namely building its own CPUs. Existing factories in Israel, Ireland and the US state of Oregon are being expanded. The three factories are currently primarily designed for 14 nm chips. Fab 42 in the US state of Arizona will be equipped in the next few weeks and could manufacture its first products from 2019. The factories in New Mexico will be designed for Flash and 3D XPoint. Intel also confirms, however, that in the future even more contract manufacturers will be used "where it makes sense". For example, TSMC could manufacture chipsets in order to relieve the Intel factories' fully utilized production capacity.