IBM recently disclosed a working 7nm test chip generating a lot of excitement in the semiconductor industry and also in the mainstream media. In this article I wanted to explore the 14nm, 10nm and 7nm nodes, the status of the key competitors at each node and what it may mean for the companies.
For many years Intel had been able to claim a process lead over everyone else in the logic segment of the semiconductor industry. In 2007 Intel introduced high-k metal gates (HKMG) at the 45nm node and it wasn’t until around 2010 that the rest of the industry introduced HKMG. In 2011 Intel introduced FinFETs at 22nm and once again it wasn’t until 2014 before the rest of the industry began to implement FinFETs. At the 14nm node it would have been reasonable to expect that Intel would have a smoother path to 14nm since they would just be shrinking FinFETs while the foundries would be introducing FinFETs and simultaneously executing a shrink. But at 14nm what has actually happened is that Intel stumbled badly. With 45nm, 32nm and 22nm introduced by Intel in 2007, 2009 and 2011 respectively, 14nm was expected in 2013, in reality it wasn’t until 2014 that Intel introduced 14nm, several quarters late.
Meanwhile Samsung skipped 20nm as a foundry node and drove to 14nm in late 2014 just slightly after Intel. Samsung’s initial process is called 14LPE and they have a follow on improved performance 14LPP process. Samsung has also licensed their 14nm process to Global Foundries creating a worldwide network of 14nm production fabs.
TSMC’s “equivalent” process is called 16nm and TSMC introduced their 16nm FinFET process (16FF) in late 2014 ramping into production in 2015 followed by an improved 16FF+ version ramping in the second half of 2015.
Now there are several comments that need to be made about this. Intel’s introduction dates are for their microprocessor process (MPU) and the foundry introduction dates are for System On a Chip (SOC) processes. An SOC process is generally more complex than an MPU process due to the need to support a wider variety of device and performance requirements. Intel typically introduces their SOC processes approximately one year after their MPU process. Secondly, the foundries have kept 20nm Back End Of Line (BEOL) interconnect pitches on their 14nm process and only shrunk the Front End Of Line (FEOL) transistor pitches. Intel shrunk both the FEOL and BEOL and did it with smaller pitches than the foundries.
Clearly the transition to 14nm technology has gone more smoothly at the foundries but what has this actually meant for the companies.
At 28nm TSMC was the clear foundry leader and that enabled them to capture Apple’s applications processors business from Samsung. At 14nm the fate of the Apple business is still playing out but appears to be split between Samsung and TSMC with Samsung getting the early designs. Qualcomm has also moved some products to Samsung from TSMC. Clearly Samsung’s early ramp of 14nm has helped them regain business from TSMC.
For Intel the impact of their late delivery of 14nm is less clear. Intel has very little foundry business and it really has no material impact on the company, at least to-date. Intel’s main business is microprocessors. For years Intel was locked in a battle with AMD and ultimately Intel’s process technology and manufacturing advantage forced AMD to sell off their fabs. Today Intel enjoys a near monopoly in the PC microprocessor business. It isn’t clear to me that being late on 14nm impacted that business in a material way, at least not from a market perspective. Being late likely had a financial impact but Intel enjoyed strong gross margins throughout 2014 so clearly even the financial impact wasn’t material.
According to Intel’s introduction dates outlined above 10nm should have been introduced in 2015 but is now looking like late 2016 to early 2017. Similar to 14nm Intel appears to be slipping badly.
Samsung has recently shown a 10nm wafer and is saying they will ramp 10nm in early 2017.
TSMC also has working 10nm silicon and has said they will start risk production on 10nm in Q2 of 2016 and full production late 2016. Basically the three companies are all likely to be neck and neck for 10nm production.
As we saw at 14nm the success of the early 10nm ramp at Samsung and TSMC will likely have a big impact on who gains Apple’s business. Both Samsung and TSMC say 10nm development is going really well. Global Foundries 10nm plans are less clear. They are reportedly developing their own 10nm process internally, they have also recently bought IBM’s semiconductor business and once that deal closes will have full access to IBM’s engineering resources.
For Intel the timing impact is once again harder to gage. Intel recently bought Altera and if Intel is going to fabricate Altera’s products then an early 10nm ramp will be important for that business. How much jeopardy Intel’s microprocessor business will be in from a slip of a few quarters is less clear. Certainly there is a financial impact but the product impact is less clear.
IBM recently disclosed they have developed a 7nm test chip in concert with their partners New York State, Global Foundries, Samsung and “equipment companies”. Among the details were silicon-germanium (SiGe) FinFET channels, EUV and a 30nm pitch. It was said in the articles that this is the only working 7nm silicon today.
The fabrication of a working 7nm test chip is certainly a significant accomplishment and if we are going to see 7nm production in the 2017/2018 time frame now is when you would expect these kinds of announcements. In fact TSMC has promised 7nm production in 2017 aiming to introduce two nodes in two years for the first time in history. The SiGe channel and 30nm pitch are both also in-line with what would be expected although based on IBM’s pitch progression I actually expected a 26nm active pitch. The successful use of EUV is also important but the throughput issues around EUV tools, pellicles and other issues still need to be solved for this to be viable in production.
Without knowing what is currently going on inside TSMC and Intel on 7nm development is hard to tell whether IBM is really ahead or just announced results first. The existence of a working test chip is certainly good news for the industry and watching this races play out over the next few years will be fascinating.
For Samsung and TSMC huge volumes of business from Apple, Qualcomm, NVIDIA and Xilinx hang in the balance with each new process node ramp. The two companies have now moved to a dead heat with Intel in new process introduction and are even ahead when it comes to SOC processes. Intel’s recent process development missteps haven’t had a significant impact on Intel yet, but over time could erode confidence in the company and embolden competitors.