Get email delivery of the Cadence blog featured here
I can't keep away from work. Saturday found me in the Cadence auditorium for the quarterly meeting of CASPA, the Chinese-American Semiconductor Professionals Association. Yes, I have noticed that I am not Chinese, but I once stepped in at the last minute to moderate a panel session on IP, and since then I seem to be an honorary member. This time the focus was on memory in general, memory in China in particular, and XMC as the company where China is pinning its hopes. The XMC part of the day will be covered tomorrow.
Tammy Kiely, Goldman Sachs
Tammy Kiely is a partner with Goldman Sachs in semiconductors, based in San Francisco. She titled her talk Migration of Memory Industry (from States, to Japan, Korea, then China). Memory started in the US (if you are old enough to remember, Intel, TI, and others all manufactured DRAM). It then migrated to Japan and now is largely in Korea with both SKI Hynix and Samsung being Korean. The migration is not complete: Micron is still in America/Japan and SanDisk is in Japan. One theme of the day is that it will migrate to China next.
Tammy started by pointing out that the one constant in the semiconductor industry is change. In 1985 NEC was #1 in semiconductors. Now it doesn't exist, a part of Renasas along with many other famous names from 30 years ago when Japan was riding high in semiconductors. In 1992, Intel became #1, where it has remained ever since. But #2 is Samsung, a Korean company. There has been a lot of consolidation, not just in Japan, with a result that now the top 15 companies control 2/3 of the industry.
Memory is also consolidated with three companies (Samsung, SK Hynix, and Micron) accounting for over 3/4 of the market compared to half 10 years ago.
Memory markets are measured in two ways: bit growth and dollar growth. One of the few constants in semiconductors is that there will be strong bit growth in memories, but unit prices will decline. Which of these trends dominate determines the health of the industry. Memory is big, an $80B industry. DRAM is in secular decline since prices are reducing faster than bit growth. NAND flash is growing faster than price decline. Everything else (such as NOR flash or E2) is characterized as legacy memory with no real future. Of course, there are also research technologies not yet in volume production.
3D NAND flash is where the bit growth is, which is why it is important for XMC. It has very high capital requirements, higher than any part of semiconductors except foundries (Tammy's rule: $1B for the shell plus $800M for every 10K wpm). There is actually a transition going on from 2D NAND to 3D NAND. As is often the case, in such transitions, having a lot invested in the old technology can be a liability and not an asset, so XMC starting in 3D without ever investing heavily in 2D is well-placed.
Handel Jones, IBS
Next was Handel Jones, who talked on Why China Should Have a Memory Supply?
Chinese companies shipped over half of smartphones and will be up to 60% this year. This is not companies like Foxconn manufacturing for Apple, this is Chinese phone companies such as Lenovo, Huawei, Xiaomi and a horde of smaller companies selling both into the Chinese domestic market and, for the large companies, internationally, too. China will also ship over 50% of ultra-high-definition TVs (I think that means 4K and above). In 5G basestations, Handel thinks that Huawei is already one to two years ahead of Ericsson and Nokia. They have 25,000 people working on 5G compared to 10,000 at Ericsson. This is an area where China is going from being a technology follower to a leader, driven by the fact that in 2025 to 2028 there will be over 1B 5G subscribers. But Huawei has a problem: they need powerful semiconductor technology and currently cannot get it domestically from China. For example, SMIC will have 14nm in 2020 working with imec, about six years behind Taiwan.
However, this is gradually changing and design starts are growing strongly in China (the light blue at the bottom of each bar).
In flash memory there is a transition going on. 2D flash suffers from rising costs and declining endurance (the number of writes that the memory can endure before failing). As a result 2D is going away and by 2020 the market will be 94% 3D (only 18% this year). Currently Samsung has 90% of the 3D NAND flash market, SK Hynix is thought to be losing money in it. The switch from 2D to 3D will catch existing companies with capital stranded in the wrong place. 2D fabs will be rapidly obsolete. This is a disruptive change.
Currently the volume is in 32-layer NAND flash, with 48-layer starting and 64-layer in a couple of years. Crosspoint (Intel/Micron) is more expensive and will probably compete more with DRAM than flash. Intel's own 3D NAND flash uses floating gate, not charge trap, which has less endurance.
Handel reckons a lot of the future growth will not be in standard products that are just half the market. The big growth will be in multi-chip modules especially for solid state disk (SDD), but also smartphones, datecenters, IoT. Cloud computing datacenters will still need hard disks (HDD), but they need the higher performance of 3D NAND for big data analytics. And by 2025, the big datacenters will largely be in China.
China is not just XMC. Samsung's main 3D NAND fab is in Xi'an, SK-Hynix has a DRAM fab in Wuxi, Intel is bulding out its Dalian fab to build 3D NAND and has packaging and test in Chengdu. Micron has assembly and test in Xi'an. In 2010, Chinese companies consumed 15% of semiconductor imports (the rest were re-exported). By 2025, it will be well over half. With China's determination to build up an internationally competitive domestic semiconductor market, there is a big opportunity. Samsung has money to invest, SK Hynix has SK Telecom behind them, but US companies won't make the investment on the scale required, such as China's $24B to XMC.
China's workforce is changing, too. In the next few years there will be a transition to being truly high technology, with an aging population but with 7M engineering graduates per year. This is a very different workforce from Foxconn, who employs a legion of unskilled and semiskilled workers whose alternatives include subsistence agriculture.
Previous: TI and UI: Texas Instruments' Experience with the Common User Interface
Next: Memory in China: XMC