China's 7nm chip surprise reveals more than Beijing might like
Should we be worried? Well, size isn't everything
Opinion After decades trailing the rest of the world in leading-edge chip making, Chinese sand stamper Semiconductor Manufacturing International Corporation (SMIC) has quietly got into the 7nm business. That's a huge and unexpected leap. Has the West's embargo of the latest fab furniture failed?
And 7nm-process silicon from China is a genuine shocker. However, whether it portends a seismic shift in the country's chip expertise that will upend world markets and Western assumptions is another matter.
Every detail in the report gives the real game away. The 7nm process is a copy of the N7 node TSMC put into mass production four years ago. China has had access to any number of ex-TSMC engineers and is spending infinite money to play chip catch-up, so the existence of a cloned fab still two cycles behind earns a "well done, I guess" by itself. As soon as you dig in, though, the true value of even this achievement starts to tarnish.
There are lots of interesting things about TSMC's N7 node, the most important being that it doesn't use the 13.5nm wavelength Extreme Ultraviolet (EUV) lithography process to print the patterns on wafers that make them into circuits.
EUV litho is both state of the art and embargoed by the West – the US successfully pressured the Netherlands government to block ASML — the only provider of EUV lithography machines — from selling them to China. So, instead, N7 uses the previous generation 193nm Deep Ultraviolet (DUV) process – a breathtaking achievement, but deeply consequential.
Maybe EUV underestimated how hard this is...
DUV lithography is old, dating back to the 1980s. For many years, it was not a limiting factor in the ever-shrinking chip geometries that powered the heyday of Moore's Law. But as the 2000s approached and with them the transition from 130nm to 65nm process chips, the fundamental problem with DUV became more pressing. Electromagnetic theory suggested it wasn't practical to make things on a chip that was smaller than the wavelength of light used to make them. But that's OK, boffins reasoned, they would just move to a smaller wavelength. The industry waited to move to EUV and its instant 14x leap in tinyness.
And it continued to wait. EUV was much harder than anyone expected. It ended up arriving 20 years late, throwing grit into the vaseline for everyone. In its absence, enormous amounts of ingenuity was poured into perverting the laws of physics as far as they could go. TSMC's N7 was the pinnacle of that process, an amazing, tottering pile of near-magic photon-wrangling audacity, a melange of multiple overlapping mask exposures and complex production tricks. That it could work at all was amazing, that it could do good yields? Unbelievable.
For a brief moment, this was TSMC's jewel in the crown. It was also DUV's last chance to shine. As N7 hit the market, EUV was finally ready.
All the weird tricks to make DUV operate far above its pay grade had broken many chipmaking rules. It used to be that once you got your new fab running, that was it. You were ready to cash in your Moore's Law bonus.
Intel's famous tick-tock cadence relied on this. A new node first got a proven chip design – tick – to be re-engineered later. Tock. Two helpings of spicy performance, but the risks of changing too much are nicely limited.
TSMC's N7 is as far from those halcyon days as you can get. Each different design needs a lot of work to get from circuit to chip. You can't take an old design and slap it in, not at the complete design level, not at the functional block level, not at the transistor level. Everything needs to be reinvented, the tools, the techniques, the lot. Every company in the game had to pick its own path through these problems while EUV sulked in its tent – or not. It's no coincidence Intel's tick-tock wound down as DUV got increasingly complex.
Another broken rule was that if you could make one logic design in a process, you were guaranteed to be able to make anything else on it, almost by default. Like doing a jigsaw. N7 is more like painting in oils. What China has looks very much like a one-off proof of concept that they can produce one particular chip in 7nm, not any chip in 7nm. It's also an ASIC crypto-miner, highly parallel, low complexity. A starter chip. We're not talking an Apple M2 SoC.
- Infosec not your job but your responsibility? How to be smarter than the average bear
- Engineers on the brink of extinction threaten entire tech ecosystems
- TSMC and China: Mutually assured destruction now measured in nanometers, not megatons
- 114 billion transistors, one big meh. Apple's M1 Ultra wake-up call
Furthermore, that chip has been in production for a year. If you have a commercial 7nm line in full swing for a year during a huge global chip shortage, you go to market as soon as humanly possible. Perhaps SMIC is fully committed to its internal market of OEMs, government and military demands, but the Chinese state loves to flaunt any ability to confound expectations and side-step embargos. Silence can be very suggestive.
The key to competitive chipmaking is scale. Once you have a commercial opening, you scale up as quickly as you can to take market in what can be a very short window. China's half-inched TSMC process cannot scale because, most likely, it doesn't have all the specialist expertise and tools that turn the process into products. It also doesn't have a window of opportunity. It doesn't even have a staging post from which it can continue to erode the US and Taiwan's multi-year lead.
Duplicating the TSMC N7 process must have been hard, expensive and high priority. But it leads nowhere. What we see when we pop that chip open is not a scarfy future, but a tiny white elephant bathed in deep ultraviolet.
China is learning to build the world's finest propeller engine, just as its competitors enter the jet age. Admire the effort, but don't cash in your chips just yet. ®