The IEEE’s International Electron Devices Meeting is like the semiconductor industry’s analog to IBC or CineGear. There, Intel revealed its ten-year manufacturing road map.
First though, what went wrong? Intel has been THE leader in semiconductor manufacturing for years, so suddenly being behind TSMC and Samsung by an entire process generation is very unusual. In fact, it’s a first. On top of that, it’s renowned for its massive economies of scale; its volumes were so large that when AMD launched the Athlon64 and Opteron and IBM the PowerPC G5, it took a full year for IBM and AMD combined to match the number of 64-bit x86 parts that Intel shipped… in one month, and Intel didn’t even have to cut back on any other product lines to do it.
Now, it’s having trouble keeping up with demand, while AMD is chewing up the market at Intel’s expense.
CEO Bob Swan came clean about what went wrong at the Credit Suisse technology conference.
- Larger than expected demand for CPUs.
In other words, Intel’s prediction for CPU demand was too low, and it didn’t allocate enough capacity to keep up with demand.
- Overly aggressive approach to 10nm process node.
Intel’s engineers tried to achieve a 2.7 scaling factor in the transition to 7nm, instead of the 2.0 scaling factor that its competitors have been achieving, and that lead to the slip.
- Smartphone modems.
Intel used the same 14nm fabs that it uses for processors to fabricate smartphone modems, which consumed more of Intel’s fabrication capacity than expected, exacerbating the effects of the 10nm process slip.
As its 10nm process spins up at long last, Intel outlined its plans to return to a regular cadence of developing a new process, enhancing it, and then launching a new one alongside the second optimized iteration of the current.
In other words, we should expect 7nm in 2021, 7nm+ in 2022, and 7nm++ along with 5nm in 2023.
It’s expecting its scaling to enable it to match the gate density of TSMC’s already ramping 5nm process with its own 7nm process. Intel has been able to achieve that in the past, but whether it can continue is an open question that only time will tell. Intel’s 7nm Extreme Ultra-Violet process is scheduled to go into production in 2021, while TSMC and Samsung are both ramping up their 7nm EUV processes right now, with both AMD (TSMC) and nVidia (Samsung) among their clients.
The road map ends at the ten year mark with the 2nm++ node in 2029.
Additionally, Intel is going to emphasize diversity beyond processors. Bob Swan pointed out that while Intel had a 90% share in x86 processors, its share in semiconductors was only 30%, which is a lot more room to grow. Intel’s work on the Xe GPU and the Nervena neural net processor, as well as products like the Agilex FPGA, cell phone modems, and Optane memory all reflect Intel’s goal of improving diversity.
With AMD firing on all cylinders in the CPU market, Intel is essentially ceding victory to AMD, at least for this round. While it’s possible that Intel might retake the x86 CPU crown from AMD at some point in the future, it’s looking at new markets rather than focusing all of its efforts on competing with Zen.
Particularly interesting is the fact that Bob Swan is taking a page from AMD CEO Lisa Hsu’s book. Where Lisa Hsu chose to hold back its GPU business, even though that lead to losing its chief GPU architect, its focus on the CPU side of its business has been paying off in spades, with AMD not only leading in performance, but also being able to make enough CPUs to take a significant share of the market away from Intel.