Does It Matter If TSMC Dethrones Intel and Takes the Semiconductor Lead?

If you’d told me five years ago that Intel was about to stumble into arguably the worst manufacturing problems of its existence, I would have been surprised. Our 2012 story on why Intel leads (led) the world in semiconductor manufacturing now looks more like a historical retrospective than a confident projection of future trends. The company’s slip — and the potential consequences of that slip — have driven a lot of analysis over the past few years, some of it written by yours truly.

But there’s another question underlying the 10nm problems that deserves attention. How much does it actually matter to the silicon industry as a whole, that Intel slipped on the 10nm node? Obviously, there have been real repercussions as far as the company’s ability to meet demand, and Intel would undoubtedly prefer to retain its reputation for being the most-advanced silicon manufacturer in the business. But the long-term impacts are surprisingly tricky to sort out. I was reminded of this by a recent Bloomberg story, titled “A Company Few Americans Know Is About to Dethrone Intel.” It’s about TSMC and its overall process node leadership, and it makes a lot of good points. Intel’s 10nm is expected to be roughly equivalent to TSMC’s 7nm, but TSMC has 7nm in-market already in mobile hardware and AMD will almost certainly ship 7nm CPUs before Intel starts moving 10nm in volume. It’s been nearly 20 years since AMD and Intel transitioned to the same node at the same time. From the original Athlon to the present day, AMD has never beaten Intel to a node. Add up the total R&D spending by all of TSMC’s customers (some of which is surely being spent at TSMC itself) and the sum is larger than what Intel itself spends on R&D.

Similarly, we still expect the shift to 10nm to have an impact on silicon performance as well. There’s a reason why AMD chose to limit its 14nm chips to 32 cores but will push up to 64 cores on 7nm. The firm projects a 25 percent performance improvement in the same power envelope or a 50 percent power consumption improvement at the same frequency. Even allowing for manufacturer fudging, that’s still a substantial boost. Similarly, we’ve seen larger GPU generational performance improvements when companies shift to new nodes as opposed to when they release new products on the same node. The jump from Maxwell to Pascal was larger than the improvement from Kepler to Maxwell, for example. So on the one hand, yes — semiconductor performance improvements matter, and given that 10nm will provide some of them, it matters to some extent. The real question is — how much?

All the Intervening Variables

On the other hand, Intel’s 10nm node is already years late. At IDF 2013, Intel talked about having chips in-market by 2015. Despite this, Intel’s business is positively booming. Data center growth is huge. The company may have famously missed mobile, but it has a 5G modem apparently well positioned for the advent of that communication standard, a growing IoT business, surging data center shipments, and plans for new 48-core Cascade Lake CPUs to ship next year with a new MCM design. In the Bloomberg article, Intel’s server division head, Navin Shenoy, argues that what customers care about isn’t process node, as such. “I’m confident that we’re going to deliver what our customers care about, which is system performance,” he said. This is a real — and important — point.

Efforts to push ARM into data centers have, to date, not enjoyed tremendous success. Amazon’s decision to build its own ARM cores is absolutely a warning sign that Intel could face long-term intensified competition from alternative CPU architectures, but the company has done a very good job fending off those challenges so far. Rory Read’s 2012 prediction that ARM would constitute 15 percent of the server market by now isn’t even within an order of magnitude of being true. The only vendor cutting into Intel’s data center share is AMD, and only by single digits. That’s still a strong result for AMD, but the near-term threat to Intel simply isn’t very large. Chipzilla has made it clear that cloud and data center is a major pillar of its business going forward, and these customers tend to be more conservative than the typical consumer market. They also tend to be less price sensitive.

It’s hard to say how all of this plays out in burgeoning markets like AI, machine learning, and self-driving cars. I think you can make a very fair argument that Moore’s law and Dennard scaling collectively functioned like a type of straitjacket. With an assumed 50-100 percent performance improvement arriving every few years, there was precious little room for alternative architectures or new approaches to computing problems. If you started in 1980 with a custom architecture that was 15x more efficient than CPUs of the day, but it took you until 1988 – 1990 to bring your product to market, chances were good that conventional processors had either cut your benefit to nothing or eaten into it to such a degree that your chip could never compete on price after economies of scale were taken into account. CPU clocks have only improved modestly since 2004, while Intel’s overall CPU performance has scaled modestly (outside of certain AVX-friendly workloads) since 2011. These two developments opened the door and the funding streams for alternate approaches to computing. But Intel has positioned itself to play in these spaces as well, with acquisitions like Movidius. It may have canceled its Knights Hill product, but the company’s GPU efforts will likely target the same space with a better, more competitive architecture.

Bloomberg’s observation that adding up the R&D budgets of all of TSMC’s customers results in more total spend than Intel seems rather banal to me and I’m not sure it has any predictive power at all. First, some of that R&D spending is going to be duplicative, reflecting things like mask and tool costs. Second, not every product TSMC builds even has an Intel counterpart. Third, the question of total R&D spending is less important than what a company is spending R&D dollars on. Fourth, R&D spend as a raw figure may not be a good predictor of final product performance no matter what. Plenty of people looked at AMD’s anemic R&D spend from 2012 – 2016 and concluded there was no chance of Ryzen being any kind of competitive product.

Intel has made process node leadership critical to its business reporting for many years, but at the Credit Suisse 22nd Annual TMT Conference, the company acknowledged that its own guidance is shifting. Here’s Intel acting CEO Bob Swan:

I think there’s a few dynamics about the company that we created and we asked you to judge us by them. And over time, if you stick with those, I think you get shortsighted. And we’ve talked about a couple. Moore’s Law leadership is all that matters. That’s not true. Product leadership is what matters. But we told your forever that Moore’s Law leadership matters. And now we’re trying to tell you, hey, we get paid on product. So, that’s a fundamental shift that I think it’s very important – it’s very important for us in terms of how we run the business and I believe increasingly important for you about how you make your decisions.

It might be tempting to dismiss this as Intel pivoting away from a metric simply because it isn’t winning in that category any longer, but the end of conventional Moore’s law scaling isn’t unique to Intel. It’s something the entire semiconductor industry is grappling with. But we’re also in a reality where no one is sure which AI/ML products will succeed, or how quickly alternative architectures will ramp, or how long it will be before autonomous vehicles are common. The path forward for conventional silicon scaling below 5nm is deeply uncertain.

It’s easy to say that sure, not being in a leadership position is going to cost Intel something. It’s much harder to predict exactly what that cost is going to be. If I personally had to guess, I’d bet on less disruption rather than more. With the gains from each new node smaller than the node before, the chances that missing a node will prove a major problem are lower as well. The factors that matter to future success are likely to be rooted in the types of products a company brings to market and how well those products are positioned to meet the needs of the new workloads we see rising by the day. Process node positioning/leadership is important to that question, but not the sole determinant.