Intel posted higher than expected revenue when it announced its third-quarter results on Thursday, making it it’s best quarter ever. The company took its time to highlight some of the milestones it has reached and also discussed its process node efforts. Intel says that it wants to recapture process leadership and go back to a two-year development cadence.
In the data center, Intel shipped 23 million Xeon Scalable processors since its launch nine quarters ago. Several customers have selected the new Cascade Lake Advanced Performance (AP) 9200 series that consists of two 28-core die, while major cloud customers have launched Cascade Lake-SP instances. Intel also announced a collaboration with Oracle to include the new Optane DC Persistent Memory into its Exadata platform.
Furthermore, Intel claims it is now number one in market share in the silicon networking infrastructure market, as customers are increasingly adopting software-defined networking (SDN) based on x86 architectures for 3G and 4G infrastructure. Intel expects over $5 billion in revenue this year (up 12% from last year) and expects to reach 40% market share in 2022 with strong 5G adoption.
In the PC group, 18 Ice Lake systems have launched already, Intel claims, with 30 total in 2019. The first of the new Project Athena laptops are also available. In the high-end desktop segment, Intel announced the Cascade Lake Xeon-W and Core X-series. Moreover, Intel achieved power-on exit of its first discrete GPU. This means that the initial validation testing of Intel’s first discrete graphics has been carried out successfully. PCGamesN reports that developer kits are out in the wild. Based on the Gen 12LP architecture, it could be a mobile GPU, but specs such as execution unit (EU) count remain to be seen.
In IoT, the Santa Clara company has shipped 12 million Mobileye EyeQ SoCs this year, up 40% from last year. Mobileye also secured six major design wins totaling 10 million lifetime units. The IoT Group (excluding Mobileye) achieved its first one billion revenue quarter.
Intel also estimates that AI is driving $3.5 billion in revenue across all of its business segments, up over 20% from last year, and it has shipped the first Agilex FPGAs. Lastly, Intel expects both the sale of its 5G modem smartphone business and the sale of the IMFT fab to Micron to close in the fourth quarter.
On the process side, Intel has three investments going on with 14nm, 10nm, and 7nm. Intel increased its capital expenditures forecast for 2019 by $500 million to $16.0 billion "as a result of increased 10-nanometer and 7-nanometer investments."
On 14nm, when demand exceeded expectations in its record 2018 year, Intel started experiencing supply constraints. The company chose to prioritize the more valuable Core and Xeon processors, creating supply constraints in particular at the lower end of the market, where the company said it had constrained its customers’ growth. The company invested over $1 billion in additional 14nm capacity in 2018 and reshuffled some 10nm capacity back to 14nm.
While Intel is now projecting 2019 to be a relatively flat year, revenue in the third quarter exceeded expectations by $1.2 billion, and reports of supply constraints again started surfacing. Intel confirmed that it no longer expects to reach supply-demand balance in the second half of the year, which means the shortage will continue.
Intel has invested in record levels of capital over the last two years totaling over $30 billion, which consists of usual production capacity as well as additional capacity investments. Intel says it has increased its 14nm capacity by 25% this year while also ramping 10nm, and expects to increase capacity by another 25% in 2020.
On the PC side, Intel said its PC client supply will be up double-digits in the second half compared to the first half of the year and expects another increase in PC supply of up to high single-digits in 2020, both to meet demand and to increase inventory. The company did not say how those PC supply numbers reconcile with the two 25% capacity increases.
Intel CEO Bob Swan cautioned: "But that growth hasn't been sufficient. We're letting our customers down. And they're expecting more from us."
Intel expects to continue to be supply challenged in the fourth quarter.
After years of problems, Intel is now bullish on its 10nm process technology. Ice Lake has hit the shelves as Intel promised in 2018 (albeit after several revisions of the projection), and Intel shipped the first Agilex FPGAs in the third quarter.
In 2020, Intel will increase its 10nm portfolio with the already announced Spring Hill AI inference accelerator, Snow Ridge 5G base station SoC, and Ice Lake-SP for servers, which is rumored to feature 38 cores and 64 PCIe 4.0 lanes.
On the client side, launches include the hybrid architecture Lakefield, the Atom Elkhart Lake and Skyhawk Lake, and the Core series Tiger Lake. But also Intel’s first discrete GPU, which Intel disclosed is called DG1. As mentioned earlier, Intel achieved power on exit for DG1. "An important milestone," CEO Bob Swan said.
The overarching problem that prevented Intel from ramping 10nm is yield, but Intel now says that yield is progressing ahead of expectations for both client and server. "We feel very good about the capacity we put in place, the products we have coming down the pipeline, and the yields that we're achieving, almost week-on-week improvement over the last six months. So for 10, we feel really good," Bob Swan expressed.
Back in April 2018, Intel had echoed following sentiment about 10nm: "We continue to make progress on our 10-nanometer process. We are shipping in low volume, and yields are improving, though the rate of improvement is slower than we anticipated. As a result, volume production is moving from the second half of 2018 into 2019. We understand the yield issues and have defined improvements for them, but they will take time to implement and qualify."
With 10nm, per Intel, getting into better shape, Intel is expanding 10nm across its factory network. Intel is in volume production of 10nm in Oregon and Israel and announced that it would start production in Arizona soon. This leaves just Ireland without 10nm production.
While many have said that Intel wants to skip 10nm as much as possible, Swan expressed optimism: "The Intel 10-nanometer product era has begun, and our new 10th Gen Core Ice Lake processors are leading the way."
Reclaiming Process Leadership
After the tumultuous 10nm transition, Intel is now bullish on its ability to move back to a regular Moore’s Law cadence for at least the next few process technologies.
Bob Swan explained: "[W]e are accelerating the pace of process node introductions and moving back to a two to two-and-a-half year cadence. Our process technology and design engineering teams are working closely to ease process design complexity and balance schedule, performance, power, and cost. We are on track to launch our first 7-nanometer based products, a data center focused discrete GPU in 2021, two years after the launch of 10-nanometer." Bob Swan further clarified that it would launch in the fourth quarter of 2021.
While it might seem doubtful that 10nm would be a one-time hiccup and the company can regain its ability to deliver on a regular cadence at 7nm, there are some arguments in favor of it. Intel says that it has learned several lessons from 10nm. It has reduced the density scaling target and the introduction of EUV should be beneficial to yield learning, as Intel has said that multiple patterning has been one of the big challenges for 10nm. It has also allowed Intel to vastly simplify the 7nm design rules.
Bob Swan said that it expects to maintain that schedule “at least for the next few nodes”, and already expressed his confidence in 5nm (which should hence launch in early 2024): "We are also well down the engineering path on 5-nanometer."
Or as he put it: "So, we're investing to recapture process leadership going forward."
With TSMC seemingly on a three-year cadence for 3nm, there might be some merit in that statement.
With Intel squarely stating it wants to go back to a two year cadence, that implies that Moore’s Law will live on for a few nodes longer. At least if Intel executes to plan.