Despite billion-dollar investments in buried markets such as the smartphone business, a CEO relieved for a sex affair and massive problems with the 10-nm manufacturing process, Intel continues to earn billions of dollars year over year. Especially for processors for data centers there are high margins to earn, a market share of 90+X percent does the rest to keep up the chip prices and to keep Intel’s balance sheet healthy.
In recent years, AMD, the second US based x86 processor manufacturer, has been making a name for itself through a shrinking operation, keeping its head above water with low-margin SoCs for consumer products such as game consoles and desktop and laptop devices. Some of the balance sheets read like documents from the horror cabinet.
However, Intel AMD has now given a penalty without a goalkeeper with its production problems.
AMD will take 10% of the total server CPU market with EPYC Rome 7 nm processors by 2020.
An analyst report suggests that Intel's server processor market share is expected to fall below 90 % by the end of 2020, meaning that AMD would be able to capture at least 10 % of the market by then. This is because AMD's EPYC processors continue to receive many orders from server vendors and cloud service providers.
Demand for AMD-based servers is increasing due to the strong price/performance ratio of the EPYC series and AMD's plan to launch the next generation of 7 nm data center processors codenamed Rome at the end of 2019.
Cloud providers like Amazon are continuously investing more in the EPYC platform than in Intel Xeon CPUs. Amazon recently announced new Amazon EC2 instances with AMD EPYC 7000 series custom processors. With AMD EPYC CPUs, these M5 and R5 instances run with higher CPU performance (2.5 GHz SKUs) and feature low-latency NVMe memory subsystems.
AMD-based instances provide additional options for customers and can be the better solution for many workloads that do not fully utilize computing resources. By optimizing the balance between compute resources and utilization, these instances offer 10 % lower costs than comparable instances.
In addition to Amazon, the Japanese company NTT Data will also equip data centers with AMD EPYC processors. There is also the upcoming Atos BullSequana XH2000 supercomputer, which will use a total of 3125 AMD EPYC Rome 7 nm processors with around 200.000 cores and 400.000 CPU threads and is expected to be ready by 2020.
In fact, given the traction AMD EPYC CPUs receive, the 10 % figure is very conservative, as we can expect an even higher server market share for AMD CPUs by 2020. AMD's market share in the fourth quarter of 2018 was 3.2% , an enormous increase from 0.8 % market share in 2017 and zero before EPYC's release.
It was expected by analysts that AMD would capture around 5 % market share with EPYC Rome in 2019, but it looks as if they will now get even more given the popularity of their EPYC product line. One of the main reasons is the very competitive price-performance ratio. AMD's current line of EPYC CPUs is very competitive in every single version and offers better performance and efficiency than the competing Intel Xeon line.
The 10 nm manufacturing deasaster
The fact that it could come to that is due to the fact that Intel has lost its historical lead of at least 2 years in chip production compared to all foundries. Instead, Samsung and TSMC have been in mass production with 7 nm since 2018, while Intel continues to fight with 10 nm. All figures are, of course, purely marketing-based; from the actual geometries, the 7 nm processes of the foundries are comparable to Intel's 10 nm process.
Intel admitted its failure at the annual Investor Meeting. "It's no secret that we tripped at 10 nm," said Murthy Renduchintala, Chief Engineering Officer at Intel in Santa Clara. But now the problems are finally solved and series production of the chips with the now second generation of the process has started. From June 2019 Intel wants to deliver the Ice Lake U for Ultrabooks to partners. This will be followed by a third 10 nm node, then a change to 7 nm with EUV, and in 2021 the first graphics chips for servers and supercomputers will be released.
Originally Intel had already announced the 10 nm process P1274 in 2013, mass production was planned for 2016. The problem was that Intel wanted too much at once. First, unlike TSMC and Samsung, no EUV was used at all, some layers have to be exposed four times (quadruple patterning). New technologies such as Contact over Active Gate (COAG) and Single Dummy Gates for the transistors were added, and last but not least some materials had to be changed, so various metal layers had to be provided with cobalt instead of tungsten or copper.
This ambitious plan failed. Normally a chip yield of 50-60 % is assumed at the start of a new process, at Intel it should have been 5-10 % according to my sources.
As a result, Intel had to constantly further optimize its 14 nm process P1272 as 14+ nm and 14++ nm: With Broadwell, Skylake, Kaby Lake, Kaby Lake Refresh and Whiskey Lake, there are now five different versions.
In order to be able to delight its customers with a significant increase in computing power, the microarchitecture and clock frequency had to be used. However, these efforts were comparatively successful; more than 20 percent as an increase in performance with the same TDP are not from bad parents. However, it must also be noted that, on the one hand, Arm achieves similar growth with its annual Cortex-A updates and Intel's clock frequency increase is at the expense of the chip area: Thus, unlike 14 and 14+, 14++ has a transistor gate pitch of 84 nm instead of 70 nm in order to achieve higher clock rates of up to 5 GHz.
For the second version of 10 nm, called 10+ nm, Intel made some changes. The manufacturer did not give details at the investor meeting, unconfirmed reports assume that some pitches were more relaxed and COAG had been cancelled.
Light at the end of the tunnel
However, there is also good news from Santa Clara: According to Intel, series production of Ice Lake U has started, and from June 2019 partners for ultrabooks are to be supplied. ICL-U uses four CPU cores with the new Sunny-Cove microarchitecture including LPDDR4 support and the also new Gen11 technology for the integrated graphics unit. The 10 nm generation called P1276 will be used as 10 nm, 10+ nm and 10++ nm until 2021. The server CPUs (Ice Lake SP) should finally appear in the first half of 2020.
Also in 2021 Intel wants to bring first products manufactured in a 7 nm process (internal P1278) - then with EUV. The experience of 14 nm and 10 nm has been incorporated, and the design rules have been greatly relaxed to simplify development. 7 nm is to be followed by 7+ nm and 7++ nm, the roadmap for this node extends to 2023.
Intel also announced figures for Ice Lake and Tiger Lake in Santa Clara: ICL-U is said to achieve twice the graphics performance of a Whiskey Lake (WHL-U) at 15 watts in 3DMark 11, TGL-U even four times. However, instead of 15 W for Tiger Lake, one slide also calls 25 W TDP, which seems realistic in view of the fourfold computing power.
According to my information, there are currently more project enquiries than AMD can support with personnel. But you obviously don't make the mistake of signing contracts like crazy whose design ins you can't support, which is good. In Silicon Valley it is reported that AMD hires "like crazy" people, but in the end the number of available specialists with the required qualifications is limited.
So the potential growth of the company in Intel's living room will ultimately depend on how many new employees you can hire and support, and over what periods of time. But better than in 2012, when after horrendous losses 15 % of employees have been layed off, 50% of the Santa Clara headquarters were sold and further restructuring measures were implemented.