Intel, which squeezes toothpaste, also has a day of "killing crazy"?

Produced by | Tiger Sniff Technology Group

Author | Maruto Mountain

Edit | Wang Yipeng

Header | Intel

Two months ago, when a truck carrying parts for the world's first 0.55NA EUV lithography machine appeared at Intel's Oregon facility, the semiconductor industry began to discuss the possibility of Intel catching up with TSMC in chip process manufacturing in the future.

However, Intel has shown the world that it has ambitions that go far beyond that.

In the early morning of February 22, Beijing time, Intel held the first Intel Foundry Direct Connect conference.

Image source: Intel

In the semiconductor industry chain, chip foundry is an out-and-out "important and boring" link, but the highlights of this event are enough to surprise people by mentioning any one of them:

The Clearwater Forest Xeon processor using Intel's 18A process (1.8nm class) has been taped out, and the 1.4nm Intel 14A process has made its debut.

It is the first to launch the system-level foundry mode of AI chips, providing full-stack optimization from factory network to software.

It has successfully won Microsoft's chip order and will use Intel 18A foundry.

Foundry services IP, EDA partners, including Synopsys, Cadence, Siemens, Ansys, Lorentz, Keysight presentation tools and IP are ready.

It is worth mentioning that Intel also announced a cooperation with ARM at the event, and will support the provision of foundry services for SoCs (system-on-chips) based on ARM architecture in the future.

"It feels as weird as running iTuns on Windows. ”

In the past week, when the world was talking about Sora's revolutionary advances, Intel quietly came up with a complete solution to deal with the explosion of AI computing power demand, once again proving the former hegemon of the chip industry.

IDM 2.0，绝境逢生？

When Intel is mentioned, people usually habitually compare it with Nvidia or AMD, but in fact, Intel's development path in the semiconductor industry is completely different from the latter two, and one of the most essential differences is that Intel is one of the few chip manufacturers in the industry that adopts the IDM model.

The so-called IDM refers to the design, manufacturing, packaging and testing to the sale of its own brand IC.

Correspondingly, there are Fabless manufacturers such as Nvidia and AMD, who only do IC design of chips, and the chip manufacturing process is completely completed by the foundry. The other type is the Fundry manufacturers represented by TSMC and GF, who are responsible for the foundry and packaging of chips.

The advantage of this division of labor model is that Fabless manufacturers do not need to bear the cost of building production lines of billions of dollars, and Fonddry manufacturers do not need to maintain a large product development team, so they can focus on process improvement.

This clear division of labor and efficient upstream and downstream relationship in the industrial chain is a prerequisite for maintaining Moore's Law to survive today. Therefore, the chip manufacturers that are still using the IDM model are actually "non-mainstream" to a certain extent.

Intel's business model is more "evil".

The current CEO, Pat Kissinger, took office in 2021 and proposed the "IDM 2.0 model". That is, while insisting on producing chips by itself, it also provides foundry services for third-party chip design companies, and at the same time hands over some process chips to other foundries to supplement their own processes.

This idea sounds great, but even people who don't know anything about the chip industry can detect a clue: third-party chip design companies, whose products may already compete with Intel, so how can they rest assured that the chip manufacturing process will be handed over to the latter?

Intel received the world's first 0.55NA EUV lithography machine, image source: Intel

For example, it's like Apple saying to Microsoft, "We're upgrading MacOS, and if you need to, we can help you develop Windows together." ”

Another problem that cannot be ignored is that the chip foundry industry is a typical "heavy asset, long cycle" industry, if the construction of capacity that can undertake third-party chip design companies, it is bound to increase the upfront cost significantly, which has been reflected in many financial reports last year, for example, in the second quarter of 2023, Intel in the overall operating margin of 33%, its IFS department responsible for the chip foundry business has a profit margin of -28%, which seriously drags down the group's performance.

Therefore, after the concept of IDM 2.0 was proposed, the outside world has never stopped questioning.

And in June last year, after realizing that the IFS division had become a heavy burden, Intel chose to split the division and operate independently in the future, and calculate profit and loss separately in the financial report.

At the same time, the semiconductor industry has also ushered in a sea change.

With the surge in demand for computing power around the world, TSMC, a major manufacturer responsible for AI chip foundry, has obvious capacity bottlenecks, such as the monthly production capacity of CoWoS advanced packaging has been maintained at 1.3-15,000 pieces (wafers) so far, and some of this capacity will be divided by chips such as Apple M1 Ultra, so in the second half of 2023, the delivery cycle of NVIDIA high-performance computing cards will generally be extended to 12-16 months.

Considering that there are currently manufacturers with advanced processes, none other than TSMC, Intel, and Samsung Electronics, and Samsung Electronics' yield problem after the 5nm process node is still a mystery, Intel has become one of the few foundry options for chip design manufacturers.

This is also the confidence of Intel to propose "AI chip system-level foundry" at this time.

Splashing the sky and riches, can Intel catch it?

Although Intel seems to have the right time and place at the moment, does the company now have the strength to show its strength in the foundry industry? If you look at market share alone, according to TrendForce statistics, Intel accounts for only 1% of the global chip foundry industry.

However, in terms of technology, Intel is still a leader in the industry.

The first is that its process is currently in a state of steady progress. In July 2021, Intel announced a plan of "four years and five process nodes", that is, to use four years to promote the five process nodes of Intel 7, Intel 4, Intel 3, Intel 20A and Intel 18A, with the aim of regaining the process leadership of global chip foundry by 2025.

According to the information that has been made public so far, two Intel 3-based products, Sierra Forest and Granite Rapids, have completed the trial production work after tape-out and design certification respectively. The "Angstrom" Intel 20A and Intel 18A are also expected to go into production this year.

Intel process node diagram, image source: Intel

It is worth mentioning that these two process nodes will be equipped with perhaps two of Intel's most important technologies in the next decade: RibbonFET gate full surround FET, and PowerVia back power delivery network.

To put it simply, the former can further reduce the size of the transistors on the chip, while the latter reduces power consumption by transplanting the power and signal lines of the chip to the back of the wafer.

Together, these two underlying technologies form the technical foundation of Intel's foundry business and ensure that Moore's Law can still be continued.

Of course, there is little need for such a sophisticated technology in the industry at present, after all, the main chip that AI practitioners rely on, Nvidia A100, has a process of only 7 nanometers.

In the short term, the AI industry may need Intel's technical accumulation in the packaging process.

At this conference, Intel announced the inclusion of FCBGA 2D+ in Intel's foundry advanced system packaging and testing technology portfolio. The portfolio will include FCBGA 2D, FCBGA 2D+, EMIB, Foveros and Foveros Direct technologies.

This series of obscure technical terms allows Intel to meet the heterogeneous integration needs of different foundry customers, and also enables "chips" from different suppliers and different process nodes to work better together.

Here it is necessary to explain the concept of a "chip": it can be simply understood as a small modular chip. Its significance for the AI industry lies in the ability to connect CPUs, GPUs, and NPUs in the same system at high speeds, greatly improving the transmission rate between heterogeneous cores, and reducing data processing power consumption while improving data access speed.

Intel 2.5D packaging production line, image source: Intel

In addition, Intel is also preparing for potential startup customers such as Groq and SiMa.AI, which have been in the limelight recently. These companies generally believe that specialized chips are more efficient than general-purpose chips, so they also focus on building chips specifically for certain architectures, but the problem is that these startups have a variety of chip layout designs.

As a result, the "EDA Big Five" (Synopsys, Cadence, Siemens, Ansys, Lorentz) in the silicon industry will be present to demonstrate Intel's readiness for certification of various software and tools.

In general, whether it is process manufacturing, packaging technology, or development tools, Intel has initially had the ability of "system-level foundry of AI chips", but can Intel really sit back and relax since then?

Obviously not, one of the most direct hidden dangers is that although the chip foundry department will be independently accounted for in the financial report in the future, the pre-construction and commissioning costs of the production line still need to be paid by Intel.

It would be an astronomical sum. According to the current project progress officially disclosed by Intel, the company will invest a total of 62.6 billion US dollars in Poland, Israel and Germany in the future for the construction of wafer foundry production lines. With funding for the chip subsidy bill delayed, these investments would almost entirely need to be borne by Intel.

Another point that has to be mentioned is that Intel CEO Pat Gelsinger has repeatedly said that after independence, the chip foundry department and the chip design department will be completely separated, and the latter will evaluate different solutions in the market when there is a foundry demand, and choose the foundry based on the cost, which is absolutely fair competition.

Kissinger did. At present, the Compute Tile modules of Intel's two processors, codenamed "Arrow Lake" and "Lunar Lake", have been handed over to TSMC foundry.

But in the long run, if the core modules of Intel's own processors are all completed by TSMC foundry, how can other Fabless manufacturers rest assured that they will hand over orders to Intel? This is probably a problem that Intel's foundry department has to consider.