This is Intel's new strategy: transforming its divisions into (more or less) autonomous companies. On February 21, the founder announced that this year he would put his factories under the supervision of a new legal entity called Intel Foundry. This weekend, the company confirmed that manufacturing its FPGAs is now the business of Altera – the same old trade name for that line of chips.
The CEO of this new company, as announced in October, is Sandra Rivera, the former Xeons boss at Intel. In both cases, the parent company Intel remains the majority shareholder.
Altera was the original name of the FPGA developer that Intel acquired in 2015. Since then, this activity has been continued at Intel in the PSG (Programmable Solutions Group) division. Altera's competitors include Xilinx, the FPGA development company that AMD bought in 2020.
As a result of this reorganization, the catalog of FPGA models produced by Altera will be clarified. At the entry level, the Agilex 3 succeeds the Cyclone and develops controller cards that can accelerate data transfers on the local network. According to Intel, this is the most promising market segment, as tens of thousands of startups or manufacturers are likely to develop a product based on Agilex 3.
In the mid-range, the Agilex 5 replaces the Arria in the design of accelerator cards in data processing. Apparently Intel wants to believe that a market for AI accelerator cards will emerge based on the Agilex 5. For this purpose, it already offers a kit that allows the implementation of the algorithms in the Agilex circuits. Most popular machine learning tools, including TensorFlow.
At the higher end, Agilex 7 and 9 are intended more for massive data conversion, in large hyperscaler data centers where they are energy efficient (the former version), and in military and aerospace equipment (the latter).
We'll see some marketing consistency with Intel's Core x86 processors, which are also numbered 3, 5, 7 and 9 depending on range.
A technical segmentation of the Intel group
Technically speaking, the division of Intel into different autonomous units follows a certain logic. Parent company Intel continues to develop x86 designs by combining them with different functional circuits (GPU, integrated HBM memory, etc.). Currently, some of these circuits are manufactured in TSMC's Taiwanese factories.
Intel Foundry, in turn, aims to revitalize the semiconductor industry sector in the West with factories in the US and Europe. This company wants to follow TSMC by filling the order books with the production of all types of chips. In particular, it is intended to become the industrial home of RISC-V processors, which are expected to compete with ARM processors with even fewer energy-consuming cores and, above all, license-free.
Eventually, Altera will develop FPGAs, chips that can be reprogrammed to produce small-scale accelerator cards (prototype ASICs manufactured on a large scale). And these FPGAs systematically have ARM cores to run generic code. For this reason, certain models are also manufactured in TSMC or Samsung factories.
Had the three units remained together, Intel could have been criticized for no longer being entirely clear about which processor technology it wanted to favor. By dividing the three technologies x86, RISC-V and ARM into three “legally autonomous” companies, the question becomes less of a question. And each company has free rein to generate revenue, even if it means stepping on the strategic toes of the other two.
And financial segmentation
Because that is clearly the other goal of this division of the group into “legally autonomous” units: the separation of accounts. Intel suggests that Altera may soon go public to increase its capital without affecting its x86 processor development or that of its factories.
It can be assumed that Intel Foundry faces a similar fate. The challenge is to find funding by presenting more targeted activities that more accurately serve the interests of the different groups likely to want to invest.
“The goal of the factory team is to fill the factories. And to deliver components to as many customers as possible around the world. We want these to be factories that serve the entire market,” Intel CEO Pat Gelsinger said as he announced the upcoming replacement of the IFS division with the standalone Intel Foundry.
In this case, revenue from this activity doubled between 2022 and 2023, thanks to the leasing of production lines to other semiconductor companies: the Israeli Tower Semiconductor and the Taiwanese UMC. And it's a shame if they just increase their production of chips based on old technologies in 65 and 12 nm while the money keeps coming in.
“This resulted in annual Intel fab revenue of $952 million in 2023. This is remarkable progress. But that’s still a far cry from TSMC’s $69 billion a year or even Samsung’s $51 billion semiconductor factories,” comments Taipei-based financial analyst Dan Nystedt. “To this day, we still don't know where Intel will get the tens of billions of dollars it still needs to invest so that its factories can catch up. »
The financial issue is extremely tense. Intel has already been forced to delay the opening of its new $20 billion chip manufacturing plant in Ohio from 2025 to 2026. In part due to a longer-than-expected delay in receiving billions of dollars in grants under the Chips and Science Act of 2022. This law fleshes out the U.S. government's award of nearly $53 billion in public aid to increase local semiconductor production . Unfortunately, part of this money goes to TSMC to establish itself in the USA.
Intel is also looking for funds in Europe, where the EU is implementing its own chips law. On the one hand, Intel has integrated the joint financing company Chips JU and, on the other hand, receives individual financing from various European countries in order to establish itself in their territory. In particular, he promises Germany to revive its industrial activity and France to establish design centers there.
Exit from the vicious spiral of loss of market share
At Intel, the uncompromising search for convincing arguments sometimes borders on the limits of reasonableness. The SPEC organization recently accused it of manipulating the performance data of its previous generation of Xeon processors.
In a different genre, Intel now names its successive advancements with numbers reversed: the Intel 7, Intel 4 and Intel 3 engravings follow one another this year before switching to a different numbering: 20A, 18A and 14A.
Intel assumes that these are nanometers and angstroms (a unit of measurement ten times smaller than the nanometer), following press releases from TSMC, which was the first to know how to engrave chips with an engraving precision of 7. 5 and now 3 nanometers. In fact, the processors that currently benefit from the Intel 7 are engraved with a precision of 10 nm.
It must be said that Intel is doing its best to avoid a vicious spiral of market share loss. It started in the early 2010s when Intel didn't want to believe in ARM processors and actually left mobile chip manufacturing to Asian factories.
This activity has incredibly enriched players like TSMC, now the world's number 1 semiconductor manufacturing company. So much so that TSMC was able to modernize its factories much faster than Intel's, allowing its competitors to offer the production of more advanced processors than Intel Core and Xeon, which still dominated the PC and server markets.
The two most telling examples are for Intel losing the market for Apple machines – which now develops its own processors – and for AMD gaining a 23.1% market share in servers in 2023, while its market share was only 20.1% lies on PCs share. While AMD had all but disappeared from the x86 chip market by the end of the 2010s.