Intel's ex-chief Gelsinger comments on controversy over 18A yield
The "18A" production process is Intel's greatest hope. A media report about it caused confusion, which the ex-chief executive is now putting into perspective.
Hundreds of chips fit on one wafer – here with Intel's Meteor Lake generation.
(Image: c't)
"18A has only 10 percent yield" – This or something similar has appeared in many reports over the past few days. This is not a shortened version; the controversial statement about Intel's upcoming chip production process really only consists of this one figure. It can be found in a media report from South Korea, without any further classification or a reasonably concrete source.
The figure was published by the online edition of the daily newspaper Chosun Ilbo, which was founded over 100 years ago and had the highest circulation in South Korea for decades. It is still considered one of the country's leading media today. The 10 percent figure can be found in an article about the dismissal of Pat Gelsinger from Intel, who was CEO of the company until the beginning of this week.
Percentage figure from August 2024
According to several machine translations, reproduced here as an example from DeepL, it says: "The 18-Angstrom process, which Intel planned to mass-produce next year, had a yield of less than 10%, according to industry analysts. As a result, the customer Broadcom canceled its orders for Intel semiconductors." This reference to Broadcom alone makes you sit up and take notice, as the company had already had its first chips with 18A manufactured in August 2024 and was not satisfied with the result. Shortly afterward, Intel announced that its own designs, namely the "Panther Lake" and "Clearwater Forest" architectures planned for 2025 , could already be manufactured with 18A in working order. As analyst Patrick Moorhead reported on X, Broadcom was also not using the then-current developer kit for its chips manufactured with 18A. Pat Gelsinger agreed with this in a reply.
Information on yield was not provided directly by either Broadcom or Intel. It is regarded as a parameter for the market maturity of a process, i.e. a method for manufacturing chips. What is overlooked in many reports on any process is that yield is directly linked to chip size. If only one processor were to be produced on a wafer, the silicon slice from which the chips are sawn, the yield could be zero percent if there is only one single defect on the entire wafer. If there is only one defect on a wafer of around 400 chips, the yield is rounded to 100 percent. The absolute number of defects remains the same, only the type of application of the process – on one or 400 chips – results in either a completely unsuitable or excellent yield.
Videos by heise
Not every chip is suitable for every process
However, the density and distribution of defects, as they occur in every semiconductor production process – no wafer is perfect – is much more important than the yield percentage relative to the individual chip area. If, as is quite common in the industry, a yield is stated as a percentage, this always refers to a specific chip design or to an average of different, very similar designs. Different processes are therefore suitable for different types of semiconductors to different degrees, even if the structure width is nominally similar. For example, if a process is well suited for high-clocked CPUs and GPUs because they are particularly expensive to sell, it is not necessarily attractive for memory components such as DRAM or flash. This is because the aim is to keep the costs for each individual functioning chip as low as possible. What matters is the distribution of defects on the wafer, not their absolute number.
Gelsinger: "Yield not understood"
Intel's former CEO Pat Gelsinger has now confirmed that the percentage alone is a worthless figure. He writes on X: "Anyone who uses yield percentages as a value for the quality of a semiconductor without the die size does not understand the yield of semiconductors." In the original, Gelsinger uses the term "semiconductor health", which in the chip business stands for the suitability of a process for series production. A "healthy" process produces so many functioning chips that it is economically worthwhile for both the manufacturer and the customer.
Intel absolutely has to achieve this with 18A, and quickly. The process is not only intended for Intel's own chips, but is also intended to establish the financially ailing company as a real competitor to TSMC. This is the world's largest contract manufacturer for chips and the model for "Intel Foundry", as which the manufacturing division was spun off as a nominally independent part of the company during Gelsinger's time as CEO of Intel.
(nie)
