Raja Koduri explains why Intel outsourced Alchemist GPU manufacturing to TSMC

In response to him, it’s as a result of Intel doesn’t have adequate manufacturing capability.

Throughout Intel Structure Day 2021, the corporate formalized the outsourcing of the manufacturing of Alchemist GPUs to TSMC: they’re primarily based on the Taiwanese founder’s N6. Intel’s choice was stunning, since in contrast to NVIDIA and AMD, which design chips however don’t manufacture them, Intel has a number of foundries. Raja Koduri, director of the AXG department (Accelerated Computing Methods and Graphics) of the corporate, defined within the columns of a Japanese media a number of the causes that prompted the agency to outsource.

The principle justification put ahead by Raja Koduri pertains to inadequate manufacturing capability: Intel would merely not have sufficient strains to provide Arc GPUs in giant numbers with out this being executed to the detriment of different merchandise. .

Definitely, so far as CPUs are involved, Intel has been doing significantly better than AMD in latest months, with a a lot greater supply. However CPU manufacturing is much less complicated than GPU manufacturing. To persuade you, let’s take the comparability given by our colleague from Tom’s Hardware FR : the 7 nm chiplets of a Ryzen 7 5800X signify an space of ​​roughly 80.7 mm² and account for 4.15 billion transistors; the GPU of a Radeon RX 6900 XT has an space of ​​520 mm2 for 26.8 billion transistors. If we take a much less environment friendly mannequin, the RX 6600 XT for instance, the GPU floor space stays 237 mm² for 11.06 billion transistors.

Intel has doubled its manufacturing capability in 3 years

TSMC N6 strains that can not be used to fabricate AMD chips

Going again to what Raja Koduri mentioned, listed here are those reported: “We first needed to decide the manufacturing capability of the method. […], and the superior course of [d’Intel] didn’t but have adequate capability. Different traits, such because the working frequency that can be utilized, are additionally essential elements, and value can be a problem. These three parts, that’s to say value / efficiency / capability, are taken under consideration in deciding which course of to make use of ”.

Raja Koduri doesn’t instantly point out the delays taken by Intel on its burning nodes in recent times; it nonetheless implies that the founder just isn’t in a position to mass-manufacture merchandise on a burning node equal to TSMC’s N6, which is assumed to be mandatory to stay aggressive with present options from AMD and NVIDIA primarily based on TSMC’s 7nm and Samsung 8nm processes, respectively. As well as, by going to occupy just a few strains of TSMC with its merchandise, Intel can be depriving its opponents of extra capability on the node in query.