Intel announces cancellation of 20A process node for Arrow Lake | Hacker News Reader

Intel announces cancellation of 20A process node for Arrow Lake

tomshardware.com

10 pointslayer811 comments

Hard to square the "we're canceling 20A because 18A is coming out way better than we thought" narrative with this report that also came out yesterday:

https://www.reuters.com/technology/intel-manufacturing-busin...

Broadcom, one of Intel's first (and largest) new foundry customers, is saying their 18A samples all failed their testing and that they "concluded the manufacturing process is not yet viable to move to high-volume production".

The announcement made by Broadcom is weird, because according to what Intel has been saying for a long time, the 18A process was expected to become ready for high-volume production only around the middle of 2025.

Therefore it is absolutely normal that it is not ready at this time.

What Broadcom could say is that for example a TSMC process that is nine months away from high-volume production would already at this time have better yields than demonstrated now by Intel and that Broadcom does not trust enough Intel that in nine months they will succeed to increase the yields at the level desired by Broadcom.

Actual announcement: https://www.intel.com/content/www/us/en/newsroom/opinion/con...

Sounds like 18A yield is so bad they are bringing in the 20A team to fix things...at the expense of using 20A in their own products. Guess we know which roulette square the company is putting it's money.

The article does not imply that at all.

No what is bad is 20A and it makes no sense to waste money on improving it, when it will be immediately replaced by the better 18A.

This is an obvious method for saving money, stop immediately any expenses for fixing 20A and concentrate all efforts to improving the yields of the 18A CMOS process, which is what Intel hopes to use to fabricate all its 2025/2026 server CPUs (Clearwater Forest and Diamond Rapids) and all its 2025/2026 laptop CPUs (Panther Lake).

The gap left by cancelling 20A can be filled more cheaply by outsourcing more to TSMC until the second half of 2025, when 18A should be ready for mass production.

From the article:

Intel again noted that it had reached a sub-0.40 D0 defect density (def/cm^2) for 18A, a critical measurement of the yield rate for a process node. A process node is usually considered production-worthy and healthy once D0 reaches 0.5 or below.

“One of the benefits of our early success on Intel 18A is that it enables us to shift engineering resources from Intel 20A earlier than expected”

The Tom’s article said it would outsource production to TSMC. Maybe in the short term?

Intel has already outsourced to TSMC the fabrication of most laptop CPUs and desktop CPUs for the year 2024/2025 (i.e. Lunar Lake and Arrow Lake S are made completely by TSMC, while Meteor Lake is made mostly by TSMC, with only one tile made by Intel).

The only remaining product that was planned to use the 20A process was Arrow Lake H for laptops, to be launched in early 2025. This announcement implies that the Arrow Lake H CPU tile, the only part of it made by Intel, will be ported to the TSMC process, so that Arrow Lake H will be completely made by TSMC, exactly like the desktop Arrow Lake S, which will be launched next month.

For the year 2024/2025, Intel has retained only the fabrication of the server Xeon CPUs and of the older CPU models that are still in production.

The plan with the 18A Intel CMOS process is that will be used not only for the 2025/2026 server CPUs, but also for the 2025/2026 laptop CPUs (Panther Lake), returning the fabrication of the latter to Intel.

According to the current rumors, the Intel desktop CPUs after next month will be produced at TSMC at least until 2026.

Meteor Lake:

- Base Tile -> Intel 16

- Compute Tile -> Intel 4

- SOC Tile -> TSMC N6

- I/O Extender Tile -> TSMC N6

- Graphics Tile -> TSMC N5

Lunar Lake:

- Base Tile -> Intel 16

- Compute Tile -> TSMC N3B

- Platform Controller Tile -> TSMC N6

The article makes it sound like for Arrow Lake, EVERYTHING, including the Base Tile, will be made by TSMC, and then Intel will just glue everything together into the CPU package.

The base tile is only equivalent with a miniaturized PCB that makes the connections between the other tiles.

Its fabrication is much simpler than that of the other tiles and it belongs to the packaging and testing operations, which are done by Intel itself.

It does not belong to the fabrication of the active parts of the CPU SoC. The Base Tile could have been fabricated by any from more than a dozen of IC manufacturers, while the active tiles are beyond the capabilities of any company besides TSMC, Samsung and Intel.

The Base Tile is just "glue" for the tiles that contain the active devices.

Reply on news.ycombinator.com