[messiah]

Autor The Exiled

Lako za AMD, oni su svoj dio napravili, sve ostalo kaj slijedi imalo bude itekak dobre temelje. Pitanje je kaj bude, ako se ispostavi da je Intelovih 10nm ipak čušpajz, mada sad naveliko pričaju (a opet nitko točno ne zna) kak je sve super i 10nm proizvodi samo što nisu. I sadašnja 10nm situacija im nije bajna, a 14nm ne budu mogli unedogled frizirati, a sve pizdarije koje su ovih godina skrivali - dolaze na vidjelo.

Evo jedan od boljih postova na redditu koje sam pročitao zašto je Intel u problemima. Ukratko, nisu znali kako će se realizirati 10nm node pa su išli u 2 smjera:
1. Multi pattern ArF immersion - pokazalo se da ima hrpetinu problema koje još nisu uspjeli riješiti
2. EUV - za koje strojeve proizvodi doslovno jedna firma na svijetu te koštaju papreno. Vrijeme isporuke je bilo 2-3 godine, a nisu na vrijeme naručili.

https://www.reddit.com/r/Amd/comment...t_with/efbtmg2

The challenge for intel has to do with the lithography process. To resolve features below 32nm there are two forks: multi-pattern ArF+immersion and EUV. Multi-pattern is achieved by performing multiple exposures with matched photomasks on the same substrate layer before the wafer resist gets resolved. This was a viable path for sub 32nm, and was the basis for the double-patterning used for Ivy-bridge and Haswell (both 22nm processes). Intel had more difficulty shrinking to 14nm, because the random factors started to creep in at that scale.

To better understand this, let's do a thought experiment:

Imagine driving a car from Los Angeles to New York and then back again, and because this is a high-precision thought experiment, let's assume that there are no other cars on the road, and the entire drive is happening in a temperature controlled environment without any debris larger than a pebble to interfere with your journey. Got it? During the first trip you intentionally drive as steady as you can, because there is a twist: during the return trip, the car must drive exactly over the same path. Remember, there are no other cars in this experiment. It's just you, the open road, and your imaginary car driving for a few thousand miles.

Sounds impossible right? There is going to be some variance and traveling exactly over the prior path cannot be realistically achieved. I am not a monster, so I will allow a little bit of error. You can drive the car back, and as long as you stay in your lane, I will count that as a success. That is what a 22nm process looks like.

Now, 14nm is going to be a bit more of a challenge. Instead of only staying in your lane for the round trip, now you are only allowed about two tire's width of difference. You also need to make two round trips instead of one. This is still possible, and with a few tweaks to the steering and suspension you might improve accuracy a bit. Maybe you can slow down a little, drive less aggressively, etc. but you need to find a way to repeat that path. And you do. You make more mistakes and it takes several tries, but you manage to pull it off! Congratulations!

Let's talk about 10nm. You'll need to make the return trip with less than one inch difference. Oh, and you'll need to make FOUR round trips, not just one. Good fucking luck!

That's multi-pattering in a nutshell. The random errors that you almost did not notice at 22nm are now gigantic obstacles to completing a successful pass. Errors are multiplied while tolerances shrink. Even when your first round trip is perfect, you're only a quarter of the way there, and you're more likely to fail with each additional trip. You might get lucky every now and then, but it's not reliable.

Now let's compare that method with EUV. It's a newer lithography process, and it offers high resolution in a single pass (no multi-patterning penalty), but there's another downside: throughput. ArF+immersion tools can process 250~300 wafers per hour. EUV tools struggle to achieve even >100 wafers per hour (this is improving with better light sources, optimizations to photoresist, etc.).

Returning to the thought experiment, EUV is like riding a bicycle from Los Angeles to New York. It's a one-way trip, and it's much slower, but the bike tires are much thinner (higher resolution), so you get a little bit more wiggle room.

These are the two forks. One is fast but error prone, the other is slow but gives you sub 22nm features in a single pass. The immersion lithography process is about ten years old, meaning that it is a mature process, with lower operating costs. The tool footprint is comparable to older platforms, making it viable for high volume manufacturing. You can pump out 22nm chips with low defects and high yields. It works at 14nm but the yields are lower, and at 10nm you start to rely more on superstition than engineering. EUV has higher resolution, but the tools are MASSIVE, and can only be installed in fabs with high ceilings, and a ballroom layout (chase and bay clean rooms cannot handle these dreadnought-sized lithography tools). EUV tools are also considerably more expensive, and have higher operating costs. The light source is "soft x-rays" and only works under vacuum, which complicates maintenance - the entire chamber has to be purged of atmosphere before resuming production.

Intel invested in both of these platforms, without knowing which would be viable for each node. They bet wrong on 10nm multi-pattering but due to costs could not afford to switch to an EUV process. To complicate matters further, ASML is the only company in the world to offer a production capable EUV lithography tool. Intel would have needed to place their orders years in advance to meet their production needs.

TL;DR - semiconductor manufacturing is fucking complicated and mistakes are unforgiving. Even with excellent long term planning, you can easily fall behind.