3 ms·

There are two parts to the story: architecture and process. On architecture, AMD deserves praise. On process, not so much. They didn't "git gud," they gave up.

There are two parts to the story: architecture and process. On architecture, AMD deserves praise. On process, not so much. They didn't "git gud," they gave up. Which clearly was the right move for them, don't get me wrong, but I think all this partying in the streets about the TSMC monoculture is... shortsighted.

arcticbull 6y ago
I think where they deserve praise related to process is the pragmatic and frankly pretty brilliant way they worked around 7nm processes yield issues in Zen 2. To improve yields they used a number of 8-core "chiplets" manufactured in 7nm tied together with a 14nm I/O die. This meant that while a certain blue competitor was still trying to get 28+ core monolithic dies, AMD could scale to huge core counts without reducing yields by tying together more chiplets. Their 64-core 128-thread Threadripper workstation parts should have yields almost exactly as good as their low-end CPUs. The Xeon W 28-core monolithic monster? Not so much. AMD did a lot of things right this time. Intel's got some catching up to do.
- jjoonathan 6y ago
  Yeah, I just hope they do catch up, instead of bailing like AMD and leaving TSMC the only advanced node player in town. (Yes, I'm aware of Samsung.)
  - arcticbull 6y ago
    Yep, for the health of the industry I want for the same thing. More competitive foundries are always better. GloFo isn't really interested in building anything super competitive moving forward either - they announced they weren't going to pursue 7nm. I think Intel's still pretty dead set of getting back into the foundry game, doing so is kind of necessary for defense contracts right? Maybe we'll see On Semi get in the game. They did pick up IBM's Fab 10 from GloFo last year.
colinmhayes 6y ago
Sometimes giving up deserves praise. It's not an easy decision to write off all the money you've invested in process, AMD realized it's what they had and made the right choice. That deserves praise.
Tade0 6y ago
It's not like TSMC is not replaceable - after all their EUV machines were designed and manufactured by ASML.

arcticbull 6y ago

I think where they deserve praise related to process is the pragmatic and frankly pretty brilliant way they worked around 7nm processes yield issues in Zen 2. To improve yields they used a number of 8-core "chiplets" manufactured in 7nm tied together with a 14nm I/O die. This meant that while a certain blue competitor was still trying to get 28+ core monolithic dies, AMD could scale to huge core counts without reducing yields by tying together more chiplets. Their 64-core 128-thread Threadripper workstation parts should have yields almost exactly as good as their low-end CPUs. The Xeon W 28-core monolithic monster? Not so much. AMD did a lot of things right this time. Intel's got some catching up to do.

jjoonathan 6y ago
Yeah, I just hope they do catch up, instead of bailing like AMD and leaving TSMC the only advanced node player in town. (Yes, I'm aware of Samsung.)
- arcticbull 6y ago
  Yep, for the health of the industry I want for the same thing. More competitive foundries are always better. GloFo isn't really interested in building anything super competitive moving forward either - they announced they weren't going to pursue 7nm. I think Intel's still pretty dead set of getting back into the foundry game, doing so is kind of necessary for defense contracts right? Maybe we'll see On Semi get in the game. They did pick up IBM's Fab 10 from GloFo last year.

colinmhayes 6y ago

Sometimes giving up deserves praise. It's not an easy decision to write off all the money you've invested in process, AMD realized it's what they had and made the right choice. That deserves praise.

Tade0 6y ago

It's not like TSMC is not replaceable - after all their EUV machines were designed and manufactured by ASML.