Oh, no. Marketing strikes again!
- Intel has silicon manufacturing (fab business) as one of its business units.
- AMD sold the manufacturing and formed GloFo (Global Foundries) to survive. AMD is now fabless.
- Foundry businesses supply the manufacturing to fabless semiconductor companies. Only two brands of foundries in the world have 7nm process in production. Samsung and TSMC. We know Samsung is more than foundry alone, a bit like Intel. Taiwan holds about 60% of the whole worlds semiconductor production. Both TSMC (about 52%) and UMC (about 8%) are located in Taiwan.
GloFo tried 7nm and gave up on it, so we have left: TSMC N7 versus Intel 14nm.
REMINDER: while decades ago, the gate dimension was related to naming of the technology nodes, this is no longer true for a long time. Brutal statement: you name it whatever you want. Therefor, my position for a long time has been: Intel made a big mistake by not naming their process similar to the naming Samsung and TSMC use because it creates a perception problem.
Not many know that 10nm Intel is more or less comparable with TSMC/Samsung N7.
(winner takes all image source: pixabay)
It is a simplified view, because the original 2.8 improvement that Intel set as goal was scaled down to 2 times because of ongoing issues with yield (not a lot of non-defective silicon output).
Mind you, the servers from Intel are still in 14nm but it is not the original 14nm either. It is 14nm++, the “object oriented” version of 14nm so to speak. Highly optimized for frequency, area and yield.
GloFo gave up on 7nm because of technical issues. A fab in N7 or N5 goes around $14 to $15B in capital investment. Only a few can afford it.
TSMC solved quite some issues by using ASML technology (EUV for the interested).
And is risk producing (or close to) N5 at the moment.
Foundry has only one goal and that is sell silicon manufacturing.
It also is the foundation for world domination in AI, the winner takes all due to the exponential curve of AI.
AMD has an advantage over Intel in the server CPUs now.
EPYC Rome has 64 cores single socket vs 56 for Intel.
Performance per Watt in N7 is better than in Intel’s 14nm++.
The price per core is better than Intel’s.
Plus they are the ONLY company that has both discrete CPU and GPU offerings.
Intel has no discrete GPU yet, Nvidia has no x86 CPU.
Plus plus, EPYC has PCIe 4 in the server market versus PCIe 3 for Intel. it makes a lot of difference for bandwidth going off-chip.
What do we need to remember:
- Everything is marketing.
- Confusing marketing numbers: keep it simple, adapt for perception purposes.
- AMD has no manufacturing and relies on TSMC (with GloFo IO in 14 or 12nm I think).