/biz/ - Business & Finance

>>23188077
add that on top of the fact that TSM yields are so damn good that the chiplet approach doesn't really make sense anymore
made a lot of sense in cutting costs with wafers from GlobalFoundries (wafers with a bunch of imperfections could still be very usable), but TSM is practically perfect that whole-wafer chips are reasonable on 7nm, and their yields on 5nm are even BETTER

AMD goes beyond just their chiplet design obviously, but a good chunk of their advantage (bang-for-buck) is disappearing. I don't expect Intel to make a comeback until late 2021, but there's definitely reasonable chance for Intel to do that.. because they also booked orders with TSM recently

tl;dr BUY MORE TSM, we loved it back in June, we love it even more now

>>21691850
i repeat the same shills points before but it's all based on Samsung fucking up.
First, their yields are shit which is why they fucked up the Qualcomm order.
Second, their transistor density is lower than TSM's, i.e. samsung's 8 nm's is comparable to TSM's 10 nm and so on. From what I remember, NVDA is using samsung chips for their lower end shit, and the only reason they got them in the first place is because TSM's capacity is limited and Samsung undercutted TSM a lot just to stay relevant with NVDA.
Third, they're basically skipping a node size to 3 nm when they're still having a lot of issues with yields on bigger nodes. Intel tried to skip a node size too and look how that turned out, just carried over existing issues they're still trying to fix.

Meanwhile TSM is flexing on the rest of the world with the news a couple days ago.
https://www.extremetech.com/extreme/313940-cerebras-wafer-packs-2-6-trillion-transistor-cpu-with-850000-cores
Something like that is only possible with a very very high yield. KInda funny considering AMD rose to prosperity because they were doing a chiplet design that goes around shitty yields (so even if you have imperfections in a wafer, you can just combine chiplets together) but looking at these news, I wonder if they really need to do that when TSM is capable of turning the entire wafer into a single functional chip.

i wish I could just stop myself from looking at how much TSM fluctuates day-to-day because I'm not gonna do anything anyways until a couple years later.