You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
As I said before, a token financial investment (like $5B) from TSMC (or a consortium led by TSMC) into IFS will be a great gesture and buy a lot of good wills.
Yes, Pat Gelsinger retired, but the damages are there to stay.
Intel's problem cannot be solved with a $5 or $10 billion cash injection, especially since the money belongs to TSMC shareholders, not just some of TSMC's executives.
Intel AZ consists of Fabs 12, 22, 32, 42 with 52/62 (on their way). 6 fabs. It tooks 25 years in round figures to get close to 6 fabs operating at once.
TSMC AZ consists of 1 fab, with 4 more on the way relatively soon. That's 5 fabs in about 10 years.
There's this industrial concept, SMED, single minute exchange of dies. How you move from project to project faster. TSMC moves from project to project about 2.5 times faster than Intel. I think this is a major factor in the whole TSMC vs. Intel discussion.
Intel AZ consists of Fabs 12, 22, 32, 42 with 52/62 (on their way). 6 fabs. It tooks 25 years in round figures to get close to 6 fabs operating at once.
TSMC AZ consists of 1 fab, with 4 more on the way relatively soon. That's 5 fabs in about 10 years.
There's this industrial concept, SMED, single minute exchange of dies. How you move from project to project faster. TSMC moves from project to project about 2.5 times faster than Intel. I think this is a major factor in the whole TSMC vs. Intel discussion.
Between Intel Arizona fab 42 and 52 there's about 10 years hiatus while TSMC constantly building new fabs and bringing new ones online. It's hard for Intel to keep internal or external resources and skills if there is a 10-year gap.
Between Intel Arizona fab 42 and 52 there's about 10 years hiatus while TSMC constantly building new fabs and bringing new ones online. It's hard for Intel to keep internal or external resources and skills if there is a 10-year gap.
Don’t forget that fab build rate is commensurate with technology cadence and volume.
Intel stalled a decade at 10nm, x86 decline and miss of AI all contributes.
TSMC of course caught and grew with mobile, added leadership HPC and now has AI also. The worlds advanced silicon and all other older node volume goes thru TSMC, of course they build fabs like crazy.
Let’s not forget they do it in Asia 7x24 with them holding all the cards. Look TSMC did in Arizona, took four years to do what they can do in Taiwan in 18 months, forget the schedule and ramp revenue miss and all the tools sitting in in a warehouse but the huge cost overruns they incurred but lost in their Taiwan profit fabs.
It will take the a decade or more to spend the 100billion they say they spend in Arizona. Look at what they did in Tainan Science Park in the past decade in comparison. The US spend is just politics pacification
"TSMC of course caught and grew with mobile, added leadership HPC and now has AI also. The worlds advanced silicon and all other older node volume goes thru TSMC,"
One thing strikes me about AI chips, the high margin ones are packed like sardines in the data centers of the world. Even the liquid cooled units are engineered for most output per watt in those barns. TSMC spent many years creating efficient chips for cellphones and showed a path to success for the AI chip merchants.
When you can't get enough juice to run things, a higher performance chip with a different architecture is a harder sell to the barn builders.
"TSMC of course caught and grew with mobile, added leadership HPC and now has AI also. The worlds advanced silicon and all other older node volume goes thru TSMC,"
One thing strikes me about AI chips, the high margin ones are packed like sardines in the data centers of the world. Even the liquid cooled units are engineered for most output per watt in those barns. TSMC spent many years creating efficient chips for cellphones and showed a path to success for the AI chip merchants.
When you can't get enough juice to run things, a higher performance chip with a different architecture is a harder sell to the barn builders.
Someone with ultra-expensive low volume chips with short production runs, and no certain repeated production schedule are the definition of a risk client for a foundry.
They have a lot of money to buy wafers at $100k+ with tiny lot numbers, and they can be gone overnight for any number of reasons.
So much capacity for risky clients is bad, and the lack of commodity high volume latest node users is worse. It gives just too much incentive to abandon the otherwise unassailable cost leadership.
