Who pays for bad gpu yields?

formulav8

Diamond Member
Sep 18, 2000
7,004
522
126
I was just thinking about yield problems with AMD and nVidia and just wondering at what point does Ati/nVidia eats the cost or Transmeta eats the cost for bad yields?

Just as a example if ati/nVidia only gets 40% yields of usable product who pays for the 60% that are not usable?? (Secondary question is what is done with the bad product?)

I'm sure that there is some contract agreement which has to do with good/bad yields?



Jason
 
Last edited:

goobee

Platinum Member
Aug 3, 2001
2,005
10
81
goobee.org
I would imagine it falls into the cost of doing business. If Goodyear released a bad tire and had to recall them, they would have to eat all of the costs. As far as I know, insurance won't cover it. The manufacturer can write off the losses too but you can't do that forever.
 

Lonyo

Lifer
Aug 10, 2002
21,939
6
81
Company A buys 200 wafers from Company Z, at an agreed price.
Company Z produces 200 wafers of whatever product.
Company A sells as many chips as can be gained from said 200 wafers.

If some of the products are "bad", they might be used in a non-perfect manner, see: most cut down products.
A 128 shader product might get used as a 96 shader one (9600GSO).
A 4 core CPU might get used as a 3 or two core one (AMD X2 or X3).
A 6MB cache CPU might get used as a 2MB cache one.

Then there's speed binning.
 

Voo

Golden Member
Feb 27, 2009
1,684
0
76
IDC had several posts that talked about that stuff in detail.. maybe I'll look around later for it.
 

aka1nas

Diamond Member
Aug 30, 2001
4,335
1
0
Company A buys 200 wafers from Company Z, at an agreed price.
Company Z produces 200 wafers of whatever product.
Company A sells as many chips as can be gained from said 200 wafers.

If some of the products are "bad", they might be used in a non-perfect manner, see: most cut down products.
A 128 shader product might get used as a 96 shader one (9600GSO).
A 4 core CPU might get used as a 3 or two core one (AMD X2 or X3).
A 6MB cache CPU might get used as a 2MB cache one.

Then there's speed binning.

I would actually think it work the other way if you are outsourcing your fabbing to a third party.
I.E.
Company A agrees to buy 200 functioning wafers at an agreed upon price, with some mutually agreed upon definition of functioning.

I can't imagine the chip designer not forcing the foundry to take on a substantial, if not total, portion of the risk for low yields. It wouldn't make much business sense to outsource such a critical portion of the business if the financial risk is not sufficiently mitigated.
 

EarthwormJim

Diamond Member
Oct 15, 2003
3,239
0
76
I would actually think it work the other way if you are outsourcing your fabbing to a third party.
I.E.
Company A agrees to buy 200 functioning wafers at an agreed upon price, with some mutually agreed upon definition of functioning.

I can't imagine the chip designer not forcing the foundry to take on a substantial, if not total, portion of the risk for low yields. It wouldn't make much business sense to outsource such a critical portion of the business if the financial risk is not sufficiently mitigated.

It's up to the chip designer to design a chip that will have high yields.

I'm sure it matters too on why chips are not functioning. If the fab messes up wafers, I'd imagine they wouldn't charge the chip company for them.

Bad wafers could also possibly be bought at a discount, which could explain why chip makers even bother with selling partially crippled products.


I'm interested to know either way. A part of the industry that I have not really read about.

Ultimately we as consumers pay for bad yields.
 
Last edited:

Daedalus685

Golden Member
Nov 12, 2009
1,386
1
0
If I understand what IDC had said it is something like this:

Fab and company agree to a price per wafer based on expected failure rate, difficulty of production etc.. Obviously if a company brings a chip that is a full wafer the Fab would quote an expected yield rate of 0%..

The company pays that price for the order, if the yields are well below what is quoted there would be penalty clauses in the deal the Fab would eat. The Fab should be well aware of what is possible before anything hits a sputtering machine. Thus I doubt anyone is ever surprised by the cost per chip except for those of us who hear well after the fact and don't get to see the contracts.

I can't imagine there are ever many "free" chips. In the end poor yields always translate into more per core for the company buying.. unless the yields are below what was agreed in a (very) thorough contract.
 

Rezist

Senior member
Jun 20, 2009
726
0
71
Exactly risks are shared in a way but the fab company has a very big performance clause. If they get high yields they get good $$$ if it's low they don't and they need to get it to what was agreed or re-do the deal.
 

aka1nas

Diamond Member
Aug 30, 2001
4,335
1
0
It's up to the chip designer to design a chip that will have high yields.

I'm sure it matters too on why chips are not functioning. If the fab messes up wafers, I'd imagine they wouldn't charge the chip company for them.

Bad wafers could also possibly be bought at a discount, which could explain why chip makers even bother with selling partially crippled products.


I'm interested to know either way. A part of the industry that I have not really read about.

Ultimately we as consumers pay for bad yields.

I'm not in the semi-con industry myself, but I deal quite a but with other forms of tech outsourcing:

If you're outsourcing production(or anything else really), it's usually mostly the manufacturer's responsibility to take into account manufacturing defect rates when they quote you a price per wafer or chip. The chip designer can only be aware in a vague way of what a 3rd-party foundry's yield rates will be.

While semiconductor manufacturing is a very sensitive process compared to other forms of manufacturing, at the end of day it's still manufacturing and the contractual side of negotiating acceptable defect rates and batch testing is a pretty mature process. The plus side that you can often still utilize a large portion of the factory seconds via speed binning is just another variable that gets factored in by both parties.