AMD on 58XX supply issues

[MrK6]

Great post IDC. I like equations and charts.

Same here. Those were very helpful and a great read, Idc. Thanks for the information.

 

[Voo]

Great post IDC. I like equations and charts.

Second that.

Thank you for that post, that's a great explanation for us laymen

 

[Idontcare]

Thanks guys for the feedback, it can be difficult sometimes for me to determine by self-assessment whether I am effective in communicating the information so I appreciate the feedback.

I wanted to add that the strong correlation between yields and defectivity and process maturity is what motivates companies like Intel and AMD to be so assertive in talking about this metric when it comes to discussing the ongoing success of the node currently under development:

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=3513&p=3

edit: To add some perspective to the numbers, GloFo's recent announcement that they were expecting to see 50% "natural" yields on a 32nm-based 24Mb sram test chip (which I had previously estimated to be around 4mm^2) by the end of the year means they are expecting to be at a D0 of around 18-19 defects per square centimeter...a long ways to go before getting down to 0.2 DD.

 

[Zebo]

Thanks guys for the feedback, it can be difficult sometimes for me to determine by self-assessment whether I am effective in communicating the information so I appreciate the feedback.
.

You're best poster in video - I don't say much just read for insight.

 

[yacoub]

I bet they would be selling the 5850 @~$200 and 5870@~300 if NVIDIA had some DX11 competition out, so they probably are reaching those 50%.

I also think they would prefer having a lot more cards to sell than charging an extra $30-50 per card as they are atm.

Yup, and i will be waiting until they do drop in price. These prices are silly.

 

[Xcobra]

As much of a noob I am to technical stuff like this, it was very helpful IDC... but then, im not surprised

 

[james1701]

In a situation like we have now, who pays for bad yields? I am sure when the companies do the contracts for chips like this, the fab puts a little cushion in for the waste, but when they have a problem in the middle of the run like this, who pays? Does AMD or Nvidia have to pay more for the wasted chips, or does the fab have to eat the cost. Does AMD or Nvidia get a discount when a fab can not produce an X number of chips in X amount of time, since they loose costumers to on side or the other, or they can't meet their contracts with other companies?

 

[jvroig]

This article gives a nice example of the impact of all this on yields.

Thanks for the explanation, and thank you for that link. (If you have other online resources about chip design and fabrication, I'd be most appreciative)

Based on what you said, so Juniper does get more yield % per wafer, and more chips per %. Thanks, it is clearer to me now, and I've got additional stuff to study.

Who pays for the bad yields, by the way? Is that a risk that customers engage in when paying for fabrication services, or does TSMC or whatever fab absorb it by returning the money in the form of discounts or credit memos?

 

[Warren21]

Who pays for the bad yields, by the way? Is that a risk that customers engage in when paying for fabrication services, or does TSMC or whatever fab absorb it by returning the money in the form of discounts or credit memos?

From the limited understanding I have of 3rd-party fab companies, I'm relatively sure ATI, nVidia and other fabless manufacturers pay on a per-wafer basis. Whatever happens to the yields from that wafer are therefore the client's headaches. That's why you see salvaged products like the Radeon 4830's and 5850's - trying to maximize their dollar. Lost yields represent wasted money.

This isn't to say TSMC and friends don't suffer if their process is immature - on the contrary, having a mature process guarantees returning customers. From what I hear, TSMC's 40nm process has been wrought with issues and even current yields are lower than expected. The problem is that only TSMC has production-ready 40nm capabilities... ATI/nV can't really shop anywhere else. To further complicate the idea of switching fabs, each fab company's process works a little differently. A few years back, I believe one of ATI's or nVidia's product lines was delayed because of the switch from UMC (?) to TSMC. Feel free to correct my inaccuracies... I remember this story vaguely (I'm not sure if this is or was related to the X1800 aka. R520 silicon bug).

 

[Idontcare]

In a situation like we have now, who pays for bad yields?

Who pays for the bad yields, by the way? Is that a risk that customers engage in when paying for fabrication services, or does TSMC or whatever fab absorb it by returning the money in the form of discounts or credit memos?

Who pays? The consumer of course, eventually at least.

But I know that doesn't answer the spirit of your questions which is more to the point of how are foundry contracts structured to deal with the fact that yields are not 100% and in fact they can vary quite a lot from wafer to wafer and over time?

As you and Warren touched on in your posts the typical foundry contract is structured based on a flat-rate per wafer (the rate itself is of course set to benefit volume orders) charge and carries with it some minimum requirements on the yields.

A 40nm contract might stipulate that the customer pays $7k per wafer and that they only pay for the wafers with yields that exceed 50%.

Consider that the foundry doesn't want to spend money raising yields if all it accomplishes is their customer orders fewer wafers because they know they are going to get more die from fewer wafers when the yields are higher.

So even though contracts are typically based on a per-wafer price with yield minimums in place the reality is of course that wafer prices are entirely negotiated with a large degree of yield expectations in place such that the customer is effectively negotiating on a price per sellable chip basis anyways.

There are extra incentives in place for the early adopters of a tech node to compensate them for the elevated risk they are exposed to by the very nature of being early adopters. AMD for example with RV740 was the canary in the coal mine, every node by definition has one, and you can bet they received some pretty favorable contractual pricing in exchange for being willing to be that canary.

In some ways it is like negotiating to buy a new car, you have some idea what the typical going rate is for the car, you talk to friends who have been car-shopping and you get msrp and kelly bluebook numbers but you don't know what your price is going to be until you go negotiate with the salesmen. And even then the price you pay isn't entirely clear because there will be warranty negotiations, and those service checkups that never really have a predefined expense list, etc.

But ultimately every dollar that TSMC wants to land in their hand comes from the hands of their customers and those dollars came from the hands of the AIBs whose hands had dollars that came to them by way of Newegg's hands who managed to get the dollars from our hands.

So who pays for these things when these things cost more money to manufacture? We do, eventually.

 

[jvroig]

Thank you for the reply, IDC. If I may ask for a little more clarification...

A 40nm contract might stipulate that the customer pays $7k per wafer and that they only pay for the wafers with yields that exceed 50%.

So in this scenario where yields are terrible, what happens to wafers whose yields are below the stipulated percentage? Does AMD get them for free? Or are they thrown out entirely? (seems like a gigantic waste, so I would think not). I'd imagine that AMD didn't agree in a contract that yields would be below, say, 50%, and when yields dipped to 40% and below, does that mean they keep on harvesting chips from TSMC for free? I can't imagine how TSMC maintains business like that. Or is it offset by the price per wafer they agreed on, meaning they set the price per wafer a little higher to compensate for the reality that a lof of the wafers would be "free", so the good wafers will be very expensive, netting a reasonable average price? Is this what you meant when you said this quote below:

So even though contracts are typically based on a per-wafer price with yield minimums in place the reality is of course that wafer prices are entirely negotiated with a large degree of yield expectations in place such that the customer is effectively negotiating on a price per sellable chip basis anyways.

 

[james1701]

Thanks for the reply IDC. I was just curious because in this instance, part of the problem came from one of the machines being no calibrated causing worse than expected yields. Since this was a problem with the foundry, and not a design issue with the chip, I wonder if AMD gets any type of refund on the wafers they already paid for, since it was not their design that caused the problem.

 

[poohbear]

Same here. Those were very helpful and a great read, Idc. Thanks for the information.

for someone called idontcare he sure seems to care people understand the issue at hand. oh the irony of the internet!!!!

 

[Idontcare]

So in this scenario where yields are terrible, what happens to wafers whose yields are below the stipulated percentage? Does AMD get them for free? Or are they thrown out entirely? (seems like a gigantic waste, so I would think not). I'd imagine that AMD didn't agree in a contract that yields would be below, say, 50%, and when yields dipped to 40% and below, does that mean they keep on harvesting chips from TSMC for free? I can't imagine how TSMC maintains business like that. Or is it offset by the price per wafer they agreed on, meaning they set the price per wafer a little higher to compensate for the reality that a lof of the wafers would be "free", so the good wafers will be very expensive, netting a reasonable average price? Is this what you meant when you said this quote below:

Understand that not every contract is the same and that I don't have details on what the exact structure is for AMD's Cypress/Juniper 40nm contracts but assuming it is structured like a typical foundry contract then it will have three yield/payout criterion. Something along the lines of:

1. If functional wafer yield is greater than 50% then AMD pays $7k to take delivery of all packaged IC's that pass AMD's parametric binning constraints
2. If functional yields are below 50% but greater than 20% then AMD pays a pro-rated amount for the wafer which depends on the yield versus the 50% threshold. (so a wafer yielding 30% would cost AMD $7k*30/50 = $4.2k)
3. If functional yields are below 20% then AMD gets the chips for free.

Now number 2 and 3 fulfills the role of providing some assurance to the customer (AMD) that the foundry (TSMC) has motivation to make sure those yields are improving (costs TSMC money to improve yields, so they need to be incented to invest that money).

Number 1 will be persistently renegotiated, as yields improve that 50% number will be moved up and up but so too will the per wafer cost unless the customer is also moving their order volumes up and up.

Something like: "Place and order for 5000 wafers when yields are 50% and we'll charge you $7k/wfr, but a year from now when yields are 90% (so you are then taking home nearly 2x the number of chips from your 5000 wfrs) we are going to charge you $9k/wfr unless you double your order volume to 10,000 wfrs"

Thanks for the reply IDC. I was just curious because in this instance, part of the problem came from one of the machines being no calibrated causing worse than expected yields. Since this was a problem with the foundry, and not a design issue with the chip, I wonder if AMD gets any type of refund on the wafers they already paid for, since it was not their design that caused the problem.

Yep, anything having to do with functional yield loss (particles, defects, etc) as well as parametric yield loss owing to transistor metrics not hitting target (static leakage being out of spec, Idrives being out of spec, etc) is lost revenue potential for the foundry if the yields are suppressed below any of the critical thresholds outlined in any given contract. The thresholds for any given contract will vary.

These contracts are not really setup to penalize anyone for profit though, AMD doesn't want to make money on 40nm because they are getting such crappy yields that the few chips they get off the wafers are free to them. Likewise the foundry is not out to price their wafers so high that no one can make profits selling the good chips even if the yields are high.

The foundry/fabless business relationship is one based on symbiosis, a common goal of survival which places both partners in a position where their own best interests are best served by helping the other partner achieve what is in their best interests as well. There is a lot of give and take in the relationship because generally everyone involved understands everyone is working as hard and as fast as they can.

Think about this: have you ever heard of any of the foundries getting involved in lawsuits with their customers? Sure it happens but they are rare.

 

[jvroig]

Understand that not every contract is the same and that I don't have details on what the exact structure is for AMD's Cypress/Juniper 40nm contracts but assuming it is structured like a typical foundry contract then it will have three yield/payout criterion.

Something along the lines of:

1. If functional wafer yield is greater than 50% then AMD pays $7k to take delivery of all packaged IC's that pass AMD's parametric binning constraints

2. If functional yields are below 50% but greater than 20% then AMD pays a pro-rated amount for the wafer which depends on the yield versus the 50% threshold. (so a wafer yielding 30% would cost AMD $7k*30/50 = $4.2k)

3. If functional yields are below 20% then AMD gets the chips for free.

Makes perfect business sense. Thanks, IDC.

 

[Borealis7]

on a side note, i suggest a public flogging of people with crossfire 58xx's. some people have the cards on a month of back order while they have 2??? outrageous!

 

[jvroig]

on a side note, i suggest a public flogging of people with crossfire 58xx's. some people have the cards on a month of back order while they have 2??? outrageous!

Sir, as one who also dreams of owning a 5850, I cannot argue with your logic. I shall make haste to prepare a cat-o-nine-tails.

AMD/TSMC promised to have this supply issue sorted out by 2nd week of December at the latest, right? Is this confirmed? I forgot. But I hope that means lower prices by January 2010.

 

[faxon]

/hides

 

[Idontcare]

AMD/TSMC promised to have this supply issue sorted out by 2nd week of December at the latest, right? Is this confirmed? I forgot. But I hope that means lower prices by January 2010.

You want supply to equal or exceed demand AND lower prices (exacerbates demand) too? You crazy man

AMD can easily make supply meet demand by second week of December. Just double the prices on existing stock and watch that demand dry up and suddenly everyone will have $800 5870's in stock for you to choose from

 

[jvroig]

You want supply to equal or exceed demand AND lower prices (exacerbates demand) too? You crazy man

When you put it that way, it does sound crazy I thought they were making sure yields by then would be up to snuff, and in my mind I assumed it *might* be enough for the 58xx.

 

[decadentia84]

So pretty much buy a 5850 now at the current price, or wait a long time for price drops...


...sigh I may have to hold onto my 8800gt much longer (or spring the 300$ now)

P.S: Great information in this thread, I appreciate you taking the time to digest and compile

 

[jvroig]

You want supply to equal or exceed demand AND lower prices (exacerbates demand) too? You crazy man

AMD can easily make supply meet demand by second week of December. Just double the prices on existing stock and watch that demand dry up and suddenly everyone will have $800 5870's in stock for you to choose from

Let's try this again, I realize I was not clear with my words when you answered me

First, let's lay down the assumptions:

1.) While the 58xx may be "awesome", the supply problem is mostly due to the less than expected yields at TSMC.

2.) By "sorted out by 2nd week of December", we mean that TSMC will start to have its 40nm yield up to the expected level. EDIT: However, we expect the prices to fall January (mid or end) 2010, not immediately in December as well, of course.

3.) We further assume "up to the expected level" to mean 20% more.

4.) AMD doesn't want the shortage as much as we hate it. We want to give them our money, and they sure want to take it, if only they could.

5.) The increase in price is merely due to the shortage, and exacerbated by the price-gouging of the retailers.

6.) We don't know the actual demand, nor the actual supply, so we don't really know how to plot the supply-demand curve here. This is important only in so much as for all we know, the 58xx series is only backlogged by 10,000 cards beyond the current rate of production, instead of 100,000 (who knows? Do you? I don't, and I can't find anybody who claims to know or does know).

Now, assuming those are "facts", and TSMC does deliver on its promise to deliver much better yields by 2nd week of December at the latest, and AMD gets more chips per wafer, and retailers finally get more stock, and consumers find more options, the free market is going to balance itself eventually, and price-gouging c/o NewEgg and other e-tailers/retailers will die down.

If the assumption is true that TSMC starts delivering the goods by 2nd week of December, and the improvement is significant, and by the start of January AMD can be more generous to retailers, why won't the prices return to normal, especially the price gouging disappearing?

Is the demand that great, or the production that limited? I have no source for either, so I cannot comment on either.

 

[BenSkywalker]

When comparing the 58xx parts to the 57xx parts in terms of availability, have people considered the small fact that the 57xx parts suck ass? Even if you are die hard 'ATi pwns jooo 4 l1fe yo!' type the 4870 smashes the 5750 for the same price just as the 4890 smokes the 5770.

 

[ronnn]

When comparing the 58xx parts to the 57xx parts in terms of availability, have people considered the small fact that the 57xx parts suck ass? Even if you are die hard 'ATi pwns jooo 4 l1fe yo!' type the 4870 smashes the 5750 for the same price just as the 4890 smokes the 5770.

Actually I would have thought even the most rabid nvidia fan would agree that these are a rather elegant solution. Low power, quiet, dx11 and decent fps. The price is high, but the other guys are currently dropping the ball. I have been thinking 5770, but still am leaning to 5850. Must admit the lower power, quiet, cool is very tempting.

 

[Ayah]

Yeah, thanks for the awesome explanations and graphs IDC.