AMD on 58XX supply issues

[thilanliyan]

http://www.driverheaven.net/news.php?newsid=322

 

[happy medium]

ok? The 5970 is here,wheres the better supply?

Out of the entire 58/5900 series on Newegg,there is one 5850 avalable for 310$.

 

[dguy6789]

The only way a better supply would be noticed is if supply exceeds demand. They could have increased the supply but still not by enough to not be sold out instantly.

 

[Tempered81]

"Yeah we're on track. Everything is going as planned. We are meeting high demand."

Bullsh*t. Make more cypress GPUS. They are obviously sold out everywhere you look, so no you aren't meeting demand.

 

[GrumpyMan]

This also keeps the price high.

 

[Idontcare]

I hope AMD is making good margins on cypress. Limited supply and huge demand, if they can't get >50% GM on an opportunity like this then god rest the souls of any AMD shareholder.

 

[GaiaHunter]

I hope AMD is making good margins on cypress. Limited supply and huge demand, if they can't get >50% GM on an opportunity like this then god rest the souls of any AMD shareholder.

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.

 

[edplayer]

The only way a better supply would be noticed is if supply exceeds demand. They could have increased the supply but still not by enough to not be sold out instantly.

good post

I'm sure there will still be people posting in here that cannot understand those two points though.

I like how they claim that they are above production targets for the 5700 series. I plan on picking one up once I see a hot deal in the $100 range, I figure in about 3 months or so (for a 5770).

 

[jvroig]

I'm a little curious about why the 58xx series are out of stock presumably due to yield problems, while the 57xx series are not so rare. Don't they use the same 40nm process at TSMC?

 

[Idontcare]

The 57xx series use Juniper whereas the 58xx and 59xx use Cypress.

Both are products of TSMC's 40nm process but Juniper is smaller than Cypress so yields will be intrinsicly higher.

Also since Juniper chips are not "harvested" cypress chips that means wafer start allocation planning for Juniper wafers is determined solely based on supply/demand expectations of 57xx series. Meaning 60-90days ago when today's Juniper chips were just getting started in the fab AMD started enough Juniper wafers such that the supply today exceeds today's demand for Juniper products.

Had they a time machine I'm sure they'd go back 90days and start fewer wafers of Juniper and more wafers of Cypress (which presumably have more lucrative margins).

 

[jvroig]

Both are products of TSMC's 40nm process but Juniper is smaller than Cypress so yields will be intrinsicly higher.

So the yields depend on the size (also complexity?) of the chip? So being smaller, Juniper's yields are better?

I knew they were different chips, but because the TSMC shortcoming was always labelled as "40nm FAIL" or something, I assumed it applied without bias to the Juniper chips as well.

How big a difference might the yields have? If Cypress is, say, 40-50%, could Juniper be significantly better at around 60-70%?. Or would the yield % remain the same at 40-50%, but there would simply be more units of Juniper produced despite the same percentage since there would be more Juniper chips produced per wafer anyway thanks to their smaller size?

Of course, you know I'm just starting out here about the wonders of chips and fabs, so pardon my "noobness".

Meaning 60-90days ago when today's Juniper chips were just getting started in the fab AMD started enough Juniper wafers such that the supply today exceeds today's demand for Juniper products.
Had they a time machine I'm sure they'd go back 90days and start fewer wafers of Juniper and more wafers of Cypress (which presumably have more lucrative margins).

I expected the 57xx series to have significantly more demand than the 58xx, since the price difference between the two is a wide chasm. From where I am, a 5750 costs about half of a 5850, while a 5770 is a little less than half of a 5870. I guess I would have needed a time machine as well if I were an AMD employee involved in the planning of production.

 

[lopri]

We deliberately timed the introduction of the dual-GPU ATI Radeon HD 5970 to coincide with the increase in supply of ATI Radeon HD 5800 series processors in order to ensure sufficient supply for both ATI Radeon 5800 and 5900 series products.

Maybe because I'm drunk but I cannot comprehend what he says above. Does that make sense to everyone else?

 

[jvroig]

Maybe because I'm drunk but I cannot comprehend what he says above. Does that make sense to everyone else?

It does to me. I understand it to mean "We launched the 5970 on this time, and not earlier, because now we are getting to increase supply of the necessary parts. Had we launched earlier, it would have aggravated the supply disaster"

 

[thilanliyan]

It does to me. I understand it to mean "We launched the 5970 on this time, and not earlier, because now we are getting to increase supply of the necessary parts. Had we launched earlier, it would have aggravated the supply disaster"

That's what I read from it.

 

[SlowSpyder]

I'm a little curious about why the 58xx series are out of stock presumably due to yield problems, while the 57xx series are not so rare. Don't they use the same 40nm process at TSMC?

What IDC said, and I also wouldn't be shocked if demand for the 5850 and 5870 is higher right now, despite their higher price. We've had 4870/GTX260 performance available for quite some time now, which is what the 5770 gives you roughly. I think there is more excitement around the 58xx cards right now. Over the life of the cards, the 5770 may sell more as it's cheaper, but at least right now around the launches I think the 58xx demand is probably really strong as enthusiasts try and find them.

 

[jvroig]

We've had 4870/GTX260 performance available for quite some time now, which is what the 5770 gives you roughly. I think there is more excitement around the 58xx cards right now. Over the life of the cards, the 5770 may sell more as it's cheaper, but at least right now around the launches I think the 58xx demand is probably really strong as enthusiasts try and find them.

I failed to consider that. You are right, from a 4850/4870, the 57xx doesn't amount to much of an upgrade unless DX11 is a matter of life and death, and it isn't.

 

[Idontcare]

So the yields depend on the size (also complexity?) of the chip? So being smaller, Juniper's yields are better?

I knew they were different chips, but because the TSMC shortcoming was always labelled as "40nm FAIL" or something, I assumed it applied without bias to the Juniper chips as well.

How big a difference might the yields have? If Cypress is, say, 40-50%, could Juniper be significantly better at around 60-70%?. Or would the yield % remain the same at 40-50%, but there would simply be more units of Juniper produced despite the same percentage since there would be more Juniper chips produced per wafer anyway thanks to their smaller size?

Yep, smaller die generally means higher yield. So not only do you get more die on the wafer when the die are smaller, but you get even more of them off the wafer as being functional and sellable as well.

There are more details than you probably ever care to read about on the topic in my posts here and here.

The simple rule-of-thumb equation for yields between two chips which are similar in design/complexity and produced in the same fab is:

Yield_Chip_A = Yield_Chip_B^(DieSize_Chip_A/DieSize_Chip_B)

Juniper's die size is 166mm^2. Cypress is 334mm^2.

So if Cypress' yields are say 50% then Juniper's yield entitlement would be:

Yield_Juniper = 0.5^(166/334) = 71%

A simplistic calculator for determining the number of dies per wafer (DPW) based on no more info than the die-size of the chips (S) and the diameter of the wafer (d):

Using d = 294 mm (3 mm wafer edge exclusion, aka WEE) and the S from above we arrive at an estimated 358 dies per wafer for Juniper and 167 dies per wafer for Cypress.

After factoring in the yields (remember we assumed Cypress yields, so this is all relative not absolute), that means we see for wafer of Juniper chips AMD nets 50%*167 = 84 Cypress chips and 71%*358 = 254 Juniper chips.

That ratio is 3:1. Now the ratio will change obviously as the background defect density drops and overall functional yields improve, and also the yields we are talking about here do not include losses from parametric yields (which are further biased against larger chips, so if anything the ratio of good Juniper chips per wafer to good Cypress chips per wafer is even larger than 3:1).

I expected the 57xx series to have significantly more demand than the 58xx, since the price difference between the two is a wide chasm. From where I am, a 5750 costs about half of a 5850, while a 5770 is a little less than half of a 5870. I guess I would have needed a time machine as well if I were an AMD employee involved in the planning of production.

Remember the shortage of Cypress isn't about the ratio of demand for Juniper:Cypress not being 1:1...the actual demand for 57xx products could be 10x more than that for 58xx, but if AMD anticipated the ratio to be 11x and they started enough Juniper wafers to support that much 57xx demand then that means AMD has excess supply of Juniper (11x > 10x) and so they could have done a better job of balancing the fixed 40nm wafer allocation by starting fewer Juniper wafers and more Cypress.

I've no doubt they did account for bigger demand of the cheaper products, it just appears the ratio they chose resulted in more 57xx supply than demand. Not a bad thing, just not optimal from a gross margins point-of-view.

 

[cusideabelincoln]

How big a difference might the yields have? If Cypress is, say, 40-50%, could Juniper be significantly better at around 60-70%?. Or would the yield % remain the same at 40-50%, but there would simply be more units of Juniper produced despite the same percentage since there would be more Juniper chips produced per wafer anyway thanks to their smaller size?

I'm going to say it's a little of both of your statements (questions), but like you I am not completely confident in this answer. Juniper gets more chips per wafer thanks to its size, and it probably does have better yield.

 

[Voo]

I'm just curious, but does anyone know how the ratio for working/dead cypress chips is (or for another chip, say 48xx or GT200 when they were new)?

I have no idea, for me 80% sounds as possible as 40%..

 

[Idontcare]

I doubt specific yield numbers have ever been put into the public domain.

 

[Voo]

I doubt specific yield numbers have ever been put into the public domain.

Yeah I thought so, but I would assume that some of the experts here, could at least tell me about what order of magnitude we're talking about

Does "bad yields" mean 30% good chips, 50? 10?

 

[zmatt]

58xx and 59xx supplies will be low until either demand runs out or when TSMC fixes their yield issues. This really isn't AMDs fault, they have no control over the quality of production coming out of TSMC. I expect Nvidia's chips to have a similar if nto more aggravated problem when they are released. Chances are they will be large dies just like their predecessors and will have a lower count of dies per wafer and a higher fail rate.

 

[Idontcare]

Yeah I thought so, but I would assume that some of the experts here, could at least tell me about what order of magnitude we're talking about

Does "bad yields" mean 30% good chips, 50? 10?

What you are asking about involves what we refer to as "functional yield entitlement" in the industry.

Functional yield will vary as a function of the die size (larger die have less yield) and as a function of the background defectivity levels of the fab. Particles that fall on the wafers from the air, etc.

Generally the functional yield entitlement for a given diesize increases as the process node matures.

When a node is just being transferred from development over to production the defectivity levels will be high and the yield entitlement will be low.

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

There are lots of ways to model yield entitlement. The equation I gave above is a very simplistic version.

Personally I like to use the following equation which captures the effect of "defect clusters" on the yield impact. (makes it a little more "real-world" like if you will)

(note this merely captures the functional yield, actual device yield is further reduced by parametric yield loss thru binning out chips which are too slow clockspeed-wise or consume to much power to hit sellable clockspeeds, etc...as such this equation represents an upper-limit estimate of device yield)

Armed with this equation and some information on the diesize of the particular IC you are interested in you can make a plot of yields for your IC versus D0 (pronounced "dee-zero").

Remember D0 is an indicator of your process node maturity, so it gives you an idea of how badly yields will be at any point during the node maturation timeline.

When we through all this together and compare Juniper and Cypress we get the following:

So now we get a "big picture view" impression of the answer to your question. When a process node is first released to production the defect density levels can be rather large, even larger than 1 DD. This would clearly penalize the yields of a large chip like Cypress, functional yields would be below 10%.

That is why the first chip AMD debuted using 40nm was the much smaller RV740 (137mm^2) as that had much better chances of yielding enough chips to make it profitable to bother trying to sell them.

I took the liberty of highlighting in yellow the region of the chart that most production fabs operate at, some do much better than 0.2 DD but this is sort of the "middle of the road" average across the industry for mature process nodes.

So what is "bad yield"? Depends on the chip size and the maturity of the fab and node. If you have a Cypress-sized chip and it is yielding >50% in a fab with a D0 of 0.2 DD then you are doing all right. If your yields are 20% then you aren't doing all that well, you should be doing better.

 

[ScorcherDarkly]

Thanks for the information IDC, that puts it in a nice perspective.

 

[thilanliyan]

Great post IDC. I like equations and charts.