Low Yields for Cell

[Itanium_Fanboi]

It looks like PS3 will never see the light of the day due to low yields.

http://www.xbitlabs.com/news/cpu/display/20060713074825.html

Maybe they should use Itanium2 instead.

 

[mjp1618]

It looks like PS3 will never see the light of the day due to low yields.

http://www.xbitlabs.com/news/cpu/display/20060713074825.html

Maybe they should use Itanium2 instead.

Ohhhhh great topic.... let's talk.

Sarcasm?

I thought you were not a big fan of either of the two processors...

 

[voxel]

A game console needs an Itanium2 like I need a hole in my head.

Funny thing is: The Cell might outlive the Itanium2.

 

[Wombat2]

The 3 GHz Woodcrests are only a shade slower than the Itanium 2 ... compare them on www.spec.org 8)

 

[K8MAN]

The 3 GHz Woodcrests are only a shade slower than the Itanium 2 ... compare them on www.spec.org 8)

Ya and they dont require a billion transistors to be that fast either.

 

[enewmen]

If the Itanium2 is SOO much better than Pentium 4 OR Core 2 (must be so much better if it costs so much more), WHY doesn't Intel make the Itanium THE ONLY core of choice?
I remember 10+ years ago when the Dec Alpha was way ahead and it's 64-bit. So the very high-end work was done on a high-end 64-bit chip/Unix while the mainstream work was done on 32-bit Pentiums/ Win95.
Now most desktop chips are 64-bit. So why even bother having a Core2 and a Itanium2 at the same time?
No software for the Itanium2 instruction set? The Windows Server 2003, 64-Bit Datacenter Edition already supports the Itanium2. Just make an x86-XP emluator like Apple did. Now the Mac OS7-8-9 stuff can run on x86-OSX.

If the x86 is too old, kill it and kill the bloated legacy support. If the Core2 is so much better performance per watt and cheaper, kill the Itanium2 and make a heavy-duty Xeon. I do not see any reason for both.
Am I missing something major?

BTW, from what I read over the months, I do not see the Cell as any threat to Intel/AMD.

 

[Heyyou27]

They would have been better off with this in the PS3:

 

[mpjesse]

Hi,

We've discussed this already. Thanks,

-mpjesse

 

[TechnologyCoordinator]

They would have been better off with this in the PS3:

Nah, at least dual core:

But my guess is that cell would like more like this:

 

[Heyyou27]

Good point.

 

[Gary_Busey]

When they first announced this thing, it was like the second coming. Now they can't even make enough of them.

 

[FITCamaro]

Unfortunately mice tend to die when you try to shove 9 of them in a box the size of maybe two. Plus they need to eat.

Details.........I know........details.......

Now if we stuck a male and a female in each wheel, then they could have sex, eat their young to survive, and keep on going for a while...........

 

[nottheking]

Like I'd be surprised about this sort of "news" whatsoever. I think it was pretty clear, given the incredible size of the chip, that yields were going to be horrific. I remember the original claim of a 4.6GHz core speed, with all eight SPEs. Now they're looking at 2.8GHz with 6 SPEs as the best they can do.

Personally, I had thought that Sony was going to wait until it could produce a shrink of the chip to 65nm, if not 45nm, though the latter would've delayed things into 2008 if not 2009.

I've not been impressed by the claims brought forth for the PS3. To be honest, I can't help but think that all of these chips, from the technology to the games and so on, will result in the PS3 garnering a laughable market share compared to its predecessors, and possibly even the Xbox 360 and Wii as well. At best, I can only see them selling 30 million units by the end of 2008, if that; even with the price at a whopping $600US, I forsee significant shortages throughout the console's lifespan.

 

[RichPLS]

Dang... if the PS3 cell CPU yeilds get any lower... they will approach the levels of AMD's 65nm chips! :mrgreen: :mrgreen: :mrgreen: :tongue: :tongue: :lol:

 

[Heyyou27]

Dang... if the PS3 cell CPU yeilds get any lower... they will approach the levels of AMD's 65nm chips! :mrgreen: :mrgreen: :mrgreen: :tongue: :tongue: :lol:

A little fanboi no?

 

[gr8mikey]

Dang... if the PS3 cell CPU yeilds get any lower... they will approach the levels of AMD's 65nm chips! :mrgreen: :mrgreen: :mrgreen: :tongue: :tongue: :lol:

Be careful with comments like this. It could result in BM showing up and posting 5 pages of BS.

 

[Mind_Rebuilding]

No no sarcasm -- On the technical merit, I do find both processors very good. On the Itanium, I thought the implementation was poor. It is a pure 64 bit solution, with completely new instruction sets --- it was Intel's attempt to compete in the high end --- and shows how horrid Intel is at collaboration. The project should have been killed after 6 months.... now after billions of dollars, it is still around, hardly any market, and no real infusion of software or support from the industry. Time to lay it to rest.

On the cell side, my opinion is that it is way ahead of it's time --- I have study a little, and just a little, to formulate this opinion and I base alot of that off of Fred Weber's comments in his interview.

No, why I appreciate this topic is that the Cell has few features unqiue to it from a manufacturing side of the equation--

a) it is large
b) it is mostly logic, it is cache lean.
c) it is being built on the 65 nm process by IBM

This says something..... what other design is large, cache lean, and will eventually be built on an IBM (or IBM like) process?

What is the die size for Cell chip? :?:

 

[Action_Man]

~230mm2.

 

[Action_Man]

Well if you want to be exact, a quick google turned up 235mm2.

I expect the K8L to peg in around 200-250 mm^2 on 65 nm

I assume this is for quad core?

 

[Action_Man]

Ugggh, action_man had it about right, 221 mm^2 according to this link:

Thats the DD1 revision, DD2 and DD3 have a bigger and better PPE and is 235mm2.

 

[Action_Man]

Ya, I follow the console scene pretty closely as it has a major effect on the PC world and vice versa. This is true and they have rather aggressive clocks, they also can't sell them for anything cheaper, celeron style.

 

[xsandman]

Yeah, but I am skewed on my scale and calculations --- right now a 512 k AM2 processor is running at 183 mm^2 (the 1 meg verison is 230 mm^2), so going to 65 nm that would be 90 and 115 respectively assuming a simple shrink. Now quad core comes along which has about the same cache leaness as a 512 k AM2 so double that we are now back to 180 mm^2, add in the extra transistors for the HT, IMC, decoders, and all the utensils, napkins, toilets, and kitchen sinks AMD is throwing in and I am not so afraid to call a K8L in that ball park....

Again, my concern is if the Cell is yielding poorly I am wondering what is wrong it is either design (too logic heavy) or process or a combination of both. K8L will be logic heavy as well, similar die size and even more prone to random particles (dimensions is smaller, smaller particles that were not an issue become an issue).
Jack

If Cell is ~235mm^2 die size and AM2 is ~180mm^2, this isn't THAT huge of a difference that the yields would be THAT much different if it was only due to process. It AMD's yield are REALLY healthy, and the die size is somewhere in the region of Cell, and Cell is THAT unhealthy, i would think there is a different reason for the yield loss.

This is no big deal really, Intel will have similar issues once the integrate 4 cores into the same die --- this, btw as you likely know, is the reason for the dual core dual package approach.

I do disagree with there being much benefit by using dual core/dual package other than being simple to implement. I have given you my opinion on this previously and am still waiting for a response. :lol:

 

[Action_Man]

Yeah, but I am skewed on my scale and calculations --- right now a 512 k AM2 processor is running at 183 mm^2 (the 1 meg verison is 230 mm^2), so going to 65 nm that would be 90 and 115 respectively assuming a simple shrink. Now quad core comes along which has about the same cache leaness as a 512 k AM2 so double that we are now back to 180 mm^2, add in the extra transistors for the HT, IMC, decoders, and all the utensils, napkins, toilets, and kitchen sinks AMD is throwing in and I am not so afraid to call a K8L in that ball park....

Ya but I'd only expect the quad cores to be server and FX's. The dual cores should be a nice size.

This is no big deal really, Intel will have similar issues once the integrate 4 cores into the same die --- this, btw as you likely know, is the reason for the dual core dual package approach.

They'll be doing it on 45nm and I think dual should still be more popular.

 

[xsandman]

You may know more than I on this topic per what we have already discussed.

There are several modeling methods for determining yield and how yield is affected. I agree with what you imply --- it is not so straight forward that you can say the large die is why Cell is yielding so low. There are upteen other factors that drive this.

What I am getting at is more of the concept of why? And to spur some speculation. One is just random defectivity -- which is influenced by two primary factors --- die size (greater probability a random will land and kill the die) and node size (i.e. smaller node now means small otherwise benign particles become killers).

Asdie from that, there is a gread deal in the design and layout. How close together are the transistors, can the contact land and hit the contact points on the source, gate, drain? These are questions we cannot answer so it is really speculation.

What can generally be said, however, is larger die yield more poorly than smaller die.... large logic areas yield more poorly than smaller logic area's.

This was actually my point.... if you look at a K8L die plot (not die shot, as it does not exist yet -- this should get BM going).... it is cache lean -- but cache is the easiest to make redundant, so that trick is diminished. Cell is cache lean, lots of logic ... so I am looking K8L -- large die, large logic, less cache --- low yielding?

Hard to tell.

I definitely dont disagree with anything that you're saying. I am just debating on how close is the relationship between the Cell to the AM2, and if they do have a close relationship (lean cache, lots of logic), then how healthy are the yields on the AM2? If the AM2 yields are healthy, then what is cell doing wrong? If the yields for the AM2 really aren't that healthy, then AMD may be in a lot of trouble.

But either way, it would benefit them greatly to figure out the yield issues. It may be that the yield is only unhealthy becuase they are trying to bin at like 3.2GHz (or was it 3.5)? I remember reading something that the yields may be healthier at 2.8GHz. If this is the case, it would not be surprising as the AM2s do not sell nearly that high, and maybe it has more to do with binsplit than actual yields.

 

[xsandman]

They'll be doing it on 45nm and I think dual should still be more popular.

Kentsfield will be on 65nm (although this will be 2 die in 1 package to get 4 cores), So they will do 4 cores on 65nm, but it probably wont be as large of a volume until 45nm.