AMD Threadripper confirmed to have 4 actual Zepplin dies.

[Doobeedoo]

I'd expect they're defective dies. Also would be odd kinda to dilude TR and Epyc platform no. And not sure how supportive even more cores would be. 

I doubt they'll make that move. 

[Ithanul]

On ‎9‎/‎15‎/‎2017 at 3:39 PM, pas008 said:

Considering they said yields were great

And cheaper to make tr and epyc in one single design and make the socket disable half in not saying anything just it could be a possibility

And unlocking wasn't always guaranteed to have great results

Think it was pII 550 into pII into quad

Double the cores lol

That was the fun days.  Hehe, I remember buying a duo, but it was a locked down quad.  Was sweet unlocking it to quad by just switching one BIOS setting.     You where not guaranteed two extra cores all the time though.

[SpaceGhostC2C]

10 hours ago, mariushm said:

Why cut dies out of a perfectly good wafer when you can just re-use dies that have failed binning (for example dies that can't even be sold as Ryzen 3 1200) ?

 

My guess is they're simply dies that have some critical errors like maybe several bad spots in the L3 cache memory (which would make them unable to get even 4 MB of  cache memory to sell these as Ryzen 3) or maybe some failure in the chip's ddr4 controller or other such critical points.

 

Well, it certainly makes sense to recycle failed dies for this purpose. However, there is no reason why dies will fail and succeed into different bins in the exact proportions needed for the intended supply.

In the limit, if we all decided to buy Threadripper and only Threadripper, either the failure rate is 50% or they would need to use nerfed working dies or some other spacer.

[vanished]

Perhaps the most concerning thing about all this is what Linus brought up on the WAN show.  It was fairly reasonable to assume someone would look into this sooner or later, and no one really cares that they're real cores that just don't work - it's a perfectly reasonable and common thing to do.

 

So AMD had absolutely no reason to lie about this, and yet, they did... why?  If they're willing to lie about such a stupid thing as this, imagine what important things they will lie about in the future.

[Jito463]

On 9/15/2017 at 2:08 PM, kiska3 said:

Here are the die shots from der8auer

So which ones of the 4 are active?

I can't even........that, that MONSTER!  That BUTCHER!!!

 

I think I'm going to be sick.

[HalGameGuru]

We see SOME etching, do we know if this is actual usable transistor etching? 

[IKnight]

Maybe they wanted to keep it a secret incase they could somehow activate them and thus catch Intel with their pants down? 

 

[ravenshrike]

On 9/16/2017 at 3:57 AM, Bananasplit_00 said:

wow no shit its 4 dies, did anyone actiually think they would cut un-used silicon just for dummies?

Well apparently Linus did.

On 9/16/2017 at 2:35 PM, Ryan_Vickers said:

Perhaps the most concerning thing about all this is what Linus brought up on the WAN show.  It was fairly reasonable to assume someone would look into this sooner or later, and no one really cares that they're real cores that just don't work - it's a perfectly reasonable and common thing to do.

 

So AMD had absolutely no reason to lie about this, and yet, they did... why?  If they're willing to lie about such a stupid thing as this, imagine what important things they will lie about in the future.

They never lied. They solely stated there's no working transistors on the spacer dies, probably because they fully burnt out faulty dies to use as spacers. I stated that this was almost certainly the case way back when even before derbauer did his first delidding of TR. Even at only 1 out of every 100 dies being unusable for any other form of Ryzen, they probably have enough supply for TR chips. Worst case scenario, they take the absolute worst of the 1200 dies and re-purpose them as spacers.

[vanished]

1 minute ago, ravenshrike said:

Well apparently Linus did.

They never lied. They solely stated there's no working transistors on the spacer dies, probably because they fully burnt out faulty dies to use at spacers. I stated that this was almost certainly the case way back when even before derbauer did his first delidding of TR. Even at only 1 out of every 100 dies being unusable for any other form of Ryzen, they probably have enough supply for TR chips. Worst case scenario, they take the absolute worst of the 1200 dies and re-purpose them as spacers.

My understanding was that they had led us to believe they were basically just a block of metal.  I'd have to go back and read the original articles.

[Liltrekkie]

2 hours ago, Ryan_Vickers said:

My understanding was that they had led us to believe they were basically just a block of metal.  I'd have to go back and read the original articles.

Here's the quote from overclock3d: "These two extra dies are not active or even contain working transistors, they are blank and as such are not "wasted Ryzen CPU dies"" 

 

So no, they did NOT lie, they clearly said they did NOT have working transistors, 

 

They said Ryzen has 80% yields, I suspect they are using the 20% failed ones for here, it's their own waste recycling program. It's a win win. Hardly any money lost. 

[vanished]

Just now, Liltrekkie said:

Here's the quote from overclock3d: "These two extra dies are not active or even contain working transistors, they are blank and as such are not "wasted Ryzen CPU dies"" 

 

So no, they did NOT lie, they clearly said they did NOT have working transistors, 

 

They said Ryzen has 80% yields, I suspect they are using the 20% failed ones for here, it's their own waste recycling program. It's a win win. Hardly any money lost. 

"they are blank" and "[they are] not wasted Ryzen CPU dies" seems like it's pretty much exactly the opposite of what's been discovered.  They are not blanks, they have (failed) transistors, and these were obviously failed dies, unless AMD has taken up the hobby of printing failed transistor dies just for fun.

[Drak3]

@Ryan_Vickers It's possible that blanks have been used in some production runs in the scenario that there are too few defunct dies to use as spacers. Or, AMD themselves never said that the die was blank.

 

Pyo.

[vanished]

2 minutes ago, Drak3 said:

@Ryan_Vickers It's possible that blanks have been used in some production runs in the scenario that there are too few defunct dies to use as spacers. Or, AMD themselves never said that the die was blank.

 

Pyo.

Well apparently they did literally say "they are blank "

[Liltrekkie]

2 minutes ago, Ryan_Vickers said:

"they are blank" and "[they are] not wasted Ryzen CPU dies" seems like it's pretty much exactly the opposite of what's been discovered.  They are not blanks, they have (failed) transistors, and these were obviously failed dies, unless AMD has taken up the hobby of printing failed transistor dies just for fun.

I believe blank means something different in the CPU die business, I had a relative who made dies, for my birthday I was given a "blank" piece of silicon wafer, that was defective, but you could obviously see transistor imprints, when asked, she said that the term blank meant they never went through the transistor phase, but had transistor outlines and traces, but were found to be defective before transistors were put in. 

Now that was over 13 years ago, but it makes me wonder if a similar thing is happening here.Maybe blank to them isn't blank to us. Another technical term maybe?

[Drak3]

1 minute ago, Ryan_Vickers said:

Well apparently they did literally say "they are blank "

Without a link, and without desire to google or bing anything, I cannot verify if it was AMD saying that, or a reporter saying that on their behalf.

[vanished]

Just now, Liltrekkie said:

I believe blank means something different in the CPU die business, I had a relative who made dies, for my birthday I was given a "blank" piece of silicon wafer, that was defective, but you could obviously see transistor imprints, when asked, she said that the term blank meant they never went through the transistor phase, but had transistor outlines and traces, but were found to be defective before transistors were put in. 

Now that was over 13 years ago, but it makes me wonder if a similar thing is happening here.Maybe blank to them isn't blank to us. Another technical term maybe?

Perhaps.  That makes a lot of sense actually and could bridge the gap in explaining this dependency.

[Liltrekkie]

5 minutes ago, Ryan_Vickers said:

Perhaps.  That makes a lot of sense actually and could bridge the gap in explaining this dependency.

That was my thought, if it's true, AMD has made a very efficient product in terms of manufacturing and waste management. AMD absorbs the costing using those failed dies in TR. I have to admit thats a very clever solution if that's actually the case. 

[vanished]

1 minute ago, Liltrekkie said:

That was my thought, if it's true, AMD has made a very efficient product in terms of manufacturing and waste management. AMD absorbs the costing using those failed dies in TR. I have to admit thats a very clever solution if that's actually the case. 

If that is the case though, that probably should have been explained in more detail at the time, as it could have avoided this whole wave of confusion.  I am reminded of the Intel slides that said threadripper/ryzen is just "glued together" desktop dies and everyone took this to be an insult when in fact it may actually be a technical term.

[ARikozuM]

On 9/15/2017 at 3:12 PM, Misanthrope said:

Not sure why they put modules that are basically off, maybe it has something to do with the interposer and getting all lanes? I.e. the core is defective to actually do calculations but it can still interconnect for the pci-e lanes and such.

Interesting question... It would be quite astounding if they found a way to have dead dies work in a "storage" format where it holds data for the other dies as needed. 

[Skanky Sylveon]

18 minutes ago, Ryan_Vickers said:

If that is the case though, that probably should have been explained in more detail at the time, as it could have avoided this whole wave of confusion.  I am reminded of the Intel slides that said threadripper/ryzen is just "glued together" desktop dies and everyone took this to be an insult when in fact it may actually be a technical term.

True, not everyone knows all of the technical terms that's used by the engineers, and I have a felling that AMD knows this.  

Whether you believe that AMD was dishonest about the dies is up for debate, the fact is that AMD failed to properly explain themselves, and that led to misinformation being spread.

[Zodiark1593]

Could also be an insurance policy in the event that Intel engages AMD in a core count war. Intel wouldn't have 16C TR to beat, but the possibility of a 32C TR to contend with.

[S w a t s o n]

2 hours ago, Zodiark1593 said:

Could also be an insurance policy in the event that Intel engages AMD in a core count war. Intel wouldn't have 16C TR to beat, but the possibility of a 32C TR to contend with.

They already are at war, Intel 18 core is the first shot.

[MorganO]

5 hours ago, Liltrekkie said:

I believe blank means something different in the CPU die business, I had a relative who made dies, for my birthday I was given a "blank" piece of silicon wafer, that was defective, but you could obviously see transistor imprints, when asked, she said that the term blank meant they never went through the transistor phase, but had transistor outlines and traces, but were found to be defective before transistors were put in. 

Now that was over 13 years ago, but it makes me wonder if a similar thing is happening here.Maybe blank to them isn't blank to us. Another technical term maybe?

that seems plausible

[dragoon20005]

No surprise that there might be a possibility of a 32 core Treadripper on the TR4 socket since the current 12 core and 16 core Treadripper are Binned parts of the Epyc CPUs

 

Sure the number of PCI-E lanes will have to be reduce from 128 to 64 lanes, but 64 lanes is way plenty for 3 to 4 way CrossFire GPUs and stupid amount of PCI-E based cards or NVME SSDs

 

RAM channels will have to be reduced from 8 to 4 channels. I don't think any consumer will have a need for 64GB of RAM on a workstation/desktop let alone 128GB.

 

Some other server grade features will need to be cut off as well like ECC memory support and dual socket support

 

if you do need multi socket support, then you should be using Epyc platform.

 

Power users like JaysTwoCent, Austin Evans, Bitwit and Paul will love all the 32 cores for video rendering at a lower cost which put all the Intel i9 and Xeons to shame.

 

Sure Ryzen and Treadripper still have teething issues with RAM support, give it 6 months to a year and you will see Ryzen CPU support 3200MHz and above.

 

 

 

Etching is the correct term here as it pretty much making an imprint of the tracers on the silicon die for the connections.

 

The reasons for which the die can fail include:

 

Imperfections present on raw initial silicon wafer (for example: too contaminated silicon crystal)

Imperfections added in the pre-production wafer's handling (for example: physical defects produced during native oxide etching)

Failures during active devices patterning (for example: insufficient/excessive doping, defects in insulation layers, improper alignment)

Failures during metals routing (for example: short/open circuits, disconnected vias)

 

There are two general factors which can lower the yield: overusing the thermal budget and failures in chemical reactions. The thermal budget is the amount of heat the die can be exposed to without being damaged (usually determined experimentally). I know nothing about chemical reactions, but I know that even slightest error in chemicals constitution can ruin not just a single die, but the entire batch.

It is very hard to obtain high yields in sub-micron processes, therefore Intel came with the "Copy-Exactly" strategy: all the controllable factors are the same across all its fabs. This ensures that once the primary fab gets to high yields, the secondary fabs will get there in the shortest possible time.

The yield numbers are trade secrets. However, it is widely known fact that it becomes more and more difficult to get high yields which more advanced technologies. The number of 60% yield seems too low to me - this might be an initial yield when the technology is not mature enough, but this is too low for large scale manufacturing.

[CostcoSamples]

On 9/15/2017 at 9:24 AM, GatioH said:

So isn’t threadripper just EYPC with 2 dummy dies?

well not exactly, but Threadripper is based on EPYC.  I don't think AMD ever intended to make Threadripper until relatively recently.  EPYC has more PCIe lanes for starters, and other stuff too.