AMD's James Prior Clarifies Threadripper's "Dummy Dies"

[Raevenlord]

Much has been said regarding AMD's Threadripper CPUs, particularly when it comes to how they are manufactured. At first, we thought Threadripper was actually EPYC in disguise, due it having what appeared to be four full-fledged 8-core modules - the same design as AMD's server-bound 32-core EPYC chips. The presence of gold-plating under all four dies seemed to confirm that these were in fact four full Threadripper dies, instead of two dies and two spacers (as AMD's statements led us to believe) for even IHS pressure on the four dies, instead of the uneven pressure that would result from the chip only having two physical dies present.





It seems the truth is, as always, somewhere in-between. On Twitter, James Prior came forward to shed some light on the issue, clarifying what exactly is going on and justifying AMD's usage of "dummy" and "inactive" nomenclatures. First up, no, readers: there is virtually no way to reactivate those unused dies. As James Prior himself said, "Threadripper is not a Epyc processor. Different substrate, different dies. 2 dies work, other 2 have no path to operation. Basically rocks." Those are some expensive pieces of rock, for sure. However, AMD's choosing of "dummy" instead of inactive seems correct here: "(...) exactly why they're not described as inactive, but dummy. Doesn't matter if they were dead, or active, they're not going to work."



Since inactive implies the capability of reactivating, yeah, dummy sounds about right. AnandTech's Ian Cutress' responded to James Prior's post with a general workflow of AMD's Threadripper dies, based on prevous AMD information and these statements from James, basically embodying what is going in our minds. See it in the image below.



View at TechPowerUp Main Site

 

[Chaitanya]

So basically AMD is getting rid of trash CPU left after binning to be used as dummy dies for TR. Thats a good way to reduce waste in factory.

 

[Totally]

It's a common sense thing which baffles me as to why this eludes some individuals.

 

[RejZoR]

Trickling down chips based on their quality is nothing new, everyone does it (instead of throwing still usable chips away, they use them for lower end products). Considering the way how they have them arranged, they are kinda forced to do this with dummy dies to make cooler pressure even on the IHS and consequently on the dies and CPU interposer. It's nothing shady or weird, it's engineering common sense.

 

[theGryphon]

When one picture is worth more than one thousand words...

 

[noname00]

"Threadripper is not a Epyc processor. Different substrate, different dies." - based on this there is no relation between Epyc and Threadripper.
As there are no (official) plans for a 32 core Threadripper, why did amd use 4 dies instead of 2? why do they need the spacers?

 

[Basard]

When one picture is worth more than one thousand words...

But when do the dies get designated for Epyc chips? Oh nevermind, I guess it's 'different'... I suppose reading helps.

 

[FR@NK]

As there are no (official) plans for a 32 core Threadripper, why did amd use 4 dies instead of 2? why do they need the spacers?

My guess would be so they could have a chip thats the same size as epyc and uses the same heatspreader. This simplifies motherboard and heatsink design.

 

[blibba]

My guess would be so they could have a chip thats the same size as epyc and uses the same heatspreader. This simplifies motherboard and heatsink design.

Yep. Using an EPYC design that was already in the works allowed them to launch a HEDT platform with very little R&D, and keeping everything standardised gives them more options later, too.

 

[Vya Domus]

"Threadripper is not a Epyc processor. Different substrate, different dies." - based on this there is no relation between Epyc and Threadripper.
As there are no (official) plans for a 32 core Threadripper, why did amd use 4 dies instead of 2? why do they need the spacers?

I am more or less convinced there will be higher core count TR CPUs coming in the future. TR might not be exactly the same to Epyc in terms of electrical connections but there is clearly a relationship between them in terms of design to allow for easier modifications should there be the need.

 

[KarymidoN]

I am more or less convinced there will be higher core count TR CPUs coming in the future. TR might not be exactly the same to Epyc in terms of electrical connections but there is clearly a relationship between them in terms of design to allow for easier modifications should there be the need.

AMD EPYC (32/64) has 128PCIE Lanes...
AMD TR (16/32) has 64PCIE Lanes...

I THINK that for more than 16/32 Cores/Threads they need more Lanes, in order to maintain the TR4 Socket boards not that expensive they had to limit the cores, JUST LIKE intel anouced on their new Chipset Z390 (Not sure if is x390 or z390) that will be only for 8c/16t cpus, clearly due to PCIE Lanes and other limitations. But i'm just guessing at this point. (sorry my bad english)

 

[Fujikoma]

Unless there is a diagram of the electrical layout for the pin connectors to chip, then the possibility exists for a higher core count among 4 dies. Allowing up to 4 dies would give AMD a lot of flexibility in competing with Intel and also allow them to produce more powerful chips on the same socket, so it would seem odd to remove that possibility. I think they're holding out due to thermals and an improved die or node shrink may allow them to go that route.

 

[Dimi]

It's a common sense thing which baffles me as to why this eludes some individuals.

I thought half working dies would trickle down to the lower end ryzen's? You'd have to produce a LOT of non working dies to keep up with threadripper production. Seems like an excessive amount of waste of resources.

 

[KarymidoN]

I thought half working dies would trickle down to the lower end ryzen's? You'd have to produce a LOT of non working dies to keep up with threadripper production. Seems like an excessive amount of waste of resources.

actually the 14nm process is new and a lot more complicated than the previous 32nm (Bulldozer/Vishera) so you can expect a lot of failing/not working well cores, and they can always complete the TR chips with just empty dies (not faulty, just blank silicon), if they Release zen2 on 14nm Again yall see the diference, but amd is probably going to ZEN2 on 7nm....

(sorry my bad english)

 

[ShurikN]

I thought half working dies would trickle down to the lower end ryzen's? You'd have to produce a LOT of non working dies to keep up with threadripper production. Seems like an excessive amount of waste of resources.

Threadripper isn't a high volume product, therefore a huge amount of dead/dummy dies is not required.

 

[OSdevr]

I thought half working dies would trickle down to the lower end ryzen's? You'd have to produce a LOT of non working dies to keep up with threadripper production. Seems like an excessive amount of waste of resources.

I'm sure they do, but AMD sells a lot more Ryzens than threadrippers. They also sell EPYCs which take 3-4 flawless dies. Even with Zeppelin's good yields there's plenty of dead dies available for TR because they don't sell as many of them.

 

[RejZoR]

Depending on demand, they probably have to purposely create some dummy dies. They have very good yields in general, around 80% iirc. So, 20% of entirely unusable chips (if we assume they count it as full working/full dead). Each working CPU requires 2 working and 2 dummy dies. I don't think they have enough "natural" dummies...

 

[OSdevr]

Depending on demand, they probably have to purposely create some dummy dies. They have very good yields in general, around 80% iirc. So, 20% of entirely unusable chips (if we assume they count it as full working/full dead). Each working CPU requires 2 working and 2 dummy dies. I don't think they have enough "natural" dummies...

They just might. I'm going to quote myself from another thread here:

The Zeppelin die is used in Ryzen, TR and EPYC, but only TR and possibly some EPYCs need dummies. Further TR is an HEDT product, Ryzen sells much better. Say AMD sells 2 Ryzens and 1 EPYC for every TR they sell. Say all 4 dies in the EPYC are good. That's a total of 10 dies. But they don't have to start using good Zeppelins as TR dummies unless their yield is higher than 80%, not 50%!

We've never seen something like this before. This sort of thing will only happen with MCMs in which all the dies are identical. It doesn't occur in traditional GPUs or CPUs because you can't recycle a bad part of a chip into another package.

 

[eidairaman1]

I thought half working dies would trickle down to the lower end ryzen's? You'd have to produce a LOT of non working dies to keep up with threadripper production. Seems like an excessive amount of waste of resources.

Ok dies that are totally junk are used, aka dead dies. Typically when a processor is made they bin chips that can clock the highest or are fully functioning as their top model, then so forth. At least they are repurposing them for structural support.

 

[sergionography]

Depending on demand, they probably have to purposely create some dummy dies. They have very good yields in general, around 80% iirc. So, 20% of entirely unusable chips (if we assume they count it as full working/full dead). Each working CPU requires 2 working and 2 dummy dies. I don't think they have enough "natural" dummies...

With 80% yield and threadripper making up the top 5% of that 80% means for every 80 ryzen chips they use the best 4 for TR. And for every 4 TR chips there are 20 Dummy(faulty) chips to use.

 

[qubit]

It would make sense to only use faulty dies that have no hope of making a working CPU as the dummys and I'm pretty sure that's what they do. If they were to run out, then a blank piece of silicon or perhaps glass of that size and shape would do.

 

[eidairaman1]

It would make sense to only use faulty dies that have no hope of making a working CPU as the dummys and I'm pretty sure that's what they do. If they were to run out, then a blank piece of silicon or perhaps glass of that size and shape would do.

It would be sweet if they started using good dies and that there is more to the tr/epyc platform than we truly know.