Retail AMD Threadripper Delidded, Four Die Present

[Guest]

Der8auer purchased a retail Threadripper 1950X and found what appears to be four active Zeppelin die soldered to the IHS.

Retail AMD Threadripper Delidded, Four Die Present : Read more

 

[mras]

Each die provides 512GB memory access.
Threadripper supports 2048GB memory, just like the EPYC.
For this to be true, combined with the last pic in above article, means that all 4 dies must work, to provide 2048GB memory for the Threadripper(and EPYC).
So, its pretty clear that all dies functions, but some might have damaged cores that means they cant use all 4 dies in all EPYC. But like EPYC, there are lower versions that doesn't provide 32 cores (4 dies), but they still provide 2048GB memory support.
So... for both to be true, a die can be none functional(as in all 8 cores works), and still provide access to the IO/mem, with the working ones.
All it takes, is that AMD make some of the cores working on the dies, and doesn't provide the core power of them, but takes the routing power of them.
In conclusion, which the video support, there isn't a thing such as a dummy die; at tops something AMD classifies as "not fully working die". Its just a way to make more profit out of EPYC CPU.
And yes, of course, all this leads to what the video also ends up at; A 32 core Threadripper will come, if demand is strong enough.
Better question will be; is it possible to re-open some of those dies to use their partly working dies, so you maybe get maybe up to 8 fully working core more pr die?

 

[InvalidError]

Each die provides 512GB memory access.
Threadripper supports 2048GB memory, just like the EPYC.
For this to be true, combined with the last pic in above article, means that all 4 dies must work, to provide 2048GB memory for the Threadripper(and EPYC).

Use 256GB DDR4 LR-DIMMs and you can have 2TB in eight DIMM slots. With EPYC and its sixteen DIMMs per socket, you can use "more affordable" 128GB DIMMs to get there instead.

 

[mras]

256GB pr ram socket, x8 sockets = 2TB mem, or 512GB pr die.
Nothing changed, each die just only controls 2 channel each.
Other articles already pointed that out, eg. from the now sister site An...
Getting a 256GB memory module is gonna be easy, however.

 

[falchard]

Next we will hear they are actually connected but failed to perform well enough for Epyc. Then someone will activate all 4 and discover they are just fine.
However, I think they are just using part of the failed chips for this initial run. Chip failure for 1st generation Zen processors was probably fairly high. Using them as spacers is a good use for them.
I expect failure rate to be better for future generations which will probably change the pricing dynamic. However, I don't think they will use them as dummy processors in Thread Ripper and just go to a spacer of some kind. Makes no sense to put perfectly good chips in a CPU which will not be used.

 

[cryoburner]

Since hearing about the "dummy dies" I have suspected that AMD may be simply repurposing defective processors for them. AMD obviously doesn't want to use the term "defective" to refer to a component of their premium processors though, especially when the limited part they make use of still funtions as needed, which may be why they're being light on details. A big trend throughout the Ryzen lineup is that they have the ability to use many chips that have problems in lower-end parts by simply disabling the cores or cache that have issues, so it stands to reason that if a chip has something wrong with it that prevents enough of its cores or cache from being used in a processor, it could still function as one of these components.

 

[InvalidError]

$999 for a half-dead processor

The spacer dies are supposedly not connected to anything and it doesn't matter if they are dead, blanks or anything else.

If TR uses a different substrate than EPYC as AMD says it does, AMD could have simply put the two cores side by side along the center line and maybe used a smaller socket. Kind of looks like AMD meant to make TR and EPYC interchangeable or at least leave the door open for doing so at a future time.

 

[aldaia]

A real bargain, considering that the full one is more than $4000
Anyway ThreadRipper is not failed EPYCs socket is electrically different. AMD is just recycling dead ryzen dies as "dummy" dies to provide structural stability for the package.

 

[firefoxx04]

AMD using chips with defective cores is nothing new. I have yet to hear about one of them performing worse than they should. The Athlon II / Phenom II days had a bunch of these. The dual cores that were really quad cores always ran fine as dual cores. Same with the 3 cores. I wouldn't worry about it. AMD is still providing a high performing chip for the price.

 

[JamesSneed]

AMD was likely being a bit coy with the "dummy" die statement as they are obviously not used but as we are seeing they aren't dummies but real dies. At least AMD built an easy path to play the one up game with Intel so they can say come out with a 24 core Threadripper if they wanted to one up the 18 Core from Intel.

 

[aldaia]

AMD was likely being a bit coy with the "dummy" die statement as they are obviously not used but as we are seeing they aren't dummies but real dies. At least AMD built an easy path to play the one up game with Intel so they can say come out with a 24 core Threadripper if they wanted to one up the 18 Core from Intel.

Strictly speaking they are dummies, AMD never specified the material they where made. It just happens that the material is failed dies. From an economic point of view makes a lot of sense. Why pay twice: 1) to dispose of wrong dies and 2) to cut dummies? if you can "recycle" wrong dies as TR dummies

 

[41light]

Looking at the reported yield of the chips.. we can understandably see that they eliminate waste by using the chips which do not function correctly.

 

[Giroro]

@MRAS
As far as I can tell Threadrypper supports only 1024 GiB of memory, not 2048.
...but that's from Wikipedia

 

[JamesSneed]

AMD was likely being a bit coy with the "dummy" die statement as they are obviously not used but as we are seeing they aren't dummies but real dies. At least AMD built an easy path to play the one up game with Intel so they can say come out with a 24 core Threadripper if they wanted to one up the 18 Core from Intel.

Strictly speaking they are dummies, AMD never specified the material they where made. It just happens that the material is failed dies. From an economic point of view makes a lot of sense. Why pay twice: 1) to dispose of wrong dies and 2) to cut dummies? if you can "recycle" wrong dies as TR dummies

No argument and when it was said they may have used dummies if they didn't have enough bad chips to fill in day one.

 

[InvalidError]

No argument and when it was said they may have used dummies if they didn't have enough bad chips to fill in day one.

Since ThreadRipper launched several months after Ryzen and a few months after EPYC, AMD likely had plenty of dead or sub-par dies to begin ThreadRipper production with - enough so that AMD had to include 'landfill' as an alternate destination for them.

Had AMD been desperate for dummy/bad dies for TR, it could have simply delayed the R3's launch - those dies would have been more than twice as valuable as TR spacers than as working chips of their own.

 

[IceMyth]

"Mounting dies to a substrate is like any other manufacturing process; there can be defects. It's possible that the Threadripper processors are simply EPYC processors with die that weren't successfully mated to the substrate. Conversely, it's also possible that AMD is using defective die for the fillers, but in either case, it's doubtful that the company is wasting functioning silicon."

So jumping the gun, and saying they are dead or fault chips! Wait a new Intel CPUs are on the corner, what better time to advertise!


"We've followed up with AMD for additional clarification and will update as necessary."

So, your review is like executing someone before the court judge if he is guilty or not!

Also, I don't understand. People complaining about 2 dies that are not functionals for a reason or another. YET they can't afford to get one or they are not interested in getting one!


For me, I am getting the 1920x/1950x even if they were faulty as long as I get the Performance vs Money I expect!

 

[deadsmiley]

It does make a lot of sense from a manufacturing point of view to use the bad dies at spacers.

1) No need to precision machine a spacer.
2) Same thermal characteristics.
3) No need to empty the trash when the waste bin is full of failed dies.

 

[Tex61]

Why not just fire it up and measure the temperature at each die or view it with an IR camera? It will be obvious which die are powered / non-powered.

 

[falchard]

Because AMD doesn't use cheap thermal paste. If you de-lid an AMD CPU, it no longer works.

 

[Tex61]

Because AMD doesn't use cheap thermal paste. If you de-lid an AMD CPU, it no longer works.

Thanks. Wasn't aware of that.