Threadripper Delidding Uncovers EPYC in Disguise: 32-Core Desktop CPU in the Works?

Megalith

24-bit/48kHz
Staff member
Joined
Aug 20, 2006
Messages
13,000
Overclocker Roman Hartung (der8auer) discusses his delidding of a Threadripper CPU in a new video: he confirms that AMD is, in fact, using four functional Ryzen dies to create a single Threadripper chip. Being that AMD’s EPYC server processors are made the exact same way, der8auer speculates that the company has plans for a 32-core desktop processor (the choice of using four dies was intentional).
 
The link to the forum is incorrect.

I have to wonder if these dummy cores are defective or not.

If they aren't, then how much do they really cost to AMD that they can use them as structural support, instead of a plain piece of silicon?
 
The link to the forum is incorrect.

I have to wonder if these dummy cores are defective or not.

If they aren't, then how much do they really cost to AMD that they can use them as structural support, instead of a plain piece of silicon?
Yeah, it wouldn't matter if it cost them a buck, it would still be a big loss since they could be sold as a real chip.
I think its very unlikely these are not garbage.
Unless you can't test silicon without building the whole thing... you CAN test it before right?
If you have to build the whole thing before testing it, then its not a matter of choice I suppose.
 
This is always why I like to wait 8-12 months before buying a new CPU type as the initial batches can be hit and miss but after a while chances are all the dies in your CPU are probably perfectly good ones and just disabled.

Kind of like the old Phenom II tri-cores. Initial batches were quads with a bad core but months later when the silicon process was sorted they were all perfect quads and ripe for tinkering.
 
I wish he would have tested the CPU after the delid. I mean have we figured out what 2 dies are enabled yet??
 
Now if only he would take apart the i9 cpu and figure out why there are two substrates on it.
 
imagine the scrambling intel would have to do if they did come out with a 32 core on TR4.
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...
 
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...

prosumer applications

like the OG Titan
 
Intel to AMD: WE HAVE 18 CORES MUHAHAHA!!

AMD to Intel: Yikes, you got us. When is that release date?

Intel to AMD: TBD

AMD to Intel: Oh, ok, please let us know when you are going to 'own' us.

AMD to Engineers: psst, get the 32 core ready. Lol, they actually think they are going to get the crown back. This is going to be Troll-tastic.

In all seriousness:

They would likely have a 24 (6x4) and 32 (8x4) CPUs.

I would be surprised if they then have another version of the 16 (4x4) as well, just to have something to do with those dies.

Of course, that likely has higher latency.

I could definitely see myself grabbing the 32 core and making a VM system for my house. Be better than maintaining the multiple PCs I have now. Just need some way to make terminals over the network.
 
Last edited:
prosumer applications

like the OG Titan
Exactly. But like i said thats less than 1% market. Dont think intel is going to rework it's strategic plan for the potential to lose a very small segment. Also of that 1% Who needs to upgrade, who is willing to cross party lines, etc.

Not saying it wouldn't be pretty cool to see and a feat, just that I don't think it will have much of an impact on the market, at least in the short term.
 
Exactly. But like i said thats less than 1% market. Dont think intel is going to rework it's strategic plan for the potential to lose a very small segment. Also of that 1% Who needs to upgrade, who is willing to cross party lines, etc.

Not saying it wouldn't be pretty cool to see and a feat, just that I don't think it will have much of an impact on the market, at least in the short term.

intel isn't going to let amd have that kind of platform and not have an answer lets not kid ourselves here.
 
From what I could tell only two of the dies were electrically connected to the package

Possibly a 32 core part could be released but the ones out today cant be unlocked
 
I could definitely see myself grabbing the 32 core and making a VM system for my house. Be better than maintaining the multiple PCs I have now. Just need some way to make terminals over the network.

There's probably some open source thin client manager software out there. Thin clients are usually a lightweight OS (XP embedded / etc) with a configurable launcher for a remote desktop app. Now a days, a Pi3 could probably do it.
 
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...

TR would be more workstation orientated at that point. Anyone using a TR isn't paying that premium for quad core type scenarios, they are using them for video editing etc etc...


And I told you lot it's likely they are just Epyc with disabled dies, it makes no sense to have a separate production, nor to Indium solder the 'deactivated' dies.
And no I don't expect you can re-enable them as they'll be laser cut.

In future, we'll see. This might be AMDs ace up the sleeve until Zen2 late 2018-early 2019.
 
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...
threadripper with 16 cores clocks better than ryzen with 4 cores, the ccx and different module seem to mitigate that compared to a monolithic die like intel does with their i9, my guess the 32 core would clock almost the same as the 16 TR, and the need for it well, staying at the top when the 18 core intel cpu comes out with 2Ghz clock...plays with mind share and appeal to enthusiasts.
 
Yeah, it wouldn't matter if it cost them a buck, it would still be a big loss since they could be sold as a real chip.
I think its very unlikely these are not garbage.
Unless you can't test silicon without building the whole thing... you CAN test it before right?
If you have to build the whole thing before testing it, then its not a matter of choice I suppose.

You can't test the full die functionality until it's complete.

At a few different points in the process you can order resistance measurements on dummy structures in the scribe lines - the correlation between performance at those points and end of line is pretty weak. If the wafer is toast you'll see it, but you won't know anything about the binning of the individual die.
 
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...

Well there is very little consumer need for 16 cores right now either.

The current ThreadRippers sales are mainly epeen sales. SOOooo... why not I bet if AMD released a limited number of 3900x 32 core thread rippers at $2500 a pop they may well sell better then expected. Its not like it would cost AMD much to add the sku anyway... it would pretty much just be an epic with a TR socket instead of the SP3 epic layout.
 
You are seriously underplaying how much of a cost and hassle savings the consolidation of a couple of PCs is. The low clocked TR is extremely viable for most devs and QA folks that actually get work done.
 
I wouldn't be surprised at all. Most of the tech companies hold back and like to keep things up their sleeves.
I've dealt with Fibre 8Gb cards that were actually 16, but disabled on the firmware level.
Few months goes by, then you get the "release" of 16Gb cards... aka same ones w/ different FRU/Part Numbers and new updated firmware.
 
For what purpose is a low clock 32 core consumer cpu? Not trolling just don't see any actual real world situations these are flying off the shelf other than the less than 1% that do have a true need or the other group trying to grow their e-peen... not sure Intel will care since for the majority of real world applications their fewer core cpus will likely "win"...
It would be for applications on the user side that see benefits from heavy multi threading. These applications do exist.
 
I wouldn't be surprised at all. Most of the tech companies hold back and like to keep things up their sleeves.
I've dealt with Fibre 8Gb cards that were actually 16, but disabled on the firmware level.
Few months goes by, then you get the "release" of 16Gb cards... aka same ones w/ different FRU/Part Numbers and new updated firmware.
Omg I know your pain. I hate when that happens.
 
  • Like
Reactions: DSee
like this
What is the significance of this?

It's not like they can be unlocked anyway right? So who cares.

It could be that they are built up as Epyc Chips, and then the ones that don't pass are disabled and sold as Threadripper, but that probably isn't the case considering the different sockets.

They could be building up 4 core chips and then disabling the ones that are damaged in the process to only use two, but I don't think that is the case either, as electrically that would be difficult to connect right.

What is probably just happening is that they have a ton of chips that were binned out for failing. Rather than wasting money on developing or buying some sort of spacer to shove in there, some engineer probably said, "Hey, wait a minute, we have all these waste dies we are just throwing out, why not use them as spacers?"

Either way, this seems like much ado about nothing.

Unless - of course - someone finds a way to unlock them like we could with the disabled dies on the old socket AM2 Phenom II's and Athlons :p
 
Intel currently has 72 core cpus with quad way hyperthreading. I was going to get one but finding the motherboard is near impossible.
 
Well, AMD specifically said that the 1900X was a dual-chip MCM with quad channel memory, but only using 2 CCXs. This means that the memory controllers aren't locked to a CCX, so it's highly possible AMD could make higher core-count CPUs while maintaining the quad-channel memory connection, it also means AMD can release CPUs with asymmetric CCX configurations per die.
 
I can see how if this is true this could save AMD a lot of money in manufacturing costs. I'm not exactly seeing a problem here. Is anyone else seeing a problem here?
 
For what purpose is a low clock 32 core consumer cpu?

Very, very little, especially at the prices the platform would command. Most people, and by that I mean 95%+ if I were to make a wild guess, are still fine with two cores (hopefully with hyperthreading). Hell, I just picked up an ultrabook with exactly that. It's great.

Further, while AMD has released an eight-core SMT CPU for the same price as Intel's current four-core SMT CPUs, the Intel part is still faster in thread-limited applications: and those applications include gaming for the most part. This is why there's still excitement for the 8700k: six cores instead of AMD's eight, but with enough extra clockspeed and instruction per clock throughput to make a difference for gaming and other thread-limited applications, and still be as fast or faster than the eight-core AMD part for most other workloads.

Now, I get the advantage of using many cores (Magny Cours?) on a 'consumer' workstation, and there are plenty of applications that consumers and professionals use that could benefit from the extra resources. They're just not a significant portion of the market.
 
Haven't AMD always just initially designed mostly Server grade stuff and just scaled it down/adjusted slightly for other markets?
AMD was trying to spin TR as custom built high end desktop product... hence the "dummy core" thing clearly these are factory 2nd Epics
 
AMD was trying to spin TR as custom built high end desktop product... hence the "dummy core" thing clearly these are factory 2nd Epics

They could very well be non-salvageable dies - just so the size and physical tolerances all match exactly; prevents non-salvage components from being a waste stream, you can continue to use the Epyc manufacturing die and process, and you don't have to invent a new manufacturing item for "inert die spacer". If they indeed are not electrically connected to the package that would support this theory.

I don't know this first hand, but my impression is the dies are binned off the wafer, and then connected to a substrate inside the package. TR dies do clock faster than Epyc dies in general - which supports the claim that TR is getting top binning. Just because they are reusing the packaging and manufacturing doesn't mean TR is an Epyc reject part or anything of the sort.
 
AMD was trying to spin TR as custom built high end desktop product... hence the "dummy core" thing clearly these are factory 2nd Epics

And in a month or two when the process is mature with far less failed silicon they will be full Epycs.;)
 
Last edited:
Either way I'm not seeing why any of this is a problem. Are you? :)

They could very well be non-salvageable dies - just so the size and physical tolerances all match exactly; prevents non-salvage components from being a waste stream, you can continue to use the Epyc manufacturing die and process, and you don't have to invent a new manufacturing item for "inert die spacer". If they indeed are not electrically connected to the package that would support this theory.

I don't know this first hand, but my impression is the dies are binned off the wafer, and then connected to a substrate inside the package. TR dies do clock faster than Epyc dies in general - which supports the claim that TR is getting top binning. Just because they are reusing the packaging and manufacturing doesn't mean TR is an Epyc reject part or anything of the sort.
 
Back
Top