Ryzen Threadripper Delidding

[Icywolfe]

AMD could make a 24 core with 64 PCIE lanes (Unlike Eypc's 128) with quad channel memory. This will murder workstations based on the intel 22 core xeon chips and murder the 18 core intel HEDT chip. 

 

It could be a [3+3] + [3+3] + [3+3] + [3+3] price this at 1999$.At maybe like a 3.1 ghz.

[raphidy]

AMD just gotta press a button and reck Intel's HEDT. Want to see AMD move when intel drops the 18cores.

[leadeater]

4 hours ago, VegetableStu said:

The EPYC and TR stack have cores in multiples of 4 so far. I'm not sure if AMD would want to disable an uneven number of cores over the four CCXes (as in e.g. 3:3:2:2, not non-multiple-of-2 equally) ._.

 

20-core sounds more possible, maybe?

Won't be 20 as that would be imbalanced CCX's per die and dies need to be balanced too. If only two dies are being used which is the most likely case then 16 is the max, two dies is most likely due to the memory channels and PCIe lanes.

 

There is a reason why EYPC comes in 8, 16, 24 and 32 core variants. 1 core per CCX and 2 cores per die so 8 core package, 2 cores per CCX and 4 cores per die so 16 core package, 3 cores per CCX and 6 cores per die so 24 core package then finally 4 cores per CCX and 8 cores per die so 32 core package.

 

Threadripper possible variants is half of EYPC due to two active dies: 4, 8, 12 and 16. What AMD choose to actually offer is up to them.

[Jito463]

2 hours ago, VegetableStu said:

I dunno, if the methodology with R3 and R5 made sense for them, I don't see why they should bomb two CCXes (even if they leave the glass in there) when they could pick which core to disable.

I tend to agree with you, however I must point out that TR has a much higher profit margin.  Meaning that they can afford to lose two dies and still make a decent profit.  It does strike me as odd that they would leave two inactive dies on TR, though.

[The Benjamins]

1 minute ago, Jito463 said:

I tend to agree with you, however I must point out that TR has a much higher profit margin.  Meaning that they can afford to lose two dies and still make a decent profit.  It does strike me as odd that they would leave two inactive dies on TR, though.

do note that this is a engineering sample and the retail models could be 2 dies.

[MageTank]

Poor Der8auer:

Woke up, only to see AMD request his video get taken down, lol. 

[Jito463]

2 minutes ago, The Benjamins said:

do note that this is a engineering sample and the retail models could be 2 dies.

That is true, and I had considered that.  In retrospect, they could have done this to prevent leaks from getting to Intel about their HEDT platform.  If they ordered it the same as their Epyc chips (which Intel did know about), then it would appear to an outsider that they were experimenting with different variations of Epyc.

[Jito463]

3 minutes ago, MageTank said:

Poor Der8auer:

Woke up, only to see AMD request his video get taken down, lol. 

Is there an embargo still on information about TR?  Did he sign any sort of NDA to get that sample of TR?

[The Benjamins]

1 minute ago, Jito463 said:

That is true, and I had considered that.  In retrospect, they could have done this to prevent leaks from getting to Intel about their HEDT platform.  If they ordered it the same as their Epyc chips (which Intel did know about), then it would appear to an outsider that they were experimenting with different variations of Epyc.

they could also have made it were they repurpose old epyc ES to become TR samples.

[MageTank]

Just now, Jito463 said:

Is there an embargo still on information about TR?  Did he sign any sort of NDA to get that sample of TR?

He said he had written permission from AMD to post it yesterday, but that they called him to have it taken down. Here's the update: 

 

[The Benjamins]

1 minute ago, Jito463 said:

Is there an embargo still on information about TR?  Did he sign any sort of NDA to get that sample of TR?

ya that seems odd, and he works with stuff under NDA before so I wouldn't expect him to screw it up.

[mariushm]

So yeah, there was some article in PC World saying two of the dies are DUMMY dies, and are there just for rigidity, so that the metal lid on top of the other dies won't flex or bend or whatever ... to have equal pressure over the whole package:

 

Link : http://www.pcworld.com/article/3211409/computers/why-ryzen-threadripper-has-two-mysterious-chips.html

 

Quote

 

So did AMD really waste two perfectly good "Zeppelin" dies? Nope.

Those other two "chips" are nothing more than spacers to help maintain the structural integrity of the gigantic heat spreader, PCWorld has learned from a source who declined to be identified.

The source didn't detail much more, but it makes a lot of sense when you consider the size of the CPU and its construction. Rather than a single monolithic die, like Intel uses on its Core i9, AMD builds its Ryzen and Ryzen Threadripper CPUs using multiple chips connected with a high-speed fabric.

If AMD had populated a Threadripper CPU with just two of the dies, the metal heat spreader would likely bend where the empty spots are.

So no, your fantasy of somehow magically turning a 16-core Ryzen Threadripper 1950X into a 32-core won't happen.

 

 

[Jito463]

32 minutes ago, mariushm said:

So yeah, there was some article in PC World saying two of the dies are DUMMY dies, and are there just for rigidity, so that the metal lid on top of the other dies won't flex or bend or whatever ... to have equal pressure over the whole package:

 

Link : http://www.pcworld.com/article/3211409/computers/why-ryzen-threadripper-has-two-mysterious-chips.html

 

 

If that's true, it would make sense.

[XCalinX]

Well the video and its description still showed up on Google, so I downloaded the pics from the link he had in the description, for our enjoyment

 

 

 

[MageTank]

1 hour ago, mariushm said:

So yeah, there was some article in PC World saying two of the dies are DUMMY dies, and are there just for rigidity, so that the metal lid on top of the other dies won't flex or bend or whatever ... to have equal pressure over the whole package:

 

Link : http://www.pcworld.com/article/3211409/computers/why-ryzen-threadripper-has-two-mysterious-chips.html

 

 

That's... odd. Seems weird that they would waste the funds soldering dies that serve zero purpose, when they could use raw glue in that area for a similar effect. Then again, I don't really know anything about the structural integrity of processor substrates, so I certainly won't pretend to be an expert, lol. Either way, if it's true, kudos to AMD for not taking the cheap way out. 

[mariushm]

The indium solder costs probably less than 1$, and "blank" (or faulty/dead) dies are basically free or worst case scenario a few dollars. Not much when we're talking about a 700-800$ or a 1000$ cpu.

 

From what I heard, each Ryzen die costs around 30-50$ to make, test, bin and prepare to be put on the base pcb.. so you're looking at around 100-150$ to make a threadripper, everything else being profit, recuperating r&d

 

Also it makes sense to reuse the same base pcb for EPYC processors and keep pinout more or less the same between EPYC and Threaripper machines, it's just easier to use the same machines to test the processors after they placed on the base pcb and have solder on top and lids glued and everything.

[MageTank]

2 minutes ago, mariushm said:

The indium solder costs probably less than 1$, and "blank" (or faulty/dead) dies are basically free or worst case scenario a few dollars. Not much when we're talking about a 700-800$ or a 1000$ cpu.

 

From what I heard, each Ryzen die costs around 30-50$ to make, test, bin and prepare to be put on the base pcb.. so you're looking at around 100-150$ to make a threadripper, everything else being profit, recuperating r&d

 

Also it makes sense to reuse the same base pcb for EPYC processors and keep pinout more or less the same between EPYC and Threaripper machines, it's just easier to use the same machines to test the processors after they placed on the base pcb and have solder on top and lids glued and everything.

It's not so much the cost of the solder itself, but more so the process involved. Seems like it would take more time (and time is money). Then again, I also do not know how the soldering process is done. If it's entirely automated, it's probably a complete non-issue. If that is the case, then this seems to be win-win if it really does improve rigidity of the substrate. 

[Ground]

Just now, MageTank said:

It's not so much the cost of the solder itself, but more so the process involved. Seems like it would take more time (and time is money). Then again, I also do not know how the soldering process is done. If it's entirely automated, it's probably a complete non-issue. If that is the case, then this seems to be win-win if it really does improve rigidity of the substrate. 

 

Well, taking the news of there being 2 active dies and 2 dummies, I would suspect that they produce that and then sent the chip through the same machine that adds the heatspreader to the EPYC CPUs - saving some cost there. No need for an extra assembly machine there, thus saving a lot of cost. 

[MageTank]

Just now, Ground said:

Well, taking the news of there being 2 active dies and 2 dummies, I would suspect that they produce that and then sent the chip through the same machine that adds the heatspreader to the EPYC CPUs - saving some cost there. No need for an extra assembly machine there, thus saving a lot of cost. 

Would love to see a CPU manufacturing behind-the-scenes video, showing this process. I imagine plenty of other people would love that kind of content as well. I am pretty good at delidding, but have never really seen how exactly they go on in the first place, lol. 

[mariushm]

2 minutes ago, MageTank said:

Would love to see a CPU manufacturing behind-the-scenes video, showing this process. I imagine plenty of other people would love that kind of content as well. I am pretty good at delidding, but have never really seen how exactly they go on in the first place, lol. 

Here you go :

 

 

[MageTank]

3 minutes ago, mariushm said:

Here you go :

 

 

Yeah, I recall watching this before. I should have been more specific. I meant a video seeing the soldering process of how they solder these CPU's to the IHS. I assume they are using some sort of solder paste, that "bakes" into a joint once heated/cooled. I know Der8auer explains the various materials used in this guide: http://overclocking.guide/the-truth-about-cpu-soldering/ but the process itself is still ambiguous to me. Is it done by hand (kinda like Der8auer did his) or is it an automated process? If it's automated, how exactly do they prep it, and in what order is it prepped? This is something I am extremely interested in learning, but can't really find the relative information. 

[Ground]

Well, let's do some simple math. Indium isn't as expensive anymore, sitting at 250$/kg (at least thats what I saw. 1L weighs around 7.3kg. Let's say the solder layer is 1 mm thick, with the Ryzen DIE being 195mm^2. This would mean that we'll need around 1.4g of Indium per DIE. This means - 0.35$ of Indium per soldered DIE, or 0.7$s extra per Threadripper CPU. I guess that difference is negligible compared to the cost of getting a new machine to assemble glue the pieces of Silicon instead... 

[mariushm]

No, they get tiny squares of solder, sheets of indium solder.  They have a sort of pick and place machine which takes two such squares and puts them over each actual ryzen die (so you get 2 squares on Ryzen processors). Another machine pours that black potting compound / adhesive where it should be (near  the edges of the lid and in the center of larger cpus like threadripper/epyc)

 

The bottom of the lid is gold plated because the indium solder binds better to gold compared to aluminum or whatever that lid is made of (i think it's aluminum with some nickel plating or something like that). So they put put the lid over and then slowly heat the lid up to around 180 degrees Celsius which makes that solder bind to both surfaces. The solder itself has a "turn liquid-ish" point at around 144-155 degrees Celsius but the lid acts as heatsink all the time, so going up to around 180 degrees is the safest.

 

The heat also activates that potting compound making it hard.. but that needs significantly more time (hours I think) to reach the final strength and stickyness.

 

[Jito463]

20 minutes ago, Ground said:

Well, let's do some simple math. Indium isn't as expensive anymore, sitting at 250$/kg (at least thats what I saw. 1L weighs around 7.3kg. Let's say the solder layer is 1 mm thick, with the Ryzen DIE being 195mm^2. This would mean that we'll need around 1.4g of Indium per DIE. This means - 0.35$ of Indium per soldered DIE, or 0.7$s extra per Threadripper CPU. I guess that difference is negligible compared to the cost of getting a new machine to assemble glue the pieces of Silicon instead... 

Why bother with all of that?  Going by Intel, they can just use the same glue that they used to stick the four dies on the substrate.  Much cheaper solution. 

 

 

 

 

 

 

 

 

 

 

 

_{*I would hope that's recognized as sarcasm/humor*}

[zMeul]

AMD "asked" him to take the video down

 

 

why? simple: Intel was 100% right when they said Threadripper was a glued together CPU