Page 1 of 2

The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:27 am
by chuckula
So delidder extraordinaire Der8auer sacrificed the Threadripper to the Youtube gods as detailed here: https://videocardz.com/71252/amd-ryzen- ... ryzen-dies

The most interesting takeaway -- ESPECIALLY FOR YOU WACO -- is that the Threadripper does indeed have four RyZen dies under its IHS.
The trick is that two of those dies aren't actually turned on, so they are more there for decoration and to balance the package around the center of gravity for maximum stability* but... THERE ARE FOUR DIES! #PicardFTW


* Yeah I know. I just had to throw that in though.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:35 am
by Glorious
Am I missing some context here?

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:37 am
by chuckula
Glorious wrote:
Am I missing some context here?


Waco has been hoping against all hope that the Threadripper is 4 RyZen dies where half of each die is turned off instead of two fully (or partially) functional RyZen dies.
It's because he really wants to put a 32 core Epyc processor in a the Threadripper motherboard and have all the cores work.
I'd just buy an Epyc server myself, but there you go.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:45 am
by Vhalidictes
Thinking about it, this kind of makes sense. It's probably cheaper overall to not have a custom design for Threadripper.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:50 am
by just brew it!
I assume they use dies that failed testing for the 2 "dummy" ones. I guess you'd want *something* in the inactive positions, to ensure that the pressure from the HSF is spread evenly across the package.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:51 am
by derFunkenstein
Glorious wrote:
Am I missing some context here?

It's a really famous ST:TNG episode (two-part) where Picard gets captured and tortured because he refuses to accept someone else's distorted, inaccurate truth.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 11:54 am
by Glorious
Oddly, I got that part.

I was baffled by the Waco disappointment angle.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 12:13 pm
by derFunkenstein
Yeah I dunno. I'm a little surprised that they're effectively wasting two dies.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 12:18 pm
by Vhalidictes
derFunkenstein wrote:
Yeah I dunno. I'm a little surprised that they're effectively wasting two dies.


It's possible that these aren't a special product at all - could be Eypc package harvesting.

Then again, even if Threadripper is its own design in some way, it's a pretty low volume part.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 12:20 pm
by derFunkenstein
I started to type a lot about how it's a low-volume part and maybe engineering costs outweighed die costs, but there's just way too much that I don't know. Like, I just don't know. Seems like a lot of wasted dies, but only if they're totally fine otherwise. Dunno. So I didn't.

Like yesterday, when I deleted my joke. I'm trying to be more discerning in what I write, but I'm often catching myself after I've already written it. Would be nice if the filter between my brain and fingers functioned. :lol:

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 12:41 pm
by the

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Thu Jul 27, 2017 12:54 pm
by Chrispy_
derFunkenstein wrote:
I started to type a lot about how it's a low-volume part and maybe engineering costs outweighed die costs, but there's just way too much that I don't know. Like, I just don't know. Seems like a lot of wasted dies, but only if they're totally fine otherwise. Dunno. So I didn't.

Like yesterday, when I deleted my joke. I'm trying to be more discerning in what I write, but I'm often catching myself after I've already written it. Would be nice if the filter between my brain and fingers functioned. :lol:


I think your first instinct was right. The overall cost of re-tooling for a two-die part likely outweighs the additional wasted die costs.

I'd also expect Threadripper to be initially populated solely with Naples rejects. Given that Threadripper has a high TDP per core compared to Naples, they can use parts that turned out defective in some of the dies and they can also use fully functional parts that need too much voltage to run as server parts.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 12:30 pm
by astrotech66
According to a story I saw today on PCWorld, the two inactive dies aren't real cores. A quote from the article:

"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."

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 12:34 pm
by chuckula
astrotech66 wrote:
According to a story I saw today on PCWorld, the two inactive dies aren't real cores. A quote from the article:

"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."


That's very interesting (and smart on AMD's part BTW, no reason to waste expensive silicon in situations where you know it will never be turned on at all).

It actually makes my intentionally joking line near the end of the first post seem logical :-)

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 1:37 pm
by Waco
Glorious wrote:
Oddly, I got that part.

I was baffled by the Waco disappointment angle.

My hope was that with four active dies, the memory controllers would be half-active on each die...and increase the chances of a cheap Threadripper motherboard being suitable for a drop-in Epyc chip (since they're cheaper than Threadrippers of the same core count).

That, and there would be hope of an update later to allow for full 32-core Threadrippers on the same motherboards. Beyond all hope I was crossing my fingers that they might have run 8 channels out to the dimms...

EDIT: I'm glad to hear they aren't just tossing out functional dies, though.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 1:41 pm
by Mr Bill
Maybe 4 dies are required because of the interposer connections and they don't want to redesign the interposer and the connection logic.

Edit: And that just begs the question are the dies designed in such a way that one could etch and enable just the portions needed to support the infinity fabric?

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 2:27 pm
by Waco
Mr Bill wrote:
Maybe 4 dies are required because of the interposer connections and they don't want to redesign the interposer and the connection logic.

Edit: And that just begs the question are the dies designed in such a way that one could etch and enable just the portions needed to support the infinity fabric?

...or if half of the socket itself is just dead.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 2:35 pm
by chuckula
Waco wrote:
Mr Bill wrote:
Maybe 4 dies are required because of the interposer connections and they don't want to redesign the interposer and the connection logic.

Edit: And that just begs the question are the dies designed in such a way that one could etch and enable just the portions needed to support the infinity fabric?

...or if half of the socket itself is just dead.


Half the socket simply not being hooked up seems like the most logical approach to me... but who knows really. The bigger LGA sockets like both LGA-3647 and the Threadripper/EPYC sockets usually have two land grid modules that define the full socket. I could see one of those modules simply being there for physical mounting purposes while the entire Threadripper chip is electrically connected to the other module.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 3:38 pm
by maxxcool
Image

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 4:51 pm
by Krogoth
The most plausible explanation is that there are early batch of recycled "Epyc" chips.. I wouldn't be too shocked that future "Threadripper" batches will end-up coming with only two dies on the packaging. I doubt you can unlock the "disabled" dies since it looks like AMD killed them after proving them to be "defective" via laser cuts.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 6:33 pm
by just brew it!
So you guys are completely discounting the PC World article astrotech66 mentioned?

Die-shaped spacers to maintain mechanical integrity of the package is the simplest (and most plausible) explanation.

The "partially defective Epyc" explanation is less plausible, because they'd only be able to "harvest" the ones that happen to have the good dies in two specific locations, to avoid having to use different package pinouts for each possible combination of good dies.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 6:40 pm
by Mr Bill
just brew it! wrote:
So you guys are completely discounting the PC World article astrotech66 mentioned?

Die-shaped spacers to maintain mechanical integrity of the package is the simplest (and most plausible) explanation.

The "partially defective Epyc" explanation is less plausible, because they'd only be able to "harvest" the ones that happen to have the good dies in two specific locations, to avoid having to use different package pinouts for each possible combination of good dies.
Missed that, DOH. :P

But wait, magically? Magic might work for Harry Potter, but a mere fantasy of magic, I guess not.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 6:45 pm
by Mr Bill
I suppose one other thing we can conclude from the distance between the packages is that AMD is not that worried about the longer traces affecting the efficacy of the infinity fabric. Also, it may be easier to effectively cool the epYc package with the heat sources "spread" out from eachother. On the other hand, Intel is seemingly not too worried about the smaller form factor of the I9 chip footprint for heat and may gain a few nanoseconds for closeness.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 7:49 pm
by deruberhanyok
Do you know, the worst part of it all, at the end there - I was going to say there were five dies. But more than that, I actually believed it.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 8:08 pm
by just brew it!
Mr Bill wrote:
I suppose one other thing we can conclude from the distance between the packages is that AMD is not that worried about the longer traces affecting the efficacy of the infinity fabric. Also, it may be easier to effectively cool the epYc package with the heat sources "spread" out from eachother. On the other hand, Intel is seemingly not too worried about the smaller form factor of the I9 chip footprint for heat and may gain a few nanoseconds for closeness.

It could also be an attempt to reduce the trace lengths from the memory controllers to the corresponding contacts on the underside of the package, and or to bring the memory controller signals closer to the edge of the package, to minimize the trace length from the CPU socket to the DIMM slots.

Edit: Some food for thought -- at DDR4 speeds, the physical length of the electrical impulses representing individual bits on the memory bus is comparable to the length of the traces between the memory controller and the DIMMs. IOW, by the time the receiving end has detected an incoming bit, the next bit is already being stuffed into the other end of the wire. Electricity does weird things at GHz frequencies... :wink:

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 8:30 pm
by chuckula
deruberhanyok wrote:
Do you know, the worst part of it all, at the end there - I was going to say there were five dies. But more than that, I actually believed it.


Actually, given the rumors about 80% yields on RyZen dies, statistically there are five dies. It's just one of them didn't make the cut :-P

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 10:46 pm
by cegras
Defective dies are better used in Ryzen 3 or 5 parts .. not as turned off spacers in Threadripper.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Fri Jul 28, 2017 11:51 pm
by Redocbew
If the PC World story is accurate, then the other two dies are not "defective", because they were never going to be part of a working CPU anyway. They're just spacers.

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Sat Jul 29, 2017 12:35 am
by Waco
maxxcool wrote:
Image


<3

Re: The Threadripper: THERE. ARE. FOUR. DIES!

Posted: Sat Jul 29, 2017 2:53 am
by Ifalna
Look again, there are FIVE lights!

*uses torture device*