AMD Ryzen Threadripper Delidded - It's EPYC

[Raevenlord]

Overclocker extraordinaire der8auer, who was one of the most vocal enthusiasts calling out for better VRM designs on Intel's X299 platform (and who worked with ASUS on redesigning the VRM cooling in its motherboards) has gone and done it: he delidded a Ryzen Threadripper CPU. And this delidding went on to deliver the goods: Ryzen Threadripper delidded is EPYC (pun intended.)

Instead of the expected MCM composed of two dies (with two CCXs of four cores per die, delivering the 16 cores we were expecting), Ryzen Threadripper is actually a much more interesting chip: it seems to be a full fledged EPYC chip, with four dies of eight cores. According to der8auer, when questioned, AMD confirmed that 16-core Threadripper 1950X CPUs are configured with two working eight core dies (four CCXs of four cores each), while the other eight-core dies are disabled by AMD.







A very, very interesting question here is whether these are actually defective or just disabled by AMD, which is something the company is naturally mum about. With AMD's history of core unlocking, and the fact that Zen-based CPUs have high yields, we can certainly hope (maybe even expect) that not all disabled cores are non-functional, and that AMD had to disable some of them so as to achieve their SKU core-count. Whether or not those will be unlockable, though, is anybody's guess. Der8auer also found that AMD Threadripper dies are gold plated on the inside, so as to improve conditions and adherence of the indium solder AMD used (some users might say that while Intel uses cheap TIM on their HEDT, X299 CPUs, AMD even sells these with gold inside.) This fact naturally also opens the question of future platform scalability - AMD can certainly decide to just up available maximum core-count on their Threadripper line of CPUs (we still have at least 4 model numbers above the 1950X, an potentially more.)



Check Der8auer's delidding video on YouTube below.

View at TechPowerUp Main Site

 

[Melvis]

Who didnt see this coming? like really? I pretty much said this in an older thread.

Maybe one day we can have a 32core CPU for our threadripper machines?

 

[jigar2speed]

Oh give me a bios that can enable the disabled Cores and i will buy this in a heartbeat.

 

[bogami]

Yes, it would be nice to still open the rest of the cores, but lacking the necessary support on the processor side. It would be best if the four useful dai would work due to thermal distribution and thus higher operating frequencies, but I think the lottery will be here . And the intel new model has twenty cores under the cap, but it's max 18 unlocked, and twenty will sell us 20 next year. EPYC has a much larger reserve .

 

[cadaveca]

The problem with solder is slowly the indium will get sucked onto the large patch of gold (used because of how indium bonds to it), and this will create holes in the solder that then lead to hot spots. This is why I like Intel using paste-based TIM... if it dries out, you can de-lid and replace, but when that solder wicks out, and the CPU runs hot only, you're forced to buy new gear.

This behavior of indium-based solders is properly documented, too, so while some see it as a boon, I see it a weakness in AMD's products, because, you know, I like science, not hype.

 

[buggalugs]

Asus didnt redesign VRM heatsinks for all their boards, just the few ROG boards that havent been released yet.

Most of Asus X299 boards on the market have the original heatsink, same with the other brands.

 

[Ubersonic]

So they are actually 32c CPUs with half of them disabled? Wow lol.

 

[cadaveca]

Most of Asus X299 boards on the market have the original heatsink, same with the other brands.

And if you believe this overheating hype, those ASUS boards you should stay away from, because, as has been said, they have inferior cooling.

 

[Totally]

The problem with solder is slowly the indium will get sucked onto the large patch of gold (used because of how indium bonds to it), and this will create holes in the solder that then lead to hot spots. This is why I like Intel using paste-based TIM... if it dries out, you can de-lid and replace, but when that solder wicks out, and the CPU runs hot only, you're forced to buy new gear.

This behavior of indium-based solders is properly documented, too, so while some see it as a boon, I see it a weakness in AMD's products, because, you know, I like science, not hype.

I never realized that, if that is a problem with Au-In bonding why don't they simply use smaller Gold patches or limit the size of the gold patch to exactly that of the die seems feasible enough to not incur any additional cost.

 

[john_]

I was about to say that maybe those rumors about excellent yields on Ryzen chips where false. But if Threadripper is in fact something that AMD expects to sell in relatively limited quantities, maybe not creating one more product line makes financial sense, even if they in fact throw away fully functional dies under that heatspreader.

But this is really fun. Intel engineers will have a hard time to find a way to create an 18 core model for a socket that was designed for 10 core models, without coming out with a laughable base frequency. Only to see AMD coming out latter with 24 or 32 cores models, probably at more reasonable base frequencies.

In the end it seems that financial problems give AMD an advantage here. With no server market, so nothing to lose, and huge financial problems, that restrict them from creating multiple product lines, they are able to give a few but really solid products to their customers.

 

[cadaveca]

I never realized that, if that is a problem with Au-In bonding why don't they simply use smaller Gold patches or limit the size of the gold patch to exactly that of the die seems feasible enough to not incur any additional cost.

I do not think that solving this issue is really that simple; they are hoping that users that would buy such a product will simply upgrade or replace things when problems arise. There is also the bit about assembly costs that need to be considered as well. The length of time it takes for these sorts of issue to occur can vary, but we've all heard of soldered chips that end up running hot after some time... and now you know the real reason why that occurs.

I was about to say that maybe those rumors about excellent yields on Ryzen chips where false. But if Threadripper is in fact something that AMD expects to sell in relatively limited quantities, maybe not creating one more product line makes financial sense, even if they in fact throw away fully functional dies under that heatspreader.

But this is really fun. Intel engineers will have a hard time to find a way to create an 18 core model for a socket that was designed for 10 core models, without coming out with a laughable base frequency. Only to see AMD coming out latter with 24 or 32 cores models, probably at more reasonable base frequencies.

In the end it seems that financial problems give AMD an advantage here. With no server market, so nothing to lose, and huge financial problems, that restrict them from creating multiple product lines, they are able to give a few but really solid products to their customers.

I don't think the reality of how much one of these chips costs to make vs. it's retail pricing is widely known; I wouldn't call them using dies on a substrate, but having them disabled, really has any detrimental costs associated with doing it.

 

[Woomack]

And if you believe this overheating hype, those ASUS boards you should stay away from, because, as has been said, they have inferior cooling.

On my ASUS X299 TUF Mark 2 ( the cheapest one ) VRM temp under full load with 7900X@4.7GHz is in high 70'. At least on air/water cooling CPU will overheat much faster than the VRM.
People who don't even own Skylake-X make the most noise about how bad it is.

So der8auer who is working with ASUS convinced said ASUS to change heatsink on APEX mobo which is not even on the market and barely anyone will buy it except those who wish to bench on LN2 and then this heatsink won't matter much anyway.

 

[qubit]

I can just see some cheaper motherboards dying from current overload if these get unlocked lol.

I'd bet money that AMD allows these to be unlocked. Very clever AMD, very clever.

 

[<-_->]

AMD ENG SAMPLE
http://imgur.com/a/gsUrv

 

[Boosnie]

The problem with solder is slowly the indium will get sucked onto the large patch of gold (used because of how indium bonds to it), and this will create holes in the solder that then lead to hot spots. This is why I like Intel using paste-based TIM... if it dries out, you can de-lid and replace, but when that solder wicks out, and the CPU runs hot only, you're forced to buy new gear.

This behavior of indium-based solders is properly documented, too, so while some see it as a boon, I see it a weakness in AMD's products, because, you know, I like science, not hype.

I wonder how many cpu in the world perform a complete thermal cycle in theyr lives.

 

[Boosnie]

I never realized that, if that is a problem with Au-In bonding why don't they simply use smaller Gold patches or limit the size of the gold patch to exactly that of the die seems feasible enough to not incur any additional cost.

because it is no problem at all for everybody cpu lifecycle.
It becomes a problem with extreme temperature shifts, like liquid nitrogen T to ambient T over a set amount of cooling-warming cycles.
Indium is a costly material, if it wasn't performing for the Au-In-Si interface at normal T it will not be used at all.
That photo of the threadripper dies exposed shows something around 8-12 dollars of indium.

 

[Totally]

I wasn't concerned with the Indium itself at all, he mentioned that the gold with enough thermal cycling eventually wicks out enough Indium between the heatspreader and the cpu die to cause problems late in the cpu's life. I was just why wouldn't they seemingly obvious fix of 'shortening the wick,' especially since the gold pad as seen don't need to be that big. So, to me this seems to be a case of planned obsolescence the more I think about it.

 

[ShurikN]

The problem with solder is slowly the indium will get sucked onto the large patch of gold (used because of how indium bonds to it), and this will create holes in the solder that then lead to hot spots. This is why I like Intel using paste-based TIM... if it dries out, you can de-lid and replace, but when that solder wicks out, and the CPU runs hot only, you're forced to buy new gear.

This behavior of indium-based solders is properly documented, too, so while some see it as a boon, I see it a weakness in AMD's products, because, you know, I like science, not hype.

And how long does it take for indium solder to "get sucked onto the large patch of gold" and create those holes in solder? What is the last Intel CPU to have solder instead of paste.

 

[Boosnie]

I wasn't concerned with the Indium itself at all, he mentioned that the gold with enough thermal cycling eventually wicks out enough Indium between the heatspreader and the cpu die to cause problems late in the cpu's life. I was just why wouldn't they seemingly obvious fix of 'shortening the wick,' especially since the gold pad as seen don't need to be that big. So, to me this seems to be a case of planned obsolescence the more I think about it.

Thermal cycling in this context (testing Si-In-Au interfaces) is defined -55C° -> 125C°.

 

[R0H1T]

I wasn't concerned with the Indium itself at all, he mentioned that the gold with enough thermal cycling eventually wicks out enough Indium between the heatspreader and the cpu die to cause problems late in the cpu's life. I was just why wouldn't they seemingly obvious fix of 'shortening the wick,' especially since the gold pad as seen don't need to be that big. So, to me this seems to be a case of planned obsolescence the more I think about it.

It always is, that's why we have cheapish TIM on Intel HEDT & a handful of PCIe lanes on RyZen, as compared to TR & EPYC.
It makes business sense, though as I consumer I'll always chose what's best for me & consider the product that provides better VFM, as compared to the other competitor.

And how long does it take for indium solder to "get sucked onto the large patch of gold" and create those holes in solder? What is the last Intel CPU to have solder instead of paste.

The 6950x or any other BDWE btw was that a rhetorical question?

 

[ShurikN]

It always is, that's why we have cheapish TIM on Intel HEDT & a handful of PCIe lanes on RyZen, as compared to TR & EPYC.
It makes business sense, though as I consumer I'll always chose what's best for me & consider the product that provides better VFM, as compared to the other competitor.
The 6950x or any other BDWE btw was that a rhetorical question?

Ok to rephrase that, how long back does the indium solder go on Intel cpus, for example i7 9xx. Are those soldered?

 

[bug]

I was about to say that maybe those rumors about excellent yields on Ryzen chips where false. But if Threadripper is in fact something that AMD expects to sell in relatively limited quantities, maybe not creating one more product line makes financial sense, even if they in fact throw away fully functional dies under that heatspreader.

This is also an easy upgrade path in the future, just unlock more cores and bam! new SKU ready to ship. Even if currently the disabled cores are faulty, in time manufacturing gets better naturally.

 

[iO]

Ok to rephrase that, how long back does the indium solder go on Intel cpus, for example i7 9xx. Are those soldered?

Every IHS since Pentium 4 was soldered.

 

[R0H1T]

Ok to rephrase that, how long back does the indium solder go on Intel cpus, for example i7 9xx. Are those soldered?

Yes, here's a good list of CPU's to know if they were soldered or glued together with the IHS ~ https://www.overclock.net/t/305443/ihs-removals-how-to-do-it-should-i-do-it-and-the-facts

Ivy was the first mainstream CPU, in a long time, that wasn't soldered, SKLX being the first HEDT which went the Ivy way. Basically there's nothing out there that suggests that solder goes kaput first, as compared to TIM, having said that we have TR & SKLX which will prove this theory one way or another.

 

[chaosmassive]

this is the reason why Intel so butthurt and go all out insulting Epyc CPU in their slide decks

think about it, one slice silicon to rule them all market segment from server, HEDT, performance, mainstream, all the way down to potato-grade CPU.
meanwhile Intel need to spend R&D and new design of silicons for each segmentation.