AMD EPYC Server Processor Thread - EPYC 7000 series specs and performance leaked

[ajc9988]

Yes interesting. When Ryzen launched there was much talk of how poor its memory contoller was. Here we see that Epyc's controller runs faster than any Xeon on the market. What could be wrong with the Xeon memory controller? Why cant it run as fast as the Epyc? Is there a bug in Xeon chips or something?

No no no. The new Intel's will support 2666MHz. The bandwidth will not be there on 2P servers though, although rumor has it they will have hexa-channel memory this round, which will increase bandwidth.

2666MHz memory too... is that the fastest Epyc supports?

Yes, but you say that like it is a bad thing. Servers use registered, error correcting memory. This is different from the ram in desktops and HEDT workstations, which can utilize faster. So, using the word "fastest" is a misnomer. The fact it is speced for an extra 266 is wonderful, over the v4 Xeons at 2400.

 

[Markfw]

Based on the way Ryzen works with memory, I would be willing to bet that EPYC will "run" with no problems, memory up to at least 3200 speed, if not 3600.

And I think it can run ECC or ECC registered.

 

[ajc9988]

Based on the way Ryzen works with memory, I would be willing to bet that EPYC will "run" with no problems, memory up to at least 3200 speed, if not 3600.

And I think it can run ECC or ECC registered.

Epyc, yes, as Supermicro sells them with registered ECC ram. The question is if Threadripper can use ECC or ECC registered. That has been a separation between the server and HEDT platforms for awhile, just like support for certain raid setups. So, still more to find out, but if AMD put ECC support on TR, I know a couple people that will be very excited...

 

[formulav8]

ECC registered.

That is a requirement for any server platform. So it definitely should.

 

[Markfw]

That is a requirement for any server platform. So it definitely should.

My point being, I think it will run ECC only as well. And TR I think supports ECC, since Ryzen is rumored to.

 

[formulav8]

I think it will run ECC only as well.

My bad

 

[formulav8]

duplicate somehow

 

[Topweasel]

In the interview between AMD and Dell. Threadripper supports ECC. Registered ECC only for EPYC.

 

[imported_ats]

Later but does not take into account clock rates. Since in the past when dealing with offerings, like The Athlon X2, Athlon, or P4, P4D, C2D, C2Q, anything that made cache or cores unusable made the Die unusable but anything that didn't run as fast could be clocked down. Wasn't till the Phenom and Nehalem that you really start to take CPU's with non-functioning cores and still sell them. So Yields always refers to fully functional dies.

No, yield has always referred to functional dies post repair/recovery.

 

[Topweasel]

No, yield has always referred to functional dies post repair/recovery.

Their are yield statements but with qualifiers. When measure "yields" without the qualifiers it would be fully functional dies.

98% where only 40% are fully functional 8 core dies and 98% dies where 80% are fully functional are completely different. But this is a statistics issue and not really a technical one.

What we do know is that 80% of the dies are fully functional and that by having a configuration that allows for as little as 2 cores per die and half the cache, there almost no dies they can't use.

 

[Veradun]

2666MHz memory too... is that the fastest Epyc supports?

It's the fastest JEDEC DDR4 standard. There is no XMP in server space. 2666 is the max possible. Period :>

 

[imported_ats]

Their are yield statements but with qualifiers. When measure "yields" without the qualifiers it would be fully functional dies.

No, yield without qualifiers is functional yield, period.

What we do know is that 80% of the dies are fully functional and that by having a configuration that allows for as little as 2 cores per die and half the cache, there almost no dies they can't use.

No we don't know that. We have a rumor that perhaps 80% of dies might be fully functional. A Twitter rumor at that. We have jack. From an account that has been wrong on multiple occasions and just had to step back its latest rumor.

 

[itsmydamnation]

who even cares? at 190mm sq the difference between 60% and 80% yield is like $9.*

Assuming $7000 wafer cost.

 

[khon]

who even cares? at 190mm sq the difference between 60% and 80% yield is like $9.*

Assuming $7000 wafer cost.

How did you come up with the idea of a 9$ difference between 60% and 80% yield ?

 

[itsmydamnation]

How did you come up with the idea of a 9$ difference between 60% and 80% yield ?

300mm wafer
is aprox 300 190mm sq dies
$7000 for a wafer on 14nm ( its probably lower these days)
300*0.6 = 180
7000/180 = $38.9

300*0.8 = 240
7000/240 = $29.2

 

[coercitiv]

In other words, the absolute yield improvement becomes less and less important as yield increases. Gloating about using up to 98% of dies is fanboy bait.

 

[lobz]

Intel is smoking crack too if they think they can keep charging that much for a CPU in a competitive environment. The answer is to wait for Intel to respond and then see where the prices settle. But I am going to go out on a limb and say that $4000 will be $2000 in the real world when the dust settles.

You have absolutely no idea about the server market, yet you so arrogantly assume everyone smokes crack but you

(Spam removed)



Insults are not allowed in the tech forums.


esquared
Anandtech Forum Director

 

[Despoiler]

My point being, I think it will run ECC only as well. And TR I think supports ECC, since Ryzen is rumored to.

All Zen variants support ECC. The only thing preventing its use with Ryzen chips are motherboard manufacturers not supporting it in their BIOS. There are a few that support it.

 

[guachi]

In other words, the absolute yield improvement becomes less and less important as yield increases. Gloating about using up to 98% of dies is fanboy bait.

Not exactly.

If AMD (or any company) can make $250 for each die and it costs $7000 for a wafer that can yield up to 300 dies then you need 28 dies to break even for a yield of 9.3%.

A yield of 40% would mean $30,000 per die minus the $7,000 cost for $23,000 gross profit. A yield of 60% would mean $45,000 per die minus the $7,000 cost for $38,000 gross profit. A yield of 80% means $60,000 per die minus the $7,000 cost for $53,000 gross profit.

In other words, a 20% increase in yield increases profits $15,000 no matter if it's a yield increase from 20% to 40% or 80% to 100%. It's the exact same amount of money. Yes, your profits have changed less in a relative sense but it's the exact same in an absolute sense.

 

[nwr]

This entire yield discussion is above my head
Even if we assume the numbers given by BnC were correct (I have a slight suspicion these numbers might have been made up, but even then they could still be correct, if only by accident) - don't we miss a lot of further information to make anything meaningful with these numbers? I for one have no clue what the contract between AMD and the fab partner looks like, if AMD is billed per wafer or chip, validated or not, what amount they are charged etc. nor do I know jack about yields of a comparable design on the same node or anything about how far the Ryzen die design itself contributes to this stellar yield and what part comes from the modular aspect of the chip. If have no idea whatsoever what to make with these 99.8%. Ok, more is better, I get that, so overall it must be really good, but what "it" specifically means... no clue.

 

[coercitiv]

Yes, your profits have changed less in a relative sense but it's the exact same in an absolute sense.

So you start your post by not exactly agreeing with me, then end it by reaching the same conclusion.

10% yields means breaking even in your example.
A jump of 10% improved yields from 20% to 30% doubles your "profit".
A jump of 10% from 90% to 100% means less than 10% increased "profit".

 

[Veradun]

This entire yield discussion is above my head
Even if we assume the numbers given by BnC were correct (I have a slight suspicion these numbers might have been made up, but even then they could still be correct, if only by accident) - don't we miss a lot of further information to make anything meaningful with these numbers? I for one have no clue what the contract between AMD and the fab partner looks like, if AMD is billed per wafer or chip, validated or not, what amount they are charged etc. nor do I know jack about yields of a comparable design on the same node or anything about how far the Ryzen die design itself contributes to this stellar yield and what part comes from the modular aspect of the chip. If have no idea whatsoever what to make with these 99.8%. Ok, more is better, I get that, so overall it must be really good, but what "it" specifically means... no clue.

Actually... capacity isn't unlimited, so yelds would help even under a per-functional-die contract.

 

[lobz]

So, yields are good enough that only 0.1% of all dies have defects in uncore that is around 100mm^2 in size and constitutes about 40% of wafer? Fott delivers premium information as usual.

Thanks, btw.

You don't have to have a fully functioning uncore to find an embedded SKU you can use the die

Sent from my VTR-L09 using Tapatalk

 

[itsmydamnation]

its not 100% accurate but over a year ago Samsung said Defect Density for 14nm LPP was at 0.2 per cm sq which results in an 70% yield. So a year on hitting somewhere around 80% fully functional chips sounds completely plausible, thats a defect density of 0.1 per cm sq. You then have to consider how many chips are defect free but get binned for power/clock reasons.

 

[itsmydamnation]

Did anyone see this ?
https://m.heise.de/newsticker/meldung/AMD-Epyc-legt-los-3748615.html?artikelseite=2
http://www.amd.com/system/files/2017-06/AMD-EPYC-SoC-Sets-4-World-Records.pdf

So the gap between GCC and "tuned" compiler for 7601 vs 2699v4 stays about the same. im guessing ICC is a "better" at spec then open64.