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

[ultimatebob]

Another article:

http://www.zdnet.com/article/amds-e...on-2-socket-performance-at-every-price-point/

I wonder when AWS and Google Cloud will start offering these servers? I'm hoping for a price cut on my hosting charges!

 

[csbin]

Memory Bandwidth：253.34GB/s

http://ranker.sisoftware.net/show_r...e4dce4d1e1d2e2c4b68bbb9df89da090b6c5f8c8&l=en

 

[estarkey7]

Memory Bandwidth：253.34GB/s

http://ranker.sisoftware.net/show_r...e4dce4d1e1d2e2c4b68bbb9df89da090b6c5f8c8&l=en

Is that a good score compared to Intel offerings?

Sent from my SM-G935T using Tapatalk

 

[scannall]

Regardless of the results, I think by now we all know "the race is on" and competition is back in the CPU world, personal all the way up to server. (Ryzen to TR to EPYC)

And the only winners are us, the customers, including businesses.

Is there a Threadripper in your future Mark? If so, a Threadripper build thread might be cool.

 

[teejee]

Is that a good score compared to Intel offerings?

Sent from my SM-G935T using Tapatalk

yes, it completely destroys Intel. All Xeon E7 v4 has 85 GB/s according to spec...


Sent from my LG-D855 using Tapatalk

 

[iBoMbY]

Is that a good score compared to Intel offerings?

It says: "1st Place: Gold Cup" ...

 

[T1beriu]

Bits And Chips realized how ridiculous their claim was that "AMD is using 99,9% of Ryzen Die" and now are readjusting their "claim" to 98% with another ridiculous explanation.

Pure BS.

 

[Veradun]

Bits And Chips realized how ridiculous their claim was that "AMD is using 99,9% of Ryzen Die" and now are readjusting their "claim" to 98% with another ridiculous explanation.

Pure BS.

Side note: I've always used "yelds" as an aggregate for "every fraking die we'll use in a SKUs", not for "every top bin fully functional die". Is the trend the former or the latter now (at least on these fora)?

 

[nwr]

It says: "1st Place: Gold Cup" ...

Some would say it's mediocre at best... ;-)
No if real, I mean holy sheet :O

 

[Udgnim]

Side note: I've always used "yelds" as an aggregate for "every fraking die we'll use in a SKUs", not for "every top bin fully functional die". Is the trend the former or the latter now (at least on these fora)?

my interpretation is 80+% for 4 core dies

99.9% for 1 to 3 core dies

real 98% because the 1 core dies are only going to be used in 8 core EPYC

unless AMD decides to do something like release 1 core APUs

their 98% doesn't make sense if yield is considered an aggregate for every die that will be used in a SKU

 

[Topweasel]

Side note: I've always used "yelds" as an aggregate for "every fraking die we'll use in a SKUs", not for "every top bin fully functional die". Is the trend the former or the latter now (at least on these fora)?

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.

 

[thecoolnessrune]

Side note: I've always used "yelds" as an aggregate for "every fraking die we'll use in a SKUs", not for "every top bin fully functional die". Is the trend the former or the latter now (at least on these fora)?

Your first statement is how I've always understood yields. Yield of a die as a percentage = Percentage you can sell. My professors all spoke of it as the percentage of dies that leave the Fab for products vs. the number of dies that could have been made from the sum of all wafers entering the fab.

While I've heard wiggle room in the statement, I've never heard it as loosely defined as "yield of only 1 subset of SKU's" without that measurement being explicitly defined prior to the number being provided.

 

[Topweasel]

Your first statement is how I've always understood yields. Yield of a die as a percentage = Percentage you can sell. My professors all spoke of it as the percentage of dies that leave the Fab for products vs. the number of dies that could have been made from the sum of all wafers entering the fab.

While I've heard wiggle room in the statement, I've never heard it as loosely defined as "yield of only 1 subset of SKU's" without that measurement being explicitly defined prior to the number being provided.

Again that confusion started when the dies had to be fully functional to sell. If you are making a 8 core die. Then you want 8 core dies. Yields are how many 8 core dies you get. Harvesting that can increase % of used dies. But yields are building something and the percentage of time you get what you are building.

 

[Veradun]

Again that confusion started when the dies had to be fully functional to sell. If you are making a 8 core die. Then you want 8 core dies. Yields are how many 8 core dies you get. Harvesting that can increase % of used dies. But yields are building something and the percentage of time you get what you are building.

Uhm, that makes sense as long as we are all speaking the same language (and B&C clearly meant that tweet the other way.

Well, move along, nothing to see here :>

 

[SlowSpyder]

Memory Bandwidth：253.34GB/s

http://ranker.sisoftware.net/show_r...e4dce4d1e1d2e2c4b68bbb9df89da090b6c5f8c8&l=en

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

 

[moinmoin]

I'd say with regard to foundries yield refers to faultless dies, their goal to drive up yield of a given process node remains the same as always. What AMD is doing there is consciously designing the dies (and the product portfolio using them) with a lot of built in fault tolerance allowing them to salvage and use apparently most of the not-faultless dies.

As the (foundry) yield for 14LPP is already plenty decent it doesn't really matter all that much. But for the upcoming 7LP this setup will be nice to rely on for everybody involved.

 

[Veradun]

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

Yep it's in the slides

 

[Phynaz]

Side note: I've always used "yelds" as an aggregate for "every fraking die we'll use in a SKUs", not for "every top bin fully functional die". Is the trend the former or the latter now (at least on these fora)?

Yield is measured two ways;

Functional yield - Chips that function correctly.
Parametric yield - Chips that function within the parameters that are set for them, such as power consumption, clock, voltage, process variation, operating tolerances, etc..

Roughly (very) functional yield is constrained by manufacturing and parametric yield is constrained by design, although the two are quite interrelated.

Here's a great paper on semiconductor manufacturing yield
https://pdfs.semanticscholar.org/142a/aacafc6b3f7adc52c158789db41c0b9324fc.pdf

One thing to take away from that paper is yield is dependent upon design. So when you see a press release or any other article that states XX% yield you can bet that an absolutely meaningless number has been given to you.

 

[piesquared]

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

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?

 

[Ajay]

Yield is measured two ways;

Functional yield - Chips that function correctly.
Parametric yield - Chips that function within the parameters that are set for them, such as power consumption, clock, voltage, process variation, operating tolerances, etc..

Roughly (very) functional yield is constrained by manufacturing and parametric yield is constrained by design, although the two are quite interrelated.

Here's a great paper on semiconductor manufacturing yield
https://pdfs.semanticscholar.org/142a/aacafc6b3f7adc52c158789db41c0b9324fc.pdf

One thing to take away from that paper is yield is dependent upon design. So when you see a press release or any other article that states XX% yield you can bet that an absolutely meaningless number has been given to you.

Great post! It like B&C is using a number that represents CPUs that fail after binning. That is, failures that occur in packaging, etc.

 

[Ajay]

I must openly admit, that observing your butthurt is a pretty epic thing./s

Classy. Guess it's your lucky day as didn't get a warning

 

[lefty2]

I think the reason that Epyc is proving so popular is the amount out I/O it provides - 128 PCIe lanes.
4 years ago 40 PCIe lanes were considered adequate for a server, but times have changed. These days servers are being packed with SSDs connected via NVMe - that chews through a lot of PCIe lanes and then you have GPU servers - each GPU uses 8 lanes of PCIe. Customers are using 2 socket servers, not because they need more cores, but to get more PCIe. In 4 years Intel has increased the number of PCIe lanes by only 10% and that has left it hopeless underspec'ed.

 

[Ajay]

I don't think we can say Epyc is popular yet. We have no sales figures, AFAIK.

 

[Kenmitch]

Debating the semantics of yields seems pointless too me. Really all that matters is how many useable dies can be sold in the end. Zen looks to have been a well thought out design. Kudos to AMD!

AMD bring on the pain! Intel really needed a wakeup call.

 

[Markfw]

Is there a Threadripper in your future Mark? If so, a Threadripper build thread might be cool.

Most likely, if the budget allows it (I have a few bills right now)

And if I don't start a TR build thread, or someone beats me to it, there WILL be one.