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

[Phynaz]

Microsoft, Samsung, Baidu ,1&1 and Dropbox have said they are using Epyc and I doubt these companies would be buying a poor product especially with the budgets they have.

Edit to post.

You're first incorrect assumption is all those companies have bought anything. They've been seeded.

 

[USER8000]

LOL,saw this mentioned elsewhere:
https://seekingalpha.com/news/32747...l-amd-epyc-reveal-shares-1_6-percent?uprof=55

Bank of America Merrill Lynch downgrades Intel (NASDAQ:INTC) from Buy to Neutral and lowers the price target from $42 to $38 after yesterday’s Advanced Micro Devices (NASDAQ:AMD) Epyc server launch event.

Analyst Vivek Arya lowers Intel’s data center sales growth for FY18 from 11% to 8% and notes that AMD’s Epyc has signed up Microsoft and Baidu among its big-name data center customers.

Meanwhile, Morgan Stanley rates AMD at Equal Weight with analyst Joseph Moore noting that AMD might have a price impact on Intel next year. Moore says Epyc launch impressive “even as moderate skeptics.”

Source: Bloomberg

Intel shares are down 1.55% premarket. AMD shares are up 5.22%.

Previously: Jefferies reiterates on AMD Buy after Epyc event; shares up over 2% (June 21)

https://seekingalpha.com/news/32747...yc-event-shares-2-percent?uprof=55#email_link

Jefferies reiterates Advanced Micro Devices (NASDAQ:AMD) at a Buy rating with $16 price target following yesterday afternoon’s launch event for new Zen MPU Epyc.

Analyst Mark Lipacis discusses the key reveals during the event including the fact that Baidu, Microsoft Azure, Dropbox, and Bloomberg all plan to put Epyc in their data centers.

Dell EMC, SuperMicro, and HP all announced Epyc-based SKUs.

Lipacis expects Epyc to capture 7% of the total server market by the end of next year.

Source: StreetInsider

Latest AMD analyst standings: 5 Buy, 5 Outperform, 17 Hold, 2 Underperform, and 2 Sell.

Median price target: $12.50.

AMD shares are up 2.06% premarket to $12.90 after closing yesterday up nearly 6%.

 

[raghu78]

You're first incorrect assumption is all those companies have bought anything. They've been seeded.

All the companies mentioned have stated they are deploying EPYC based servers. AMD has been seeding EPYC to server OEMs and cloud service providers for close to a year now. Server solutions have long development cycles and require rigorous validation. These companies needed to have been working with AMD for over a year to deploy EPYC in Q3/Q4 2017.

https://globenewswire.com/news-rele...orms-and-Global-Server-Ecosystem-Support.html

As the world’s largest Chinese language search engine and leading AI-Tech company, Baidu prides itself on simplifying a complex world through technology,” said By Dr. Zhang Ya Qin, president of Baidu. “The AMD EPYC processor powered one-socket server can significantly increase our datacenter computing efficiency, reduce TCO and lower energy consumption. We will start deploying with the launch of AMD EPYC and I look forward to our cooperation leading to scaled EPYC adoption this year, and ongoing innovations.”

“We’ve worked to make Microsoft Azure a powerful enterprise grade cloud platform, that helps guide the success of our customers, no matter their size or geography,” said Girish Bablani, corporate vice president, Azure Compute, Microsoft Corp. “To power Azure, we require the most cutting-edge infrastructure and the latest advances in silicon which is why we intend to be the first global cloud provider to deliver AMD EPYC, and its combination of high performance and value, to customers.

Bloomberg

“At Bloomberg, we handle the flow of information for professionals in the capital markets. We look forward to AMD’s leadership in the open standards communities, such as OCP, NVMe, GenZ, and CCIX, to help accelerate the industry-wide adoption of these innovative data center, storage, and interconnect solutions,” said Justin Erenkrantz, head of compute architecture for Bloomberg, the global finance, media and tech company based in New York City. “With higher density and lower latency NVMe storage a primary feature delivered by the AMD EPYC processor, we fully expect to realize the next level of performance and cost efficiency.”

Dropbox

“Dropbox customers expect fast, reliable access to the content they ask us to manage for them and EPYC delivers on those requirements. We have worked closely with AMD during our evaluation of EPYC in our environment and see significant potential in lowering total-cost-of-ownership while improving performance in single-socket designs," said Akhil Gupta, vice president of infrastructure at Dropbox. "Our evaluation systems take advantage of the industry-leading 128 lanes of PCIe on EPYC for storage performance and capacity. Dropbox is exploring deployment options for EPYC later this year, and I believe the future looks bright for the relationship with AMD and EPYC.”

LexisNexis

At LexisNexis® Risk Solutions, we believe in the power of data and advanced analytics for better risk management. As a trusted data analytics provider for organizations seeking actionable insights to manage risks and improve results while upholding the highest standards for security and privacy,” said Flavio Villanustre, vice president, Technology, LexisNexis Risk Solutions, RELX Group. “LexisNexis processes more than 90 million transactions per hour with HPCC Systems®, a proven, open source solution for Big Data. LexisNexis Risk Solutions and AMD have teamed up to optimize the HPCC Systems platform to take advantage of the benefits of thread density, core performance, memory bandwidth and the industry leading 128 lanes of PCIe per socket of the AMD EPYC processor.”

 

[Ancalagon44]

Congratulations to AMD on yet another good (paper) launch!

If you look at how far back they were, I mean they were nowhere in the server market. No market share at all. Now, they have a product that actually competes! Sure, Intel is faster in IPC, but AMD might actually win some benchmarks with this. When is the last time you saw an AMD processor win a benchmark against an equivalently priced Intel CPU? *

* Yes I know that in certain cases 8 core Bulldozer CPUs could beat 4 core Intel CPUs.

 

[Phynaz]

Snip...

None of what you posted contradicts my post.

 

[Topweasel]

I was wondering what half the replies in this thread were aimed it. So temporarily turned off my "ignore"... and found a rake of "lolfail" posts.

With that username, s/he was already hovering near my ignore list before reading any of their posts. A few weeks later, a solid addition to the ignore list was made.

Do it. It'll save you hassle!

I ran into the same thing. Responses with no quotes. Weird stuff that didn't need to be defended. Then I figured out Fail was on some crusade again.

 

[Topweasel]

Also good news with the advent of the 8c EPYC. AMD can fit 99.9% of their Zeppelin dies into SKU's.

https://twitter.com/BitsAndChipsEng/status/877495715316060164

 

[raghu78]

None of what you posted contradicts my post.

The mentioned companies are buying EPYC unlike what you seem to indicate. A server does not land on your lap the minute you wish it to. For AMD's customers to deploy EPYC servers in Q3/Q4 2017 they must be getting production shipments of EPYC now.

 

[Phynaz]

The mentioned companies are buying EPYC unlike what you seem to indicate. A server does not land on your lap the minute you wish it to. For AMD's customers to deploy EPYC servers in Q3/Q4 2017 they must be getting production shipments of EPYC now.

Quote where they said they are buying them please.

You still don't understand marketing speak.

Anyway, of course they may buy some Epyc Servers. The question is how many.

Do you think the major cloud providers will making a move away from Intel? In what timeframe?

 

[lolfail9001]

Also good news with the advent of the 8c EPYC. AMD can fit 99.9% of their Zeppelin dies into SKU's.

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.

 

[Markfw]

Quote where they said they are buying them please.

You still don't understand marketing speak.

Anyway, of course they may buy some Epyc Servers. The question is how many.

Do you think the major cloud providers will making a move away from Intel? In what timeframe?

What speaks volumes to me and obviously to those who know and INVEST, is (quoted from link above)

"Bank of America Merrill Lynch downgrades Intel (NASDAQ:INTC) from Buy to Neutral and lowers the price target from $42 to $38 after yesterday’s Advanced Micro Devices (NASDAQ:AMD) Epyc server launch event.

Analyst Vivek Arya lowers Intel’s data center sales growth for FY18 from 11% to 8% and notes that AMD’s Epyc has signed up Microsoft and Baidu among its big-name data center customers.

Meanwhile, Morgan Stanley rates AMD at Equal Weight with analyst Joseph Moore noting that AMD might have a price impact on Intel next year. Moore says Epyc launch impressive “even as moderate skeptics.”

Source: Bloomberg

Intel shares are down 1.55% premarket. AMD shares are up 5.22%.

Previously: Jefferies reiterates on AMD Buy after Epyc event; shares up over 2% (June 21)"

AND

"Jefferies reiterates Advanced Micro Devices (NASDAQ:AMD) at a Buy rating with $16 price target following yesterday afternoon’s launch event for new Zen MPU Epyc.

Analyst Mark Lipacis discusses the key reveals during the event including the fact that Baidu, Microsoft Azure, Dropbox, and Bloomberg all plan to put Epyc in their data centers.

Dell EMC, SuperMicro, and HP all announced Epyc-based SKUs.

Lipacis expects Epyc to capture 7% of the total server market by the end of next year.

Source: StreetInsider

Latest AMD analyst standings: 5 Buy, 5 Outperform, 17 Hold, 2 Underperform, and 2 Sell.

Median price target: $12.50.

AMD shares are up 2.06% premarket to $12.90 after closing yesterday up nearly 6%."

 

[TheGiant]

That is nice.

I hope Intel will wake up and bring us a new C2D with big style !

It looks like I might buy an AMD CPU as an upgrade

 

[raghu78]

Quote where they said they are buying them please.

You still don't understand marketing speak.

Anyway, of course they may buy some Epyc Servers. The question is how many.

Do you think the major cloud providers will making a move away from Intel? In what timeframe?

When companies like baidu and Microsoft say they are deploying EPYC its clear that they are buying EPYC . Now if you want to take the discussion further saying how much thats another point altogether. But your earlier statement that these companies are not buying EPYC is false. As for AMD EPYC sales the increase in market share would not be a step function but a gradual increase. Anyway AMD has said they expect 10% market share by 2019. I am thinking they were conservative. AMD should get to double digit market share in 2018.

The real impact will be when 7nm Zen 2 arrives. AMD Zen was built on a 14LPP process designed for mobile SoC. GF 7LP is a process designed for very high performance and 5 Ghz operation. GF 7LP is also much more comparable to Intel 10nm in terms of density. I have put my calculations based on GF statements on the % of area shrink.

https://forums.anandtech.com/thread...infet-process-and-fx-7-asic-platform.2508687/
https://forums.anandtech.com/thread...and-fx-7-asic-platform.2508687/#post-38937252

Intel 14nm CP = 70nm , MMP = 52nm => Cell Area = 70 * 52 = 3640
GF 14 LPP CP = 78nm , MMP = 64nm => Cell Area = 78 * 64 = 4992

Intel 10nm CP = 54nm, MMP = 36nm => Cell Area = 54 * 36 = 1944
GF 7LP CP = 56nm, MMP = 40nm => Cell Area = 56 * 40 = 2240

Intel 14nm was 73% of GF 14LPP in terms of cell area. Intel 14nm was also very high performance which allowed 5 Ghz CPU designs. Intel 10nm is only 87% of GF 7LP in terms of cell area. Transistor performance of 7LP is also very competitive with Intel 10nm with high performance version of 7LP optimized for 5 Ghz operation. At 14LPP AMD had to make tradeoffs in terms of density and performance while at 7LP they will have the best combination of both performance and density. With Zen 2 at 7LP and Zen 3 at 7LP+ with EUV I see AMD being able to reach >20% server market share. btw for AMD this is not a sprint race but a marathon. AMD is back in HPC but more importantly they are back to stay.

 

[Hitman928]

Yesterday's momentum carried into today for AMD's stock which is up 8% so far today.

 

[Kenmitch]

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.

Not saying it's true or not true....But the way Zen/Ryzen/Threadripper/EPYC was designed looks to a very future looking way of doing things. It's possible that by design it was for the exact purpose of yields and waste in the past.

Looks to be a disruptive product no matter what you call it in the end. Not really sure why the hate for AMD carries on so strongly with some members. It's somewhat ironic as if nothing more it forces Intel to get out of the stagnation mode they've been in for a decade.

Maybe it's more of a AMD did it 1st situation?

 

[scannall]

Not saying it's true or not true....But the way Zen/Ryzen/Threadripper/EPYC was designed looks to a very future looking way of doing things. It's possible that by design it was for the exact purpose of yields and waste in the past.

Looks to be a disruptive product no matter what you call it in the end. Not really sure why the hate for AMD carries on so strongly with some members. It's somewhat ironic as if nothing more it forces Intel to get out of the stagnation mode they've been in for a decade.

Maybe it's more of a AMD did it 1st situation?

I don't really get it. It seems some people are really mad that AMD has produced a compelling product. I buy whatever product is best for ME, and whatever my goal is. To have the 'sports team' loyalty to a company and product brand just doesn't make any sense.

 

[formulav8]

'm a software engineer working on the code optimizer in one of the major C++ compilers out there and I think what AMD did with the results is justified. The Intel compiler is indeed "optimized for SPEC" - it employs optimizations that are either illegal in a language such as C++ or not applicable to pretty much any real program outside the SPEC benchmarks that is larger than a few hundred lines of code. You have to use a magical combination of flags to get anywhere close to the numbers they publish, and if you try that on other programs you either don't see any improvements or might introduce runtime bugs. GCC is overall the most suitable compiler for systems and server software right now.

Thanks for the info. I still think AMD could have showcased the CPU differently for PR reasons.

 

[lolfail9001]

It's possible that by design it was for the exact purpose of yields and waste in the past.

Oh, make no mistake, that design is probably the best trade-off between yields, development cost and performance AMD could theoretically have. But there's a difference between praise and making up BS for the sake of praise. I take issue with latter.

Anyways, only thing remaining to learn is what performance does it show off, waiting for STH review here.

Thanks for the info. I still think AMD could have showcased the CPU differently for PR reasons.

I still do not see the reason why AMD did not just use their own optimizing compiler and Intel's certified results in straight comparison. Rather, i suspect it but obviously i will be called an AMD hater again if i name it.

 

[raghu78]

Not saying it's true or not true....But the way Zen/Ryzen/Threadripper/EPYC was designed looks to a very future looking way of doing things. It's possible that by design it was for the exact purpose of yields and waste in the past.

Looks to be a disruptive product no matter what you call it in the end. Not really sure why the hate for AMD carries on so strongly with some members. It's somewhat ironic as if nothing more it forces Intel to get out of the stagnation mode they've been in for a decade.

Maybe it's more of a AMD did it 1st situation?

AMD's approach with Zen and Infinity Fabric is something Intel is going to follow with EMIB ( Embedded Multi-die Interconnect Bridge ) and server first at 10++ . The approach of stitching multiple dies over a single monolithic die has become necessary to avoid low yields and to improve time to market. AMD is likely to also improve the stitching of dies by going for more power efficient die packaging solutions like silicon interposer, SLIM (Silicon Less Integrated Module) , EMIB.

https://en.wikichip.org/wiki/heterogeneous_integration
http://www.3dincites.com/wp-content/uploads/slim-swift-customer-overview-may-13-2015.pdf

 

[Oddzz]

Those 3 research papers where mentioned in this Forbes article about Epyc. I thought they could be of interest for some of you:
AMD EPYC Brings New RAS Capability
SPEC CPU 2017 & Changing Performance
Power / Performance Determinism

 

[Shlong]

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.

From reddit: https://www.reddit.com/r/Amd/commen...finity_fabric_and_to_high_yields_amd/dj7ctka/

It's not just possible, but quite probably to be over 95%.

The Ryzen 8-core die should have a perfect yield of 86.1% on 14nm LPP at Global Foundries (defect density of 0.08 per cm2).

That leaves 41 partially damaged die. 60% of them should be usable as six core CPUs considering 2/3rds of the die is a CCX. That is to say that the L3 and uncore are completely undamaged.

That leaves 25 die with damaged L3 or uncore. 50/50 split between what is damaged. AMD sales those with damaged L3 as a Ryzen 5 1400.

That leave 12/13 die with potentially fatal damage. That's a 95.7% effective yield.

The GPIO area is fairly large, taking about 15~20% of the uncore area(including related SRAM). That same percentage of die should be usable as desktop Ryzen (potentially even as 8-core CPUs).

We now have 10 die that are bad from a wafer of 295 die. Or a nearly 97% effective yield.

Effective yield could jump to 99.9% from that baseline with a defective density only marginally better than 0.08/cm2.

That would mean 294 die used per wafer - which means AMD is spending like $30 per die instead of $34 per die using only the perfect ones.

 

[lolfail9001]

From reddit: https://www.reddit.com/r/Amd/commen...finity_fabric_and_to_high_yields_amd/dj7ctka/

Uhm, since when is 2/3rds of the die are CCXs? Die size is ~195mm^2, 2 CCXs total 88mm^2. Unless i've been missing on something, 195*2/3=130, if anything 2/3rds of the die are L3+non-CCX parts. Similar with split.

Of course in the end it does not quite matter in this math because it is a difference of 4 dies, or about 1.3% if i assume that otherwise it was flawless and sure at such defect density 95-97% yield sounds reasonable. But difference between 97% and 99.9% yield is actually massive and i hope we all understand why.

P. S. where the hell did looncraz pull that defect density figure from?

 

[Space Tyrant]

I still do not see the reason why AMD did not just use their own optimizing compiler and Intel's certified results in straight comparison. Rather, i suspect it but obviously i will be called an AMD hater again if i name it.

For the same reason that you or I did not bench these CPU's with our own optimizing compiler; we don't have the resources to compete with Intel. AMD is competing in the CPU market here and carefully allocating its limited engineering resources.

If you're interested in a tuned combination of compiler and CPU, Intel's set just might be for you. If, on the other hand, you're only interested in how well the hardware stands on its own based on software you buy or make with some other compiler, you wouldn't be interested in the otherwise unused Intel compiler options.

We're talking about the CPU's here; the highly optimized-for-Intel compiler should only be relevant in proportion to its market penetration in your areas of interest. It's basically irrelevant to my interesets. Like many here, I only care about their hardware.

 

[crashtech]

Quote where they said they are buying them please.

Since you are asking this, you must certainly be ready to provide substantiation of your claim as well.

 

[Markfw]

Quote where they said they are buying them please.

You still don't understand marketing speak.

Anyway, of course they may buy some Epyc Servers. The question is how many.

Do you think the major cloud providers will making a move away from Intel? In what timeframe?

Soo this post ?

Analyst Vivek Arya lowers Intel’s data center sales growth for FY18 from 11% to 8% and notes that AMD’s Epyc has signed up Microsoft and Baidu among its big-name data center customers.

So if you sign up with AMD, I think its for sure you are confirmed to be buying from them. How many ? I don't know, but the investors sure think AMD will outsell Intel in a market that includes EPYC. That tells me they will buy a lot more than before.