AMD's EPYC CPU'S with Benchmarks & Pricing!

[PCGuy_5960]

3 minutes ago, Spannerhands said:

 

Underwhelming... I hope that they perform better, because that score is pretty bad..... A 44C/88T Broadwell server can almost match it

 

[Spannerhands]

6 minutes ago, PCGuy_5960 said:

Underwhelming... I hope that they perform better, because that score is pretty bad..... A 44C/88T Broadwell server can almost match it

Well its not really "pretty bad" is it... Scores inside top 10 best world record on cinebench with no optimisation brand new computer architecture fresh out of the box...  I would say that's pretty good... 

[PCGuy_5960]

Just now, Spannerhands said:

Well its not really "pretty bad" is it... Scores inside top 10 best world record on cinebench with no optimisation brand new computer architecture fresh out of the box...  I would say that's pretty good... 

I didn't say that it's a bad score, but it should score a lot higher given that it has 64 cores and a 44 core machine can almost match it...

 

I mean, it does have 45% more cores than a dual 2699 v4 PC, the score should be at least 30% higher. Let's hope that optimizations can improve performance

[The Benjamins]

5 minutes ago, PCGuy_5960 said:

I didn't say that it's a bad score, but it should score a lot higher given that it has 64 cores and a 44 core machine can almost match it...

 

I mean, it does have 45% more cores than a dual 2699 v4 PC, the score should be at least 30% higher. Let's hope that optimizations can improve performance

I think it is interesting that he got a 10% uplift from 2012 to 2016. could scaling become a issue, it has 40 more threads to manage?

[PCGuy_5960]

1 minute ago, The Benjamins said:

I think it is interesting that he got a 10% uplift from 2012 to 2016. could scaling become a issue, it has 40 more threads to manage?

I don't know.... I believe that AMD's infinity fabric may not scale very well with more than 8 cores.

[leadeater]

Just now, The Benjamins said:

I think it is interesting that he got a 10% uplift from 2012 to 2016. could scaling become a issue, it has 40 more threads to manage?

 

Just now, PCGuy_5960 said:

I don't know.... I believe that AMD's infinity fabric may not scale very well with more than 8 cores.

Last I saw a Cinebench run with EYPC and core utilization monitoring it was having issues actually using all 64.

[Jahramika]

33 minutes ago, PCGuy_5960 said:

Underwhelming... I hope that they perform better, because that score is pretty bad..... A 44C/88T Broadwell server can almost match it

 

Whats that like $9000 for a 44 core CPU?

[PCGuy_5960]

1 minute ago, leadeater said:

Last I saw a Cinebench run with EYPC and core utilization monitoring it was having issues actually using all 64.

Source?

BTW, a 64 core Haswell machine scores 6100+ in Cinebench:

So EPYC should be able to get 6200 or higher..... 

[PCGuy_5960]

Just now, Jahramika said:

Whats that like $9000 for a 44 core CPU?

2x22core CPUs, but those are Broadwell Xeons. Skylake Xeons will be the real competitor for EPYC

[Dylanc1500]

15 minutes ago, PCGuy_5960 said:

2x22core CPUs, but those are Broadwell Xeons. Skylake Xeons will be the real competitor for EPYC

Shove 8 of the Xeon 8180's, all of them all core boosting to 3.2, in one box and call it good.

[PCGuy_5960]

1 minute ago, Dylanc1500 said:

Shove 8 of the Xeon 8180's in one box and call it good.

Unless Intel decides to make a 34 or 36 core CPU

[The Benjamins]

21 minutes ago, PCGuy_5960 said:

2x22core CPUs, but those are Broadwell Xeons. Skylake Xeons will be the real competitor for EPYC

wccf says in this article that this german source got 6879

http://wccftech.com/amd-epyc-7601-benchmarks-analysis/

https://www.heise.de/newsticker/meldung/AMD-Epyc-legt-los-3748615.html

 

That would put it were I would expect it. comapred to the 4x Xeon 8357 V3 with a score of 6139

[PCGuy_5960]

1 minute ago, The Benjamins said:

wccf says in this article that this german source got 6879

Yeah, makes sense....  Although, this was also in their article:

So Intel still beats AMD with 33% fewer cores (that are clocked 22% higher)

 

Idk, I hope that optimizations can improve EPYC's performance

[Taf the Ghost]

Cinebench has issues scaling somewhere over 16c.  This came up when when some of the Xeon scores started cropping up recently.  At that many cores, other issues start cropping up.  Blender seems to hit a wall somewhere in the 20s, I believe, as well.

 

When you get out to that many cores, other things beyond the core's performance are cropping up. Still, fun to see server hardware crunching stuff like that.

[The Benjamins]

Just now, PCGuy_5960 said:

Yeah, makes sense....  Although, this was also in their article:

So Intel still beats AMD with 33% fewer cores (that are clocked 22% higher)

 

Idk, I hope that optimizations can improve EPYC's performance

ya intel will win in the top end but again AMD can price them lower, and deliver a lot more I/O.

 

The big point is that they can compete again in server CPU compute tasks. it is looking awesome.

[Dylanc1500]

3 minutes ago, The Benjamins said:

ya intel will win in the top end but again AMD can price them lower, and deliver a lot more I/O.

 

The big point is that they can compete again in server CPU compute tasks. it is looking awesome.

The only issue I personally have with the I/O on EPYC is when you have a dual system you still only have 128 lanes. The way the system communicates between sockets doesn't allow for scaling up I/O.

[The Benjamins]

1 hour ago, Dylanc1500 said:

The only issue I personally have with the I/O on EPYC is when you have a dual system you still only have 128 lanes. The way the system communicates between sockets doesn't allow for scaling up I/O.

Ya but at least it gets you more then any other 2 CPU config.

[leadeater]

4 hours ago, Dylanc1500 said:

The only issue I personally have with the I/O on EPYC is when you have a dual system you still only have 128 lanes. The way the system communicates between sockets doesn't allow for scaling up I/O.

I don't mind since 128 is still a lot and it makes motherboards feature consistent across single socket and dual socket.

[Belgarathian]

At those launch prices that's almost ~50% of Intel's RRP in some cases.... 

[The Benjamins]

5 minutes ago, leadeater said:

I don't mind since 128 is still a lot and it makes motherboards feature consistent across single socket and dual socket.

due to my bad foresight, I am considering upgrading my server to the gigabyte Epyc board with a 16 or 24c Epyc. and moving from a 1U chassis with external sas drives to a 2U chassis with 8 HDD bays.

 

this is all due to the fact I bought into 10Gbpe RJ45 not realizing SFP+ would of been a better route for switch options. so when I want to more to 10Gbps I need to get a $500+ switch or change my server (replace motherboard and all)

 

going the Epyc route will cost more but at lease it would correct my mistake better, and be more versatile.

[leadeater]

53 minutes ago, The Benjamins said:

due to my bad foresight, I am considering upgrading my server to the gigabyte Epyc board with a 16 or 24c Epyc. and moving from a 1U chassis with external sas drives to a 2U chassis with 8 HDD bays.

 

this is all due to the fact I bought into 10Gbpe RJ45 not realizing SFP+ would of been a better route for switch options. so when I want to more to 10Gbps I need to get a $500+ switch or change my server (replace motherboard and all)

 

going the Epyc route will cost more but at lease it would correct my mistake better, and be more versatile.

Keep an eye on supermicro motherboards too, they might come out with something that could have better features on it like an actual SAS controller.

[hoohoo]

On 6/20/2017 at 1:19 PM, SC2Mitch said:

Oh, my point still stands, does any manfuacture do that kind of nuts amount lol and what do you need 2tb of DDR4 memory for, Chrome Machines?

You operate an online company that does e-commerce, or perhaps offers an online comment/community system. Your products, customers, comments, users, etc are in a DBMS. Your DBMS has grown to a terabyte or more. You want to use in-memory tables because disk tables are too slow. You need several TB of RAM.

 

If you are running a comment/community system then your DBMS will be many TB. You want to use in-memory tables because disk tables are too slow. You need several TB of RAM.

 

You are either of the above and you decide to use noSQL key/value stores to cache your DBMS. You need several TB of RAM.

 

You are a researcher modelling something that has a large phase space (physics, chemistry, web analytics, bio-sci, finance, ...) your data sets will easily be many hundreds of GB to many TB. You need several TB of RAM.

 

You are a cloud operator, and yes, you need to host millions of virtual machines. You also need them encrypted, and you need boot-level control over which VM images are allowed to boot where*. You need several TB of RAM.

 

Now, then again, maybe you are a callow young gamer: _you_ only need about 16 GB RAM.

 

*) You need enterprise RAM capacity and you need enterprise boot crypto. Currently AMD has the upper hand on both compared to Intel.

[AnonymousGuy]

8 hours ago, hoohoo said:

*) You need enterprise RAM capacity and you need enterprise boot crypto. Currently AMD has the upper hand on both compared to Intel.

Not at all. Intel has been in the enterprise data protection game a lot longer than AMD and there's a lot more aspects to it then just SEV and SME from AMD which already have theoretical attack vectors, and it's not even released.

[leadeater]

5 hours ago, AnonymousGuy said:

Not at all. Intel has been in the enterprise data protection game a lot longer than AMD and there's a lot more aspects to it then just SEV and SME from AMD which already have theoretical attack vectors, and it's not even released.

Those are or will be required features for virtual hosts that will have data with regulatory access restrictions like medical data. There is a lot of data that cannot be hosted out of country here or can only if there is an in country copy.

 

The local hosting providers have dedicated virtual hosts and storage arrays for Government regulated data and are very isolated from general customer VMs/data. It's not all about classified/secret data some of it falls under strict protection rules due to Privacy Acts.

 

If you want Gov hosting contracts you're going to need those types of features, it won't be an option or selling point but a requirement. There will also be non Gov entities looking for those protections too.

 

This isn't exclusive to AMD, new Intel based servers have similar features too.

 

Intel hasn't really made anything more secure, good system architects using good practices and isolation techniques have been keeping things secure. Good security and protection often has very little to do with hardware features.

[Whispre]

On ‎6‎/‎21‎/‎2017 at 6:27 AM, tom_w141 said:

Going back to this point real quick yes they do. It is in the best interest of any business to save cost wherever possible. I should know I do it for a living. (Spoiler alert all workstations here are Xeons but obviously they are old and on their next upgrade cycle they likely won't be Xeons unless there is a compelling reason to spend the money ofc.

20 years as a senior SysAdmin in a multibillion dollar company, and I can say that Price is not the top of our list... more like:

 

1. reliability

2. compatibility/feature set/ supporting hardware features

3. performance

4. power draw

5. availability (need to be able to order/build without delays)

6. support (vendor and application)

7. Price