AMD Epyc 7000 series revealed

[aakopa]

Just now, domandric034 said:

No no no. 6 NVMe slots + fan that cools those SSD's and you have choice for all RAID's and it doesnt have hang up on motherboard about RAID types

But one nvme drive needs 4 lanes, so 4x would already use up all those 16 lanes offered by the slot

[corsairian]

64 threads, 128 PIC E lanes, 128gb of ram (256gb maybe?) 

 

I want an EPYC. The 16 core one is sticking out to me. This may be my first time going AMD. 

 

I really want to see a Dual Processor board.. Maybe quad processor boards?! super micro? Lol. 

[domandric034]

Just now, aakopa said:

But one nvme drive needs 4 lanes, so 4x would already use up all those 16 lanes offered by the slot

Oh ok. Bad math but still.

 

Just now, DXMember said:

all RAIDs?? even RAID-7 and RAID-53 ?

RAID 0, RAID 1, RAID 5, RAID 10, RAID 50, RAID 6......

[DXMember]

Just now, domandric034 said:

So he can shit on Intel better and have better dreams of shitting on Intel Optane.

na, Optane is just too good, ain't nottin' shit on Optane... except for Intel by gating it behind dongles

[domandric034]

Just now, DXMember said:

na, Optane is just too good, ain't nottin' shit on Optane... except for Intel by gating it behind dongles

Don't worry. AMD will come up with something to shit on optane. They wont let it pass untouched or unshitted.

[DXMember]

4 minutes ago, corsairian said:

64 threads, 128 PIC E lanes, 128gb of ram (256gb maybe?) 

 

I want an EPYC. The 16 core one is sticking out to me. This may be my first time going AMD. 

 

I really want to see a Dual Processor board.. Maybe quad processor boards?! super micro? Lol. 

they were live benchmarking EPYC (Naples back then) against Xeons - AMD was running 512GB, Intel is limited to 384GB,

the best part was when they loaded up a benchmark that utilized 480GB and said Zen is like mega infinity times faster at this workload cuz the damn Intel can't even load that shit

 

[Dylanc1500]

38 minutes ago, DXMember said:

they were live benchmarking EPYC (Naples back then) against Xeons - AMD was running 512GB, Intel is limited to 384GB,

the best part was when they loaded up a benchmark that utilized 480GB and said Zen is like mega infinity times faster at this workload cuz the damn Intel can't even load that shit

 

Where do they get that Intel is limited to 384? With the availabity of 3 128gb DIMMs per channel, at six channels, comes to a whopping 2,304gb. That's just for a single chip system. I'm a little slightly confused by that statement. If you don't mind, can you please clear that up?

[2bitmarksman]

8 hours ago, DeezNoNos said:

Oh god. intel what u gonna do

Your profile pic is a great summation of that statement

[AnonymousGuy]

3 hours ago, DXMember said:

you are missing the:

and that 8 core EPYC is still EPYC 7000 Processor which will have 8 Channels DDR4 and 128 PCIe Lanes

Each die only has access to 2 channels.  AMD is sorta lying about it being 8 channels.  I also had a long discussion in another thread about 128PCIe lanes oversubscribing the DDR4 bandwidth as well.  

 

EDIT: to elaborate on the oversubscription problem, Skylake Xeon's have 128GB/s of dram bandwidth for 48 lanes of PCIe requiring 96GB/s bandwidth.  Epyc each die only has 42.7GB/s dram bandwidth for 64GB/s PCIe bandwith in 32 lanes.

 

These things aren't going to sell particularly well due to the inefficiencies of having 4 packages to reach 32C.

[Jito463]

4 hours ago, DXMember said:

there's a 10 core Threadripper with 3+2 + 3+2 CCX configuration

as well as 14 core Threadripper with 3+4 + 3+4

It would make more sense if they were 2+2/3+3 or 3+3/4+4, so that each die has a matched pair of CCX's.

[Taf the Ghost]

2 minutes ago, Jito463 said:

It would make more sense if they were 2+2/3+3 or 3+3/4+4, so that each die has a matched pair of CCX's.

They can run unbalanced cores inside a single package, it's just a matter if they see the use of supporting it.  (There were engineering samples of 3+1 and 4+2, apparently.) It still seems like the 10c & 14c parts might not happen.  We'll see, but, at this point, I'd be rather surprised if we get them.

[leadeater]

12 hours ago, rattacko123 said:

That TDP looks a bit too high for the low core count models, especially considering the low clock speed. 
The clockspeed is lower than intels which based on threadripper leaks shouldn't be happening, and if true AMD has little chance of beating intel. Unless of course the server processors are way downclocked for some reason even though they are supposed to be better binned.

Current Intel Xeons with 10 less cores have the same base clocks but boost a bit higher, not on all cores though.

 

Edit:

And they have to maintain the 4 dies on the low core count models for the 8 channel memory I suspect which would increase the TDP.

[leadeater]

12 hours ago, Bananasplit_00 said:

true but id love to see someone try to get one of the 180W chips in a 1u chasies with air cooling, i dont know if there is even a way to do that lol. probably a huge ass vapor chambre and then just heatsink in the rest of the 1u lol

There are 1u servers right now with two 160W Xeons in them, also TDP doesn't mean much at all considering Intel and AMD measure them differently and they don't even reflect power usage either.

[rattacko123]

5 minutes ago, leadeater said:

Current Intel Xeons with 10 less cores have the same base clocks but boost a bit higher, not on all cores though.

 

Edit:

And they have to maintain the 4 dies on the low core count models for the 8 channel memory I suspect which would increase the TDP.

The issue here is the same clocks. Since Intel has higher IPC, AMD might be at a disadvantage here despite having better memory controllers. I was expecting AMD to have higher clock speeds on their server CPUs compared to Intel, based on how efficient the 1700 is.

Intels new platinum Xeons have like 2.5Ghz base clock or something with 28 cores, which I guess is an invalid argument since they are incredibly expensive anyways  
 

[leadeater]

Just now, rattacko123 said:

The issue here is the same clocks. Since Intel has higher IPC, AMD might be at a disadvantage here despite having better memory controllers. I was expecting AMD to have higher clock speeds on their server CPUs compared to Intel, based on how efficient the 1700 is.

Intels new platinum Xeons have like 2.5Ghz base clock or something with 28 cores, which I guess is an invalid argument since they are incredibly expensive anyways  
 

I was hoping for higher base clocks too but I was rather cautious about it since I knew the chips would have to be power limited which would restrict base and boost clocks.

[SpaceGhostC2C]

2 hours ago, AnonymousGuy said:

Each die only has access to 2 channels.  AMD is sorta lying about it being 8 channels.

 

(...)

 

These things aren't going to sell particularly well due to the inefficiencies of having 4 packages to reach 32C.

 

I'm not sure Intel agrees with you on the topic of multiple IMCs:

 

[ARikozuM]

12 hours ago, Bananasplit_00 said:

true but id love to see someone try to get one of the 180W chips in a 1u chasies with air cooling, i dont know if there is even a way to do that lol. probably a huge ass vapor chambre and then just heatsink in the rest of the 1u lol

It's usually easier to funnel air through a server than it is with a desktop.  2U will likely be the best spot for these CPU's. 

[leadeater]

6 minutes ago, SpaceGhostC2C said:

 

I'm not sure Intel agrees with you on the topic of multiple IMCs:

 

I'm also not sure how actually having 8 channels can be a lie, it has them...

 

Do people actually want AMD to just photocopy Intel designs?

[Lays]

12 hours ago, TOMPPIX said:

can't wait for the 15000€ price tag. 

implying prices matter when people are looking at these types of applications lol

[AnonymousGuy]

 

6 minutes ago, leadeater said:

I'm also not sure how actually having 8 channels can be a lie, it has them...

 

Do people actually want AMD to just photocopy Intel designs?

They're trying to imply that their 8 channels is "better" than Intel's 6 channels, so it's not a lie exactly but it's definitely deceptive.  Just like they're trying to sell it as a "32C" when really it's more like a 4 socket 8 core system in terms of how much optimization you'd have to do to actually get good performance out of it.

 

AMD marketing man...don't trust it.

[Clanscorpia]

12 hours ago, revsilverspine said:

not to mention they could also pull another dick move and make it a 4-cpu config only thing, so you have to buy 4 of them to use them on a $5000 motherboard.

But thats the optimal config for a server, why on earth would a consumer be buying a 34 core CPU

[leadeater]

Just now, AnonymousGuy said:

 

They're trying to imply that their 8 channels is "better" than Intel's 6 channels, so it's not a lie exactly but it's definitely deceptive.  Just like they're trying to sell it as a "32C" when really it's more like a 4 socket 8 core system in terms of how much optimization you'd have to do to actually get good performance out of it.

 

AMD marketing man...don't trust it.

Well we don't know how it performs yet at all and architecturally it's not like a 4 socket system, 4 dies != 4 sockets. It's how you connect them that's important and what kind of cache access each die has etc, all the fine details we don't have.

 

And saying it's not 8 channels is no different so saying 4WD isn't a proper 4 wheel drive system and AWD drive is, both drive 4 wheels and one is better but neither are lyiing.

[AnonymousGuy]

12 minutes ago, leadeater said:

Well we don't know how it performs yet at all and architecturally it's not like a 4 socket system, 4 dies != 4 sockets. It's how you connect them that's important and what kind of cache access each die has etc, all the fine details we don't have.

 

And saying it's not 8 channels is no different so saying 4WD isn't a proper 4 wheel drive system and AWD drive is, both drive 4 wheels and one is better but neither are lyiing.

We do know the details of their architecture, because we already have Zeppelin and they're not redesigning it for Epyc.  The only thing we don't know yet is the latency of their interconnect, but there's no pixie dust that can make it 0nS even if it has all the bandwidth in the world.

 

And yes, for this purpose 4 dies is the same as 4 sockets.  Zeppelin chip #1 isn't going to have direct access to L3 cache on Zeppelin chips #2-4 as it would on a monolithic design.  Can you optimize your workload to keep work on as few CCX and die domains as possible?  Yeah, you *could*, but no one wants to because it's a pain in the ass.  What is going to happen is Skylake X has perfect scaling from 1-28 cores and Epyc is going to be all over the place depending on what you're doing.

[leadeater]

9 minutes ago, AnonymousGuy said:

We do know the details of their architecture, because we already have Zeppelin and they're not redesigning it for Epyc.  The only thing we don't know yet is the latency of their interconnect, but there's no pixie dust that can make it 0nS even if it has all the bandwidth in the world.

 

And yes, for this purpose 4 dies is the same as 4 sockets.  Zeppelin chip #1 isn't going to have direct access to L3 cache on Zeppelin chips #2-4 as it would on a monolithic design.  Can you optimize your workload to keep work on as few CCX and die domains as possible?  Yeah, you *could*, but no one wants to because it's a pain in the ass.  What is going to happen is Skylake X has perfect scaling from 1-28 cores and Epyc is going to be all over the place depending on what you're doing.

Um the way the Infinity Fabric is designed yes it does have access to the L3 cache of any die, that is why it exists.

 

Zeppelin is EPYC, code name vs product name. Nothing detailed as ever been show about how the 4 dies are interconnected other than us being told it's using the Infinity Fabric and how that works.

 

There is no doubt that the Intel products that have the same amount of cores and cost more will perform better, there will even be ones with less cores and perform better. What we don't know is how each will stack up at the same price, that is where it is important.

 

Many servers do not get used anywhere near the peak performance capability so sitting around debating the best theoretical synthetic performance is not useful, if I had a choice between a 20K AMD server and a 30K Intel server and both will more than serve my needs why would I pay 10k more? Do I get 10k more benefit?

 

If AMD can deliver the required performance at the right price why should we be saying that their architecture is inferior? It would have meet all the requirements so where is the issue?

[JuNex03]

6 hours ago, DXMember said:

they were live benchmarking EPYC (Naples back then) against Xeons - AMD was running 512GB, Intel is limited to 384GB,

the best part was when they loaded up a benchmark that utilized 480GB and said Zen is like mega infinity times faster at this workload cuz the damn Intel can't even load that shit

 

 

Lol "Insufficient Memory to Load²". Its like they're saying it's not just shit, but Shit²