AMD Launches EPYC Enterprise CPUs Available Today @ [H]

could just use pci-e extensions.

I don't see how is that viable for the market that Motherboard is aimed for. as lutjens I'm guessing the same, lot of issues with GPUs and RAM. specially with server grade, Pro cards that are typically big ass passively cooled.
 

I don't mean to be unreasonable but that Intel slide you quoted for the next Xeon (Purley CPUs) shows 48 lanes/CPU as the plan. So 4 more lanes than what we are getting from Skylake/Kaby Lake-X. So I think that my point that Intel and AMD seem to disagree about the number of lanes that people need still stands. Just wanted to get some perspective from people who actually use server equipment as to whether Intel or AMD is more right about the number of lanes people need.

EDIT: added quote.
 
Last edited:
This is probably partially because AMD plans to keep the same platform for longer than Intel. This has its advantages and disadvantages (as we seen with AM3+).

Will the 8C support the same # of lanes?
 
Thanks for the responses. Just wanted to kind of explain where my thought process was coming from. So current Intel HEDT offerings have something like 20 lanes and the upcoming Skylake/Kaby Lake-X have a maximum of 44 lanes on the top end chips. My thought process was if that Intel thinks that 44 lanes is enough for their top end offering-- 64 lanes on threadripper should be sufficient. Same thing with the memory as the top offering from is limited to quad channel. Clearly I was mistaken.

Is Intel just making a mistake then about the number of lanes and memory support? Is buying EPYC a matter of future proofing or do you all think that you will be using all the lanes/memory capacity immediately?

The outgoing Intel 6950X has 40 PCIe lanes on the CPU and 8 available via the X99 chipset. I don't know where you are getting "20" lanes from. There are 28 lane CPUs like the 6820K. With the X99 chipset, you get an additional 8 lanes for a total of 36 PCIe lanes. What I've seen thus far indicates that Skylake-X Core i7 78xx series chips will have 28 PCIe lanes while the i9 7900X will have 44 PCIe lanes. Kaby Lake-X is the useless fucker with only 16 PCIe lanes. X299 itself provides up to 24 PCIe lanes, bringing the total for the Core i9 7900X to 68.

That's in the HEDT segment. The server market could play out differently. I don't keep track of the server chipsets beyond the rebranded ones for the Xeon, which are essentially the same as the X99, X299 chipsets. On the server side, you can't really ever have too much memory. 8-channel memory can add to memory performance, which is important in some cases. That said, I think it's more for future proofing and simply to win benchmarks over Intel's Quad-Channel memory controllers. As for the PCIe lanes, it may be more important to extend the amount past where Intel wants to as the market starts going towards PCIe based storage devices. In servers, you can't ever have too much storage capacity..
 
For those of you planning on buying Epyc as a home PC (especially those interested in the 16 core Epyc), I am just wondering what you plan to do that would benefit from the features. Unless you plan to go with the more expensive 32 core one, the main difference is the memory support. Threadripper supports quad-channel (max 128 GB) but Epyc supports 2TB of RAM. There are also more PCI lanes but I would have thought that 64 lanes from threadripper would be enough. Just wondering what people are doing at home that requires 2TB RAM and 128 PCIe lanes.
Virtualization primarily and more specifically storage and virtualization combinations. Current storage systems utilize a lot of memory. In addition to that available IO bandwidth decreases with each VM you add.
 
OpterOWN has returned as EPYC.

I'm excited to see the CPU wars again. It's been too long.
 
Good to see the way this is going, I thought infinity fabric was some marketing hoo-haa, but looks like it's actually good.
 
I thought Xeon was Intel's server chip? I'm not as familiar with the server side of things being more used to the consumer side of things. These Xeon chips were just released in Feb of this year with 32 lanes. This is an $8,000 chip with only 32 lanes released a few months ago. Other Xeons were released in March with between 20-32 lanes as well.

https://ark.intel.com/products/96900/Intel-Xeon-Processor-E7-8894-v4-60M-Cache-2_40-GHz
https://www.theverge.com/circuitbre...6/intel-xeon-e7-8894-v4-processor-9000-server

Those scale up to 8 sockets. That's why they got less. Else you get 40 on the Broadwell platform. but Purley have now shipped for around 10 months to the top dogs and will get released for the rest on July 11th.

For a 2 socket system you tend to have at least 80 lanes without chipset. Or the ability to move roughly 80GB data a second. For most servers you have a few NIC ports and a raid controller or some NVME PCIe based SSDs, usually x8.
 
I don't mean to be unreasonable but that Intel slide you quoted for the next Xeon (Purley CPUs) shows 48 lanes/CPU as the plan. So 4 more lanes than what we are getting from Skylake/Kaby Lake-X. So I think that my point that Intel and AMD seem to disagree about the number of lanes that people need still stands. Just wanted to get some perspective from people who actually use server equipment as to whether Intel or AMD is more right about the number of lanes people need.

EDIT: added quote.

4 lanes will be used for chipset. Same with AMD. If you skip the chipset you got 4 lanes extra directly from the CPU.
 
So which today are we talking about when you say of "Available Today?"

Next month? Next year? Yesterday? There's certainly no pre-orders with a delivery date of next week at Newegg, like the X299 platform.
 
Last edited:
So which today are we talking about when you say of "Available Today?"

Next month? Next year? Yesterday? There's certainly no per-orders at Newegg like the X299 platform.
They said you could guy them today. Check with Dell or HP?
 
AMD is off to a decent start in their goal to re-enter the datacenter and high performance computing market. I think AMD will have moderate success with Zen and EPYC. AMD has said they expect to get to 10% server market share in 2019. My opinion is thats a conservative estimate and feel AMD will get to double digit market share in H2 2018 . But I think the 7nm Zen 2 and Rome in 2019 is when AMD starts to really make serious inroads into the server and HPC market and get to >20% server market share. There are reasons for that

1. Firstly AMD will be using GF 7LP which is much more competitive vs Intel 10nm in terms of transistor density and performance. GF 7LP has a high performance version optimized for 5 Ghz operation. This is exactly what AMD needs to make a serious impact in high end desktops and servers.
2.Secondly Zen being a brand new architecture and Infinity Fabric being a completely new system design have lots of room for improvement. There is typically a lot of low hanging fruit in a brand new architecture. AMD should be able to address the biggest weakness of Zen while continuing to improve on its strengths.

Overall there is excitement in the PC and server industry that AMD is bringing competition and true choice to customers. Thats a welcome change from the past 5 years where Intel was the only option. :)
 
1. Firstly AMD will be using GF 7LP which is much more competitive vs Intel 10nm in terms of transistor density and performance. GF 7LP has a high performance version optimized for 5 Ghz operation. This is exactly what AMD needs to make a serious impact in high end desktops and servers.
I swear, your process overhyping will end just like it did with Polaris, with you having a meltdown over disappointment of AMD not meeting your expectations. There's trying to hype up a turd and there's trying to overhype a sincerely decent product. Guess which one is worse.
 
I swear, your process overhyping will end just like it did with Polaris, with you having a meltdown over disappointment of AMD not meeting your expectations. There's trying to hype up a turd and there's trying to overhype a sincerely decent product. Guess which one is worse.

You might want to say that GF 7LP is a turd without any proof but I think GF has improved its execution vastly over the 14nm node where they had to license Samsung 14LPP node after they failed to deliver their in-house 14XM node. The IBM semiconductor acquisition has provided GF the talent and know-how to deliver on a cutting edge high performance process. If you want to argue that GF is lying on information provided to their customers and the press then thats your opinion. Anyway I will just let others decide for themselves

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

GF 7LP is ready for customer designs and AMD has confirmed 7nm tapeouts later this year.
 
Last edited:
  • Like
Reactions: N4CR
like this
Contacted supermicro and HP, they said no available pre orders yet. I'm sure they have them but only for contracts.

I didn't ask HP but supermicro said "no date" for individual systems.

Supermicro only lists some uninteresting boards for now, nothing else. These are for pre order. Unless it's in some advertising i missed.
 
You might want to say that GF 7LP is a turd without any proof but I think GF has improved its execution vastly over the 14nm node where they had to license Samsung 14LPP node after they failed to deliver their in-house 14XM node. The IBM semiconductor acquisition has provided GF the talent and know-how to deliver on a cutting edge high performance process. If you want to argue that GF is lying on information provided to their customers and the press then thats your opinion. Anyway I will just let others decide for themselves

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

GF 7LP is ready for customer designs and AMD has confirmed 7nm tapeouts later this year.

i think using samsungs 14LPP node was probably the way to go on their end. gave them more time to work on 7nm instead of trying to shoe horn designing two processes at the same time. we'll see how it turns out but seems quite promising on paper at least.
 
I swear, your process overhyping will end just like it did with Polaris, with you having a meltdown over disappointment of AMD not meeting your expectations. There's trying to hype up a turd and there's trying to overhype a sincerely decent product. Guess which one is worse.

He have been wrong every single type about his hyping. Not just process, also products and sales.
 
In CB15 MT 2P.

EPYC 7601 32C - 6879, but its questionable since the result first was between 5600 and 6000.
Xeon 8168 24C - 7212.
Xeon 8180 28C - 8301.
 
In CB15 MT 2P.

EPYC 7601 32C - 6879, but its questionable since the result first was between 5600 and 6000.
Xeon 8168 24C - 7212.
Xeon 8180 28C - 8301.
Rumor. seems most around the web are quite skeptical and not really putting a lot of weight in these numbers yet.
 
The most impressive for me is 1 EPYC delivers 128 PCIe lanes and 8 Channel DDR4. That's just nuts for a 1S setup where enterprise have to pay fees based on socket counts. Intel neuters it's 1S forcing companies to pay more for 2S and on top of that, more software costs while their workloads do not require that 2nd Xeon.

1S EPYC with 6x GPU accelerators & NVMEs full bandwidth is no joke and allows for a nice density jump that's even beyond Intel's Skylake-X (6 channel DDR4 and 44 PCIe).
 
In CB15 MT 2P.

EPYC 7601 32C - 6879, but its questionable since the result first was between 5600 and 6000.
Xeon 8168 24C - 7212.
Xeon 8180 28C - 8301.

If you analyze the numbers, it's obvious Cinebench is not optimized for 2 socket configuration.

Compare the results for single Intel Xeons and then 2 socket config, the scaling is pretty awful. And that's on released part that's mature.

If you want to be negative, this is the more interesting quote from Heise.de

"At the Linpack, Epyc expected to be less than the competitor, with 1050 to 1520 GFlops, it comes to about two thirds."

Looks like 1050 -> 1520, they are getting 50% scaling on that 2nd socket. Some issues??
 
Looks like 1050 -> 1520, they are getting 50% scaling on that 2nd socket. Some issues??

Considering you normally get ~95%? Hell, yeh. And you can assume EPYC isn't bandwidth starved either.
 
AMD brings competition and Kyle feels good about AMD for the first time in a long time. That is a plus in my book. :) Better than the two who seriously know nothing about what they are saying though, just AMD is the bad for them. :D
 
"At the Linpack, Epyc expected to be less than the competitor, with 1050 to 1520 GFlops, it comes to about two thirds."

Looks like 1050 -> 1520, they are getting 50% scaling on that 2nd socket. Some issues??
It does not talk about socket scaling, it talks about FLOPS comparison between 2P systems, please.

Same goes to you, Shintai.
 
AMD brings competition and Kyle feels good about AMD for the first time in a long time. That is a plus in my book. :) Better than the two who seriously know nothing about what they are saying though, just AMD is the bad for them. :D

tech sites on release day is the mothership of entertainment, you get two sides jousting for the high ground trying to vindicate their corporate of choice over the other while getting to play steven hawkins for a day, in the confines of those four walls. The intellectual testosterone is off the charts.
 
AMD brings competition and Kyle feels good about AMD for the first time in a long time. That is a plus in my book. :) Better than the two who seriously know nothing about what they are saying though, just AMD is the bad for them. :D
Very well said. These characters are found doing the same crap in any AMD thread both here and at anandtech forums.
 
Very well said. These characters are found doing the same crap in any AMD thread both here and at anandtech forums.
Quick, lets draw discussion away from benefits of AMD and focus on something irrelevant/highly questionable synthetic benchmarks
Their tactics are getting very drawn out, stale and boring as fuck.
Even seen one of the share shills here on WCCF recently shitting up an AMD article comment section. Nice one, Juan, hope those sheckels are worth it.


If there are no frills boards which can OC and those CPU prices are legit, this is a very enticing platform over Threadripper and a bit of a mecca of HEDT sanity for us crazy [H] motherfuckers, when the other camp wants to give you 20-40 lanes (lol).
I'd just settle for a 16C for now and spring for 7nm next year. Onboard 10Gb and heaps of Sata channels is worth it already..
I am a little worried that these are speed binned though, with TR getting the higher binned cores and Epyc getting the lower clocked but good cores.... typically in past I found exact opposite with Opteron. They had better clocking potential and I guess better sillicon quality/validation. That said, there were much smaller clock differences between desktop and server parts.
 
Last edited:
Quick, lets draw discussion away from benefits of AMD and focus on something irrelevant/highly questionable synthetic benchmarks
Their tactics are getting very drawn out, stale and boring as fuck.
Even seen one of the share shills here on WCCF recently shitting up an AMD article comment section. Nice one, Juan, hope those sheckels are worth it.


If there are no frills boards which can OC and those CPU prices are legit, this is a very enticing platform over Threadripper and a bit of a mecca of HEDT sanity for us crazy [H] motherfuckers, when the other camp wants to give you 20-40 lanes (lol).
I'd just settle for a 16C for now and spring for 7nm next year.
I am a little worried that these are speed binned though, with TR getting the higher binned cores and Epyc getting the lower clocked but good cores.... typically in past I found exact opposite with Opteron. They had better clocking potential and I guess better sillicon quality/validation. That said, there were much smaller clock differences between desktop and server parts.

How anyone can get so bent out of shape about trivial differences is laughable, it is great that now I can go pick up an intel or AMD CPU and get a good experience. But I expect the boy band brigade to tell me otherwise. The irony is most have never actually used a Ryzen or relevant intel part to get a real feeling, it is basically the pie chart heroes.
 
Back
Top