Intel CEO Bracing For EPYC Impact, Aims to Keep AMD Under 20% of Server Market Share

Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

ksec

Senior member
Mar 5, 2010
420
117
116
Wouldn't business for H2 2018 include contracts for 2019 (thus potentially include Zen 2/Epyc 2) then? Everybody already knew/expected that AMD will have a slow ramp up in everything OEM/ODM, especially in the server market.

Well no, Because they didn't have the time to test out Zen 2 yet. I think H2 2018 is still a little early.

Yep! We could not purchase AMD gear, because the architects said no. It wasn’t due to Intel malice, but general stupidity among people. You have to remember, most people shouldn’t be in the position they’re in.

To be honest. If I was in their position I wouldn't have said yes as well. Server are mission critical, changing something isn't as simple as flip a switch. Ryzen on Business Desktop is entirely different matter.

The reason why it is a lot easier to sell if Apple, IBM, Microsoft Azure, Google, or other Fortune 500 are using it internally is not simply just because no one takes the blame when larger enterprise are using it. It is because you trust these much larger animals to have done lots of testing and they see it fits there needs. You can trust these company to have done their parts so your inadequate testing are at least backed by their data.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
well yeah, that's a fair point. I don't see them hitting that in one year. I'm thinking BK was talking more long term than short term, unless he's suggesting more about the internal problems at Intel.

I think 20% server share by end of 2019 is perfectly achievable for AMD as they ramp 7nm Rome by Q2 2019.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,620
136
Well no, Because they didn't have the time to test out Zen 2 yet. I think H2 2018 is still a little early.
H2 2018 didn't even start yet. And plenty contracts for stuff in early 2019 are bound to be done in late 2018, thus H2 2018.
 

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
I think You misunderstood, i meant 2 or 4 dies on MCM, connected via secret sauce.



AMD inter die latency is ~same as Intel inter CPU latency in dual multi socket , so as long as Intel is willing to drop monolithic, they can simply take a 2 dies and connect them with QPIs (maybe rename them with more flashy name ). I am not versed in "EMIB" advantages, but cores being within centimeters instead of decimeters in socket) and communicating in "substrate" would allow for more BW and/or latency redution.

I'd say main challenge is power density and delivery to socket. Inte's AVX512 mode has ridiculous power usage per core and without 400w or so per socket AVX2/AVX512 clocks would drop to the bottom.


Actually AMD inter die latency is better than intel inter socket. So if your VM only use 4 core, AMD actually have better performance/ipc
 

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
Well no, Because they didn't have the time to test out Zen 2 yet. I think H2 2018 is still a little early.



To be honest. If I was in their position I wouldn't have said yes as well. Server are mission critical, changing something isn't as simple as flip a switch. Ryzen on Business Desktop is entirely different matter.

The reason why it is a lot easier to sell if Apple, IBM, Microsoft Azure, Google, or other Fortune 500 are using it internally is not simply just because no one takes the blame when larger enterprise are using it. It is because you trust these much larger animals to have done lots of testing and they see it fits there needs. You can trust these company to have done their parts so your inadequate testing are at least backed by their data.

To be honest, I won't trust intel unless their get io bottleneck shorten out, it's killing our database and VM performance. Especially when we need HPET to always on to prevent time drifting but we're transitioning to ryzen now.
 

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
Actually AMD inter die latency is better than intel inter socket. So if your VM only use 4 core, AMD actually have better performance/ipc

Check it out to educate Yourself:

https://www.servethehome.com/amd-epyc-infinity-fabric-latency-ddr4-2400-v-2666-a-snapshot/

To be honest, I won't trust intel unless their get io bottleneck shorten out, it's killing our database and VM performance. Especially when we need HPET to always on to prevent time drifting but we're transitioning to ryzen now.

Running HPET in 2018? Kinda interested to hear about it as we are running Xeon Golds and never had any problems with clock drift.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
Check it out to educate Yourself:

https://www.servethehome.com/amd-epyc-infinity-fabric-latency-ddr4-2400-v-2666-a-snapshot/



Running HPET in 2018? Kinda interested to hear about it as we are running Xeon Golds and never had any problems with clock drift.
They said that the difference is average 8%. If also said the Intel was only about 8% different (higher) If it could run 3200, which I think it should do easily (the TR chips do 3600, I have some) Then it would be faster than Intels QPI. But your point is well taken that Intel still holds that one lead.
 
  • Like
Reactions: Drazick

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
They said that the difference is average 8%. If also said the Intel was only about 8% different (higher) If it could run 3200, which I think it should do easily (the TR chips do 3600, I have some) Then it would be faster than Intels QPI. But your point is well taken that Intel still holds that one lead.

Not sure i can connect TR with EPYC and with this conclusion from article linked:

Since comparisons are all the rage these days, and we boldly made the statement that AMD’s on package latency looked much like Intel’s inter-socket latency (in a quad socket configuration.) To an OS, both an Intel Xeon Scalable quad socket system and a single socket AMD platform will look similar with four NUMA nodes.

And what we are trying to prove anyway? Intel has 2 monolithic chips connected with up to 3 links of QPI. So going from socket to socket is pretty much uniform latency. And on AMD? It's complicated.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
The only difference between TR and EPYC is that EPYC officially supports ECC, and TR has to have a motherboard that supports it, AND it only has 4 channel ram instead of 8. The base workings of them (and the IF) is exactly the same, and they suffer/benefit from the same things. The memory speed increase makes the CPU almost as fast in a linear way.

SO if EPYC could run 3200 or even 2933, by the articles own logic, it would equal or surpass the Xeons QPI speed.
 

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
SO if EPYC could run 3200 or even 2933, by the articles own logic, it would equal or surpass the Xeons QPI speed.

Can't really "surpass" anything when you need to go extra hop in the socket in worst case. EPYC can't overclock memory, so it is stuck with 2666 and latencies are what they are.
 

Genx87

Lifer
Apr 8, 2002
41,095
513
126
People still buy servers and fill out data centers in 2018? :)

Dropped that 18 months ago. Only servers we buy are to run VMs for domain controllers, print server, and maybe a local DFS site. Rest is going up in Azure. Cheaper and easier to deploy\scale.

I think this is where Intel is worried. Especially if they have not fixed the spectre stuff and their chips are under performing. Corporations buying up cloud computing to find out their VMs or App service is now not getting it done. Either reduce cost of service, or increase back end to provide same level of service. Either way it is a cost for somebody in the chain.
 
Last edited:

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
Check it out to educate Yourself:

https://www.servethehome.com/amd-epyc-infinity-fabric-latency-ddr4-2400-v-2666-a-snapshot/



Running HPET in 2018? Kinda interested to hear about it as we are running Xeon Golds and never had any problems with clock drift.

It's comparing inter-socket latency, and you can always use TR platform since it's also support ECC memory, and you can also overclock ECC memory in TR, and it will significantly reduce latency.

If you use over provisioning core in vm, some VM will time drifting, especially if you use Windows server 2008 vanilla or r2(we still have legacy 32 bit app).
 

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
People still buy servers and fill out data centers in 2018? :)

Dropped that 18 months ago. Only servers we buy are to run VMs for domain controllers, print server, and maybe a local DFS site. Rest is going up in Azure. Cheaper and easier to deploy\scale.

I think this is where Intel is worried. Especially if they have not fixed the spectre stuff and their chips are under performing. Corporations buying up cloud computing to find out their VMs or App service is now not getting it done. Either reduce cost of service, or increase back end to provide same level of service. Either way it is a cost for somebody in the chain.

Because our application is latency sensitive, so cloud computing is not an option, heck we need our own dedicated line to connect between different branches.
 

ksec

Senior member
Mar 5, 2010
420
117
116
H2 2018 didn't even start yet. And plenty contracts for stuff in early 2019 are bound to be done in late 2018, thus H2 2018.

Well Zen 2 EPYC will only start to sample at late H2 2018 and shipping in 2019. To have those tested and signed contract in 2018 sounds a little premature. I hope I am wrong though.
 

BigDaveX

Senior member
Jun 12, 2014
440
216
116
I think 20% server share by end of 2019 is perfectly achievable for AMD as they ramp 7nm Rome by Q2 2019.

Maybe that's possible, but the end of 2020 would probably be a more achievable goal. Server buyers are notoriously slow to move to new platforms, Intel or AMD; if you look at Dell's website for instance, a lot of their current server line-up is still based on Broadwell-E.
 

zinfamous

No Lifer
Jul 12, 2006
110,515
29,100
146
Well Zen 2 EPYC will only start to sample at late H2 2018 and shipping in 2019. To have those tested and signed contract in 2018 sounds a little premature. I hope I am wrong though.

I don't think Zen 2, Epyc or Ryzen, is out until 2Q or 2H 2019. Ryzen 2/Epyc 2 = "Zen +"

...it's confusing.
 

dacostafilipe

Senior member
Oct 10, 2013
771
244
116
Well Zen 2 EPYC will only start to sample at late H2 2018 and shipping in 2019. To have those tested and signed contract in 2018 sounds a little premature. I hope I am wrong though.

Sampling is part of the testing procedure and that testing also includes external sources. Mostly it's platform vendors, but it's possible that big clients have access to that hardware long before the launch.

CERN had a lot of information about Zen long before the launch.
 

ksec

Senior member
Mar 5, 2010
420
117
116
Sampling is part of the testing procedure and that testing also includes external sources. Mostly it's platform vendors, but it's possible that big clients have access to that hardware long before the launch.

CERN had a lot of information about Zen long before the launch.

The is assuming they can sample, and test and close the deal all at the same time, likely all in Q4. I still doubt that will provide revenue in H2 2018. Again I would love to be wrong.


I don't think Zen 2, Epyc or Ryzen, is out until 2Q or 2H 2019. Ryzen 2/Epyc 2 = "Zen +"

...it's confusing.

Well it is not that confusing ( Compared to Intel )

There won't be Zen+ for EPYC. EPYC 2 will be using Zen 2.
Ryzen 3 or whatever it is called will be using Zen 2 as well.
 
  • Like
Reactions: dacostafilipe

Genx87

Lifer
Apr 8, 2002
41,095
513
126
Because our application is latency sensitive, so cloud computing is not an option, heck we need our own dedicated line to connect between different branches.

You can have dedicated links into the cloud. Probably have better latency into Azure or AWS than between your offices ;)
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
The is assuming they can sample, and test and close the deal all at the same time, likely all in Q4. I still doubt that will provide revenue in H2 2018. Again I would love to be wrong.

The process will take longer than that; at least with platform vendors and that will gate the whole process.
 

ultimatebob

Lifer
Jul 1, 2001
25,135
2,445
126
Has your Senior Management being invited to any Intel all expenses paid "conferences" in recent months.......

Some of us have been around long enough to remember AMD's Opteron product line, and how unstable it was compared to their Intel brethren. It was particularly flaky with VMWare ESX early on, which would explain why no CTO in their right mind would use a first generation EPYC processor as a production hypervisor.

Personally, I'd buy one to run dev/test workloads and see how it performs there. If it works out there for a year, THEN we can buy some for production.