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

[itsmydamnation]

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.

I was around then and ESX 3 was flaky as on intel as well........ We ran just as much opteron as we did Xeon during those days, dont remember any difference in platform stability. The fact they sold everything they could make while being more expensive then Xeon's (atleast in australia) shows that no one else really had these problems.........

I dont know why people are pretending this is all new to AMD, IF is the continuation of hyper transport, The topology of a 2 socket EPYC is the topology of a 4 socket mangny-Cours/BD (https://pdos.csail.mit.edu/6.828/2016/lec/AMD.pdf).

Intel has a complete new uncore for scailable family, do you recommend the same approach?

 

[Markfw]

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.

I personally owned Opterons of that era, and I had no problems at all. I also was in a high level IT position, which I have since retired from. I am not sure where you heard your information.

Oh, and you may not remember that in that era also, Anantech was run 100% by Opterons, and Anand himself swore by them.

 

[scannall]

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.

That's not an unreasonable thing to do actually. It's a new architecture, and learning it quirks, where it works well, or not at all is just smart.

 

[scannall]

I was around then and ESX 3 was flaky as on intel as well........ We ran just as much opteron as we did Xeon during those days, dont remember any difference in platform stability. The fact they sold everything they could make while being more expensive then Xeon's (atleast in australia) shows that no one else really had these problems.........

I dont know why people are pretending this is all new to AMD, IF is the continuation of hyper transport, The topology of a 2 socket EPYC is the topology of a 4 socket mangny-Cours/BD (https://pdos.csail.mit.edu/6.828/2016/lec/AMD.pdf).

Intel has a complete new uncore for scailable family, do you recommend the same approach?

It would likely be smart to do that on the new Intel's as well. Particularly in light of Meltdown.

 

[ultimatebob]

I personally owned Opterons of that era, and I had no problems at all. I also was in a high level IT position, which I have since retired from. I am not sure where you heard your information.

For me, it was first hand experience. I had to reboot a fair number of flaky HP ProLiant servers with Opteron's in them back when I helped run a data center around 2008. I can't really blame HP, either... we had a bunch of HP ProLiant DL360's with Xeon processors as well, and the only thing that tended go bad on those were the hard drives.

 

[jpiniero]

I don't see anything Intel really has to counter Epyc for the next 2 or 3 years.

Despite all the delays Intel has had with 10nm, there's always going to be the threat that Intel will actually fix it and suddenly get a lot faster as they basically skip 2-3 generations of products. But yeah, unless Intel craps out a miracle with Cooper Lake AMD is going to get at least a year.

 

[wahdangun]

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

But still, I'm not living in United States.

 

[ehume]

vulnerabilities

Him even talking like that may mean they won't even try the illegal route like they did in the Opteron/Athlon 64/X2 era.

Won't be near as easy as it used to be to do the stuff they did before. Especially like what Mr. Anand was told from motherboard makers.

I guess we'll see.

Q's:

1. What did MB mfr's tell Anand?

2. What if when Intel fixes theirSpectre/Meltdown vulnerabilities, they take a performance hit that puts them at parity with Ryzen?

 

[formulav8]

1. What did MB mfr's tell Anand?

2. What if when Intel fixes theirSpectre/Meltdown vulnerabilities, they take a performance hit that puts them at parity with Ryzen?

Here is a post for some of Anand's experiences. He's mentioned before in other articles but don't remember which ones.

Here is a quote from the article.

I've known about this sort of stuff for quite some time, in fact, I'd say that out of the 48 pages AMD's legal team put together there's a lot missing. AMD told me that they aren't putting all cards on the table, but here are a couple of other things that I've seen personally:

I can't even begin to count the number of times where motherboard manufacturers have told me that they could not:

1) Send an AMD motherboard for review
2) Promote an AMD motherboard
3) Let us take pictures of an AMD motherboard

Out of fear of Intel retaliation. Remember the original Athlon days when no motherboard manufacturer would dare make a board for the K7? All of the frightened manufacturers were afraid of them losing their Intel chipset allocation if they supported the K7.

The same sort of stuff happened during the i820 days. Intel's first RDRAM based chipset was a complete flop, yet they offered no real SDRAM alternative. VIA did however, and Intel punished those manufacturers who didn't promote their i820 platforms or who too eagerly embraced VIA's solutions.

The list goes on and on.

i'm sure some here could probably remember when Asus didn't even put their name on their Athlon boards. Not only was the mobo box plain (white IIRC) and lacking their brand, the motherboard itself didn't have Asus written anywhere apparently do to fear of Intel retaliation.

For your 2nd, not sure what the total performance impact will be when all of the fixes have been applied. Meltdown measurably hurt Intel in some IO areas. And with all of the Spectre/Meltdown/NG/etc applied, they will lose CPU performance. Even if each fix only impacts performance to a small degree, it will likely reduce their performance to some measurable level when all are added up. But even so, do to their strong Mhz advantage, they should retain the overall per core performance for multiple models.

 

[SickBeast]

Intel is starting to look desperate. Their 28 core CPU demo was a complete disaster. I'm surprised they didn't use liquid nitrogen.

 

[beginner99]

2. What if when Intel fixes theirSpectre/Meltdown vulnerabilities, they take a performance hit that puts them at parity with Ryzen?

i mean fixes for meltdown and some spectre variants are here and available and it can cripple performance. The main problem being IO. And remember networking = IO. So if you are running something with many, many small requests, your performance will tank. If you most do calculations with little IO, you will notice barely nothing.

I think there was benchmark made public from a multiplayer gaming server and basically cpu usage doubled after patches but this is also pretty much the worst case scenario (many tiny requests).

 

[CatMerc]

Funny thing about server adoption, the first gen will be the slowest one.
EPYC 2 will not take as long to certify (shared platform and iteration of existing tech), and from what I'm hearing from a local cloud provider, AMD has almost unbelievable claims for EPYC 2. They're definitely creaming their pants.
No I don't know what the claims are, just that my friend is excited, don't ask me.

20% by 2020 doesn't sound unrealistic considering how explosive in sales this market tends to be after verification is done.

 

[wahdangun]

Funny thing about server adoption, the first gen will be the slowest one.
EPYC 2 will not take as long to certify (shared platform and iteration of existing tech), and from what I'm hearing from a local cloud provider, AMD has almost unbelievable claims for EPYC 2. They're definitely creaming their pants.
No I don't know what the claims are, just that my friend is excited, don't ask me.

20% by 2020 doesn't sound unrealistic considering how explosive in sales this market tends to be after verification is done.

Man you make me regretting to jump ship this early to zen ecosystem.

 

[CatMerc]

Man you make me regretting to jump ship this early to zen ecosystem.

As I said, I don't know why its making them cream their pants. Could be something completely unrelated to consumer workloads. Don't fret yet

 

[csbin]

Intel -1.3% on analyst downgrade with 18% downside

https://seekingalpha.com/news/3364492-intel-minus-1_3-percent-analyst-downgrade-18-percent-downside

Northland Capital downgrades Intel (NASDAQ:INTC) from Market Perform to Underperform.

• Price target: $45, an 18% downside to Friday’s close.

• Intel shares are down 1.3% premarket to $54.40.

 

[richaron]

i mean fixes for meltdown and some spectre variants are here and available and it can cripple performance. The main problem being IO. And remember networking = IO. So if you are running something with many, many small requests, your performance will tank. If you most do calculations with little IO, you will notice barely nothing.

I think there was benchmark made public from a multiplayer gaming server and basically cpu usage doubled after patches but this is also pretty much the worst case scenario (many tiny requests).

You mean this Fortnite one from here?

^I guess 1, 2, and 3 are different machines with some sort of load balancing. Shows what their patched system needed to match the unpatched ones. Though keep in mind this is from very early in the patching process and performance could have gone up or down from here (I think this is just for Meltdown).

 

[wahdangun]

As I said, I don't know why its making them cream their pants. Could be something completely unrelated to consumer workloads. Don't fret yet

But, the thing is if I'm care about consumer workload, I won't considering jumping ship (besides horrible IO penalties).

 

[Topweasel]

For me, it was first hand experience. I had to reboot a fair number of flaky HP ProLiant servers with Opteron's in them back when I helped run a data center around 2008. I can't really blame HP, either... we had a bunch of HP ProLiant DL360's with Xeon processors as well, and the only thing that tended go bad on those were the hard drives.

We had some Dell Opteron's from those days. We still have one running as a esxi host for a our low priority training network. Outside a raid controller going down with one of the systems they were bullet proof.

 

[ultimatebob]

What I want to know is when we'll be able to get cloud hosted Epyc servers. Most people don't buy their own infrastructure anymore, they rent it from the likes of Amazon or Microsoft Azure.

 

[Topweasel]

What I want to know is when we'll be able to get cloud hosted Epyc servers. Most people don't buy their own infrastructure anymore, they rent it from the likes of Amazon or Microsoft Azure.

Is that really the case? I mean I can see moving most outward facing services to the cloud, but all?

 

[jpiniero]

Is that really the case? I mean I can see moving most outward facing services to the cloud, but all?

It is trending that way.

 

[IEC]

What I want to know is when we'll be able to get cloud hosted Epyc servers. Most people don't buy their own infrastructure anymore, they rent it from the likes of Amazon or Microsoft Azure.

You can rent a cloud instance using Epyc CPUs from Tencent (one of the largest cloud/Internet presences in China) already:
https://cloud.tencent.com/act/event/AMD.html

 

[ultimatebob]

You can rent a cloud instance using Epyc CPUs from Tencent (one of the largest cloud/Internet presences in China) already:
https://cloud.tencent.com/act/event/AMD.html

That's cool, but I was really hoping that one of the big three (Google, Amazon, Microsoft) would offer it here in the states. Cheap compute horsepower is a good thing to have!

 

[Mopetar]

Funny thing about server adoption, the first gen will be the slowest one.

I'm not even sure why Epyc didn't take off more than it did. As soon as the performance numbers came out, we jumped on getting a new server specced out and ordered. Ended up with a 64C/128T box that cost just a little bit more than what Intel was asking for their highest-end 28C/54T Xeon alone.

I can get that big data centers aren't going to take an immediate plunge on a new product, but I figured that there were plenty of places that occasionally need a new server and the value proposition on Epyc was so much better.

The relative lack of spectre/meltdown issues was just extra gravy.

 

[maddie]

You can rent a cloud instance using Epyc CPUs from Tencent (one of the largest cloud/Internet presences in China) already:
https://cloud.tencent.com/act/event/AMD.html

This is why I said China is the key in an earlier post, especially with the tech license agreement. Each time is different and comparisons to the last time we had this competitive situation is wrong. Intel has little/less influence over Chinese firms and if AMD is truly superior in servers, then other competitors will have to follow or be left behind and lose customers.