News [Anand] Intel's Enterprise Extravaganza, Cascade Lake Launches

[topmysteries5]

So you handicap the 2950x, just so the Xeon can get close ? And 16 cores vs 22 ? I have 4 Xeons, but they are 14 core@2.5 ghz. My 2700x beats them with only 8 cores@stock ! A Ryzen 3xxx series or Rome will wipe the floor with them.

lol,
Chill man, just have fun .
comparison is 16c vs 18c turbo unlocked. I was talking about IPC, not overall performance. Performance at same clock speed on last year cpu vs 3 to 4 yrs old cpu. I know 2950x wins, but there is not big difference like 40-50%, its (xeon) around 5-10% slower at same 3.4ghz clock speed




Calling other members fanboys, is not allowed.


esquared
Anandtech Forum Director

 

[Markfw]

I didn't say it would be competitive in power usage or heat though, I was talking raw performance in DPFP. Which is the market aimed with the AP line.

If money is no object, you go with a supersomputer, who cares about the electric bill. If you run a data center, power and heat are your enemy, and you can live or die by how much capacity you have and your power bill, that includes power consumption and AC.

 

[JasonLD]

If money is no object, you go with a supersomputer, who cares about the electric bill. If you run a data center, power and heat are your enemy, and you can live or die by how much capacity you have and your power bill, that includes power consumption and AC.

That probaly depends on what kind of workload you are dealing with. If you happen to utilize avx-512 heavily, then you have a good reason to get CL-AP.

 

[Mopetar]

You don't need a dual EPYC to beat this thing. Just one 64 core would beat it. The 3,8 is only one core.

I was wondering about that. The AT article doesn't specify that whether the listed turbo is one- or all-core turbo. It seemed entirely too good to be true for any kind of sustained all-core load.

 

[JasonLD]

I was wondering about that. The AT article doesn't specify that whether the listed turbo is one- or all-core turbo. It seemed entirely too good to be true for any kind of sustained all-core load.

Should be one core turbo. Since everything else is listed that way it seems. All core could be between 2.8-3.2.

 

[jpiniero]

If money is no object, you go with a supersomputer, who cares about the electric bill.

That's sort of the point I was making - the aim of the AP line is at supercomputers. Efficiency wise, using GPUs is definitely the way to go but customers who would consider the AP line aren't considering GPUs.

 

[jpiniero]

I would say that the only really impressive thing is the Optane DIMMS... given the right workload it would make a gigantic difference and there really isn't anything on the market that can compare. I'm sure they will charge waaay too much for them however.

 

[LTC8K6]

It's 2 8200 chips right? So we know what the turbo speeds are, don't we?

8176 does 2.8 all core turbo, so probably 2.7?

Okay, 8176 is 2.8/2.4/1.9 all core normal/avx/avx512

Guess 2.7/2.3/1.8?

 

[Markfw]

It's 2 8200 chips right? So we know what the turbo speeds are, don't we?

8176 does 2.8 all core turbo, so probably 2.7?

Okay, 8176 is 2.8/2.4/1.9 all core normal/avx/avx512

Guess 2.7/2.3/1.8?

Well, mind you this is a guess, but doubling the cores ? has to be more than a 100 mhz hit.

 

[LTC8K6]

Well, mind you this is a guess, but doubling the cores ? has to be more than a 100 mhz hit.

True, but it seems to come with massive cooling, and it's BGA only.

 

[IntelUser2000]

True, but it seems to come with massive cooling, and it's BGA only.

The -AP is basically a successor to Xeon Phis so it makes sense.

 

[beginner99]

I didn't say it would be competitive in power usage or heat though, I was talking raw performance in DPFP. Which is the market aimed with the AP line.

That's sort of the point I was making - the aim of the AP line is at supercomputers. Efficiency wise, using GPUs is definitely the way to go but customers who would consider the AP line aren't considering GPUs.

Quoted your first post then read your second one which was what I wanted to reply initially. Who buys Xeons for DPFP throughput? You would almost certainly be much more efficent doing that on Radeon instincts and besdies that how big is the market for these CPUs? Most servers are some lame db server or for virtualized app servers. Nothign special.

 

[jpiniero]

Quoted your first post then read your second one which was what I wanted to reply initially. Who buys Xeons for DPFP throughput? You would almost certainly be much more efficent doing that on Radeon instincts and besdies that how big is the market for these CPUs? Most servers are some lame db server or for virtualized app servers. Nothign special.

Some groups seem to prefer CPU only development. If you look at the Top 10 in the Top500 list, 5 are using nVidia but the other 5 are CPU based (3 Intel, 1 PowerPC and one Chinese Custom)

 

[DrMrLordX]

The -AP is basically a successor to Xeon Phis so it makes sense.

Glad someone made this point.

Nobody's going to be looking at a Platinum 9200 unless they were already in the market for Phi, which is/was an AVX-512-based computing monstrosity. It's direct competitors are/were datacenter dGPUs. Not EPYC/EPYC 2. I expect existing Phi customers to want these things since most of their code should port over nicely. As an added bonus, there's no question as to whether or not the 9200 can boot on its own. All the oogy-ness of dealing with past Phi products should be essentially eliminated. 400W for a datacenter compute . . . thing . . . isn't necessarily awful, if that's really what you want. That being said, the 9200 is kind of a stopgap until Intel can roll out their own datacenter dGPUs.

 

[jpiniero]

Well, if you think about it, the 9282 is basically two 8280's in one CPU package, except you lose 8S support and I believe the AP only has 40 lanes total. I don't think there's even any memory bandwidth advantage since it's using UPI to connect between the two dies (and not say EMIB). The only real advantage is the space savings it looks.

 

[Markfw]

Well, if you think about it, the 9282 is basically two 8280's in one CPU package, except you lose 8S support and I believe the AP only has 40 lanes total. I don't think there's even any memory bandwidth advantage since it's using UPI to connect between the two dies (and not say EMIB). The only real advantage is the space savings it looks.

To me, its just a desperate attempt to counteract the 64 core Rome EPYC, and a bad attempt at that.

 

[Kippa]

As far as I am aware these are Xeon based chips, what about HEDT by intel in 2019? Last year they just did a refresh with the same core numbers but just with a few extra 100 mhz. What is the intel HEDT likely to be later on this year as I haven't heard much about them.

 

[IntelUser2000]

To me, its just a desperate attempt to counteract the 64 core Rome EPYC, and a bad attempt at that.

The counter is Cooper Lake, not the -AP line.

 

[Markfw]

The counter is Cooper Lake, not the -AP line.

And when is that going to be seen ?

Sorry, but ever since Intel has blown the 10 nm thing, and their CEO said "well we are trying not to lose more then 20% share to AMD in the enterprise market", I have not seen one thing that impressed me, or made me think they can even compete in the next 2 years, enterprise or consumer. Both the X3175 launch, and this appear to be a joke to me, not real competition at all.

 

[jpiniero]

To me, its just a desperate attempt to counteract the 64 core Rome EPYC, and a bad attempt at that.

The AP line has been rumored for several years now, once the original Aurora project got canned because of the Phi's ineffectiveness. It may be pressed into duty to compete with Rome, but it pretty clearly was designed to be a Phi replacement. Anything designed to compete with Rome would have been 10 nm based.

The counter is Cooper Lake, not the -AP line.

Cooper looks like it'll be another 400+ W monster though. May not be too much different from the AP line, except using EMIB to connect the two dies.

As far as I am aware these are Xeon based chips, what about HEDT by intel in 2019? Last year they just did a refresh with the same core numbers but just with a few extra 100 mhz. What is the intel HEDT likely to be later on this year as I haven't heard much about them.

Seems there will be Cascade Lake-X at some point.

Sorry, but ever since Intel has blown the 10 nm thing, and their CEO said "well we are trying not to lose more then 20% share to AMD in the enterprise market", I have not seen one thing that impressed me, or made me think they can even compete in the next 2 years.

The Optane DIMMs (and it's exclusivity) might save them somewhat in some markets. The $$$ will make it a niche however.

 

[IntelUser2000]

And when is that going to be seen ?

That's true, but the -AP line has its markets. It's a very dense server platform. It won't go away with GPUs.

With Cooper Lake, -SP line will get MCM.

 

[Markfw]

That's true, but the -AP line has its markets. It's a very dense server platform. It won't go away with GPUs.

With Cooper Lake, -SP line will get MCM.

And EPYC with 64 cores is not as dense ? And the EPYC will fit in a 4p motherboard (do they make those for EPYC ?) At least in 2P.

 

[IntelUser2000]

Servethehome explains it quite well. It allows for compact, dense HPC deployments.

https://www.servethehome.com/intel-xeon-platinum-9200-formerly-cascade-lake-ap-launched/

They say it isn't supported by most OEMs because it doesn't fit their market, and will also end up being lower volume. For them, it'll be the -SP line. That's why Cooper Lake is coming.

 

[SarahKerrigan]

And EPYC with 64 cores is not as dense ? And the EPYC will fit in a 4p motherboard (do they make those for EPYC ?) At least in 2P.

Epyc with 64 cores doesn't presently exist. It will, of course - but today the competition is 32 cores. Epyc does not presently support 4-socket or 8-socket configurations; there have been rumors that Rome will, but if AMD has confirmed them, I haven't seen it.

 

[Zucker2k]

Has AMD even implemented AVX-512 in Rome? And these Intel chips since Skylake-X are 2x512bit capable. Or?