Intel announces Cascade Lake AP

[Taf the Ghost]

1 hour ago, leadeater said:

And will the OS see it as 2 CPUs or 1?

 

NUMA NUMA time 

 

Did someone say NUMA?

 

 

[Taf the Ghost]

1 hour ago, 0x00000001 said:

Linpack : up to 3.4X vs single EPYC 7601

StreamTriad : up to 1.3x vs single EPYC 7601

..... up to ... vs single EPYC 7601

Price : up to 9999x vs single EPYC 7601

It has 50% more cores per package. 30% better. Haha

 

On topic, part of me wants to forehead slap over this. They aren't even using the 28 core parts. No, it's the 24 core parts they had trouble selling in current Xeon stack. Though it's actually a smart business decision. The 24 core parts go for a lot less, so I full expect these to go for over 16k USD each. Gotta pump that margin up.

 

And it all but confirms we're getting 64 core Epyc parts. Intel wouldn't launch this 2 days before the AMD IR meeting if there wasn't something important they wanted to get ahead of.

[Amazonsucks]

3 hours ago, 19_blackie_73 said:

are we back to the early days of powerpoint design? looks 80s

Presentations for HPC, mainframe and datacenter stuff is usually black text on a white background with the occasional picture or diagram. Since thats mostly what im used to seeing, this intel thing didnt strike me as anything out of yhe ordinary until you mentioned it

[Sauron]

About time! It's painfully obvious that modular designs are the future, this is a step in the right direction.

[Amazonsucks]

3 hours ago, leadeater said:

Yea but Intel's glue is a much better special glue, it's Blue so better.

 

Holy crap that's a bit over kill on the ram channels, but I guess that's the side effect of gluing 2 dies together that have 6 channels each.

 

It'll probably be UPI over EMIB, the dies themselves have to use UPI to talk across to other dies/packages as that's the dedicated interface for it and Intel hasn't done a complete re-architecture to change that.

http://www.cpushack.com/2015/02/22/nec-sx-ace-quad-core-vector-supercomputing/

 

Back in 2013 that 4 core CPU had 16 channels of DDR3 2000 with 256GB/s.

 

https://en.wikichip.org/wiki/ibm/microarchitectures/power9#Memory_Hierarchy

 

IBM POWER 9 has 32 channels currently.

[S w a t s o n]

INTEL IS SHOOK.

 

I thought cascade-AP was going to be 3 dies on a package not two but still. For me this confirms that Zen2 will come in a 64 core variant rather than the originally rumoured 48 core limit.

 

Spoiler

 

[The Benjamins]

Has no one ask how big this socket is going to be, I think it will be a estimated 6000-7000 pins, the thing will be huge.

 

WCCF says BGA 5903

https://wccftech.com/intel-cascade-lake-advanced-performance-48-core-xeon-cpus-announced/

[GoldenLag]

5 minutes ago, The Benjamins said:

Has no one ask how big this socket is going to be, I think it will be a estimated 6000-7000 pins, the thing will be huge.

Think of the VRM that will be on the board just for redundancy. 

 

Its gunna be big

[sazrocks]

Hold up;

Quote

CONFIGURATIONS
LINPACK: AMD EPYC 7601: Supermicro AS-2023US-TR4 with 2 AMD EPYC 7601 (2.2GHz, 32 core) processors, SMT OFF, Turbo ON

Uh, is that a problem?

 

Also the stream triad thing is really old for AMD:

Quote

Stream Triad: 1-node, 2-socket AMD EPYC 7601, http://www.amd.com/system/files/2017-06/AMD-EPYC-SoC-Delivers-Exceptional-Results.pdf tested by AMD as of June 2017

 

[Taf the Ghost]

2 hours ago, GoldenLag said:

Think of the VRM that will be on the board just for redundancy. 

 

Its gunna be big

It won't quite need double of a Xeon Platinum, but at least 60% more VRM capacity is going to be needed. 12 channel memory going to take up a lot of space (unless it's only 12 DIMM slots available?), along with a huge space for the socket.

 

While this will be the "single socket performance king" or something, this will require unique boards & designs (not refreshes from Purley), spacing will be interesting and cooling is a big issue. TCO on these aren't going to be great.

[sazrocks]

And reading the fine print on the Cascade-AP slide, it says that not all security updates were installed on the intel system (meltdown/specter?) and that some platform-independent optimizations were only given to Intel. 

 

Can Someone who knows more than me about this stuff say if this means anything?

[Brooksie359]

6 hours ago, leadeater said:

And will the OS see it as 2 CPUs or 1?

 

NUMA NUMA time 

 

I would hope anyone who is buying such a processor would be fine with running in numa or better yet in Linux. 

[The Benjamins]

2 minutes ago, VegetableStu said:

BGA? soldered!? o_o

Ya I find that odd, but it might be the only way to keep the size down to fit 2 of these on one board.

[Taf the Ghost]

1 hour ago, VegetableStu said:

BGA? soldered!? o_o

 

1 hour ago, The Benjamins said:

Ya I find that odd, but it might be the only way to keep the size down to fit 2 of these on one board.

Might also be for custom cooling solutions, though if it's BGA, it's probably only being offered by a few system integrators anyway. 

[leadeater]

3 hours ago, Brooksie359 said:

I would hope anyone who is buying such a processor would be fine with running in numa or better yet in Linux. 

Multi socket servers are already multiple NUMA zones anyway, Windows handles NUMA just fine as well so long as all NUMA zones actually have memory like they are supposed to. I really like TR and EPYC but the 2970WX/2990WX are just stupid designs.

[leadeater]

8 hours ago, Amazonsucks said:

http://www.cpushack.com/2015/02/22/nec-sx-ace-quad-core-vector-supercomputing/

 

Back in 2013 that 4 core CPU had 16 channels of DDR3 2000 with 256GB/s.

 

https://en.wikichip.org/wiki/ibm/microarchitectures/power9#Memory_Hierarchy

 

IBM POWER 9 has 32 channels currently.

Yes and those are much different, I was commenting about this Xeon having 12 channels as being over kill. The Power9 doesn't actually have massive memory bandwidth, not like the 32 suggests. There's 8 DMI interfaces at 28.8 GB/s for a total of 230GB/s which is about the same as 12 DDR4 channels but with higher latency, IBM went with the design they did for super high amounts of memory per CPU.

 

Edit:

Pretty much why I think it's over kill is because you're going to have a really hard time fitting in the number of memory slots required in a dual socket multi blade design, those are 1U half width and are already super tight on space.

[porina]

10 minutes ago, leadeater said:

Pretty much why I think it's over kill is because you're going to have a really hard time fitting in the number of memory slots required in a dual socket multi blade design, those are 1U half width and are already super tight on space.

Isn't part of the offering here is that you now get in 2 sockets what you had to use 4 sockets for previously? Even if the ram space is the same, you just reduced 2 sockets. IMO ram bandwidth hasn't been keeping up especially with HCC CPUs and this at least prevents it from getting worse.

[WereCat]

 

 

LOL

[cj09beira]

12 hours ago, GoldenLag said:

How many channels will Zen 2 have then with 8 prpbable core dies (+_+)

8 is the rumor but it will be 3200 mhz 

5 hours ago, Taf the Ghost said:

It won't quite need double of a Xeon Platinum, but at least 60% more VRM capacity is going to be needed. 12 channel memory going to take up a lot of space (unless it's only 12 DIMM slots available?), along with a huge space for the socket.

 

While this will be the "single socket performance king" or something, this will require unique boards & designs (not refreshes from Purley), spacing will be interesting and cooling is a big issue. TCO on these aren't going to be great.

for how long though, amd's 7nm is coming soon, and it might beat it (it better as it will have more cores and be on 7nm), if it doesn't they could always release a higher power version to beat it, 

also it seems zen 2 will get 2 times the avx performance but without avx 512 which is great news 

[Taf the Ghost]

1 minute ago, cj09beira said:

8 is the rumor but it will be 3200 mhz 

for how long though, amd's 7nm is coming soon, and it might beat it (it better as it will have more cores and be on 7nm), if it doesn't they could always release a higher power version to beat it, 

also it seems zen 2 will get 2 times the avx performance but without avx 512 which is great news 

It's an Intel "new product" to get announced just before AMD announces something. This AP line is probably just slightly above vaporware at the moment.

 

As for Zen 2, improving the AVX2 ability was a pretty low-hanging fruit, so we'll see. Be interesting if the speculation it's actually 4x 128bit units, rather than a dedicated 256bit unit plays out.

[Amazonsucks]

1 hour ago, leadeater said:

Yes and those are much different, I was commenting about this Xeon having 12 channels as being over kill. The Power9 doesn't actually have massive memory bandwidth, not like the 32 suggests. There's 8 DMI interfaces at 28.8 GB/s for a total of 230GB/s which is about the same as 12 DDR4 channels but with higher latency, IBM went with the design they did for super high amounts of memory per CPU.

 

Edit:

Pretty much why I think it's over kill is because you're going to have a really hard time fitting in the number of memory slots required in a dual socket multi blade design, those are 1U half width and are already super tight on space.

POWER9 and its Centaur memory buffer do more than allow for 8TB/socket capacity. It also allows the CPU to be memory agnostic since the Centaur chip handles whatever memory is on the CDIMM itself.

 

Im not sure how 230GB/s is low, considering thats about double anything Intel has available.

 

IBM has the scale up POWER9 variant that uses the buffered CDIMMs but it also has a normal 8 channel DDR4 scale out version and that has 120GB/s of memory bandwidth. In reality, the Centaur buffers also significantly increase capacity AND bandwidth.

 

Especially since each CDIMM has 16MB of L4 cache on the buffer as well. In a system with 128 DIMMs that equals 2GB of L4 cache just on the DIMMs themself.

 

Thats one of the reasons they use up to 512GB CDIMMs in their z14 mainframe as well. Allows for a 40TB(32TB user accessible) capacity as well as ridiculous throughput.

 

Unless you compare it to something like the Fujitsu SPARC XIfx which uses HMC and has 480GB/s memory bandwidth or its successor the Fujitsu A64fx which is 1TB/s using HBM the POWER9 is pretty fast.

[Đỗ Đức Huy]

Well, one might say that Intel is having the CLAP now 

https://www.urbandictionary.com/define.php?term=the clap

for the non native 

[Arika]

11 hours ago, mr moose said:

TDP

oh, now you've gone and done it, you mentioned TDP in an Intel thread. are you trying to summon him?

[leadeater]

40 minutes ago, Amazonsucks said:

Im not sure how 230GB/s is low, considering thats about double anything Intel has available.

It's not low, it's just not 32 channels of bandwidth.

[leadeater]

2 hours ago, porina said:

Isn't part of the offering here is that you now get in 2 sockets what you had to use 4 sockets for previously? Even if the ram space is the same, you just reduced 2 sockets. IMO ram bandwidth hasn't been keeping up especially with HCC CPUs and this at least prevents it from getting worse.

Those small blades don't come in 4 sockets because there isn't enough space for it.