Intel Offers More Cascade Lake-AP Performance Numbers(96 core Versus 64 core EPYC 7601)

csbin

Senior member
Feb 4, 2013
838
351
136
Cascade Lake-AP 96 core AVX512 (1.8Ghz?)
EPYC 7601 64 core AVX2 128bit (2.xGhz?)
EPYC Rome 128 core AVX2 256bit (2.xGhz?)

https://www.anandtech.com/show/13586/intel-offers-more-cascade-lakeap-performance-numbers


Intel%20Pre-SC18%20Press%20Update-page-006.jpg



results based on projections ?:coldsweat:
AMD use intel compiler o_O

Intel%20Pre-SC18%20Press%20Update-page-013.jpg



LisaSu_NextHorizon_Close_07.jpg
 
Last edited:

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
They must really be panicking if they are releasing *more* internal numbers so soon after the original release.

Everyone with more than two brain cells to rub together knows internal numbers mean slightly less than nothing - from both AMD and Intel.

Astonishingly, you'd think it was Intel desperately trying to grab mindshare from AMD - throwing more numbers into the public domain so quickly is akin to shouting "hey, over here, don't forget about us" after not getting the initial reaction hoped for.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
well cascade lake with 12CH memory can be a real benefit
the price and power is important now
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
well cascade lake with 12CH memory can be a real benefit
the price and power is important now

Yep, as a CFDer in a past life, with that memory bandwidth it definitely can.

But if power limits means it actually runs quite a bit slower in the FPU, then the bandwidth may not matter.
 

piesquared

Golden Member
Oct 16, 2006
1,651
473
136
Has intel said yet if these include the liquid coolers or if they will be extra?
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
Isn't Cascade Lake-AP supposed to be the replacement for Phi? Why are they benching it against EPYC when it would be GPU products positioned most directly against it?
I also thought about this. I mean it performs the best in very parallel BW heavy tasks, that take advantage from AVX-512. Many of such tasks run even better on GPUs, which offer more BW and parallelism.

The main problem with gpus is rewriting your codebase in CUDA, or the fact that the problem might not be easily be splittable to chunks thst fit into the GPU memory. But these look like niches of a niche market
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
To make themselves look better ?

For the same reason Mayweather fought McGregor.

Yeah, but that also means that Intel is creating some confusion among potential buyers. Unless Intel is just going to spam the market with Cascade Lake-AP for awhile as their premier Xeon product for pretty much everything.

The current HPC paradigm is highly-parallel, efficient CPUs hosting compute cards, whether they be GPUs (Summit) or Xeon Phi cards (Tianhe). Cascade Lake-AP sort of takes the "you can have your cake and eat it too" approach further than Phi ever did, by being fully-capable x86 "general purpose" CPUs that just happen to have a little grunt in the HPC arena as well.

Still, I don't see Cascade Lake-AP stacking up very well against anything from Nvidia.

I also thought about this. I mean it performs the best in very parallel BW heavy tasks, that take advantage from AVX-512. Many of such tasks run even better on GPUs, which offer more BW and parallelism.

The main problem with gpus is rewriting your codebase in CUDA, or the fact that the problem might not be easily be splittable to chunks thst fit into the GPU memory. But these look like niches of a niche market

Most of the HPC market has embraced CUDA and/or OpenCL. I have to admit, it would be stupidly easy to code for Cascade Lake-AP, though. Easier than Phi ever was.
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
Most of the HPC market has embraced CUDA and/or OpenCL. I have to admit, it would be stupidly easy to code for Cascade Lake-AP, though. Easier than Phi ever was.

Yeah, CUDA isn't anywhere near as big an issue as it was ~10 years ago. I also don't see much point in Cascade lake-AP. The fact that it is 2x24-cores "glued together" lessens the value a lot compared to what it would otherwise have (at least a lot of cores connected in a cache-coherent manner). Now Zen2 seems to fill most of it's niches, unless memory bandwidth is absolutely everything.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
Yeah, CUDA isn't anywhere near as big an issue as it was ~10 years ago. I also don't see much point in Cascade lake-AP. The fact that it is 2x24-cores "glued together" lessens the value a lot compared to what it would otherwise have (at least a lot of cores connected in a cache-coherent manner). Now Zen2 seems to fill most of it's niches, unless memory bandwidth is absolutely everything.

As I mentioned in one of the earlier threads, there were rumors that the Core Xeon Phi replacement was going to include 3 additional AVX-512 units instead of just 1. Probably on 10 nm. So Cascade Lake-AP would be something they hacked together since they obviously can't deliver that.
 

dark zero

Platinum Member
Jun 2, 2015
2,655
138
106
Intel is doing the wrong moves again despite they wanted to change for the better.... Oh boy. That won't end good.
 

Topweasel

Diamond Member
Oct 19, 2000
5,436
1,654
136
Yeah, especially given the 10 nm debacle.
Yeah that's what 3 major product lines being axed another 2 getting pushed back. You hope that they haven't had to shelve actual developments. But that assumes they were really trying to push it in the first place. It's probably still another 3 years till we Intel's real reaction to Epyc and Ryzen.
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,075
136
2x CLX-AP Benchmark?o_O
source : chinese sns
So if 2.5ghz is the base clock are we looking at ~400watt TDP?

I wonder if intel are just going to target HPC with these or are they going to be priced to compete with Rome across all 1S and 2S markets?

I admire Intel for making lemonade with lemons, but if Rome is 2.25ghz base on 64 cores in 180watt tdp its still beating it at significantly less power.