Intel Releases Benchmark Comparison to AMD EPYC and We Weigh In

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
marcoi

marcoi

Well-Known Member
Apr 6, 2013
1,533
289
83
Gotha Florida
I like this quote from the article: "The real answer is “it depends” which is pretty cool from an industry perspective as we now have competition."
Competition always brings better tech, which is what I like to see moving forward.
 
E

Edu

Member
Aug 8, 2017
55
8
8
33
cheezehead

cheezehead

Active Member
Sep 23, 2012
728
176
43
Midwest, US
Edu said:
It's worth noting that some new Epyc related patches are in the Linux 4.15 kernel: "Improved NUMA node balancing for EPYC processors and we are seeing better EPYC server performance with our early benchmarks of this new kernel." The New Features Of Linux 4.15: AMDGPU DC, RISC-V, EPYC Benefits, VR Improvements - Phoronix

Also, it seems to me that Intel are highlighting the use cases where Epyc is at it's weakest (i.e. database tests where the infinity fabric is saturated)
Click to expand...
Of course, all all/most vendors do this.
 
Jeggs101

Jeggs101

Well-Known Member
Dec 29, 2010
1,529
241
63
Edu said:
It's worth noting that some new Epyc related patches are in the Linux 4.15 kernel: "Improved NUMA node balancing for EPYC processors and we are seeing better EPYC server performance with our early benchmarks of this new kernel." The New Features Of Linux 4.15: AMDGPU DC, RISC-V, EPYC Benefits, VR Improvements - Phoronix

Also, it seems to me that Intel are highlighting the use cases where Epyc is at it's weakest (i.e. database tests where the infinity fabric is saturated)
Click to expand...
Phoro is big AMD fanboys. "Improved balancing" doesn't really fix the root architecture. If you need a 16 core VM it doesn't help you much.
 
  • Like
Reactions: cheezehead, Evan and Pepe
E

Edu

Member
Aug 8, 2017
55
8
8
33
In the article it says: "If you are doing big NVIDIA CUDA AI/ deep learning training you will first run into the roadblock of P2P not working on AMD EPYC, you will then see the impacts of this. One of the best known autonomous automobile/ fleet companies and another major hyper-scaler told us this is why they are not using EPYC with NVIDIA GPUs for deep learning at present."

Well, look at this:
M-Class
An EPYC server with 10 Tesla GPUs!
 
Patrick

Patrick

Administrator
Staff member
Dec 21, 2010
12,513
5,805
113
@Edu I saw that. 8x GPUs on the single root not 10. You are right, I oversimplified the reasoning. There are other concerns. For example, you often do image manipulation/ data prep on the CPU before pushing to the GPUs so you want as much on the connected die as you can, not up to 8 cores on EPYC.

I have no doubt people will try it, but as I talk to some of the big names doing this stuff, I am not hearing that it is close and that a small server vendor coming out with a box will change it.
 
E

Edu

Member
Aug 8, 2017
55
8
8
33
Patrick said:
@Edu I saw that. 8x GPUs on the single root not 10. You are right, I oversimplified the reasoning. There are other concerns. For example, you often do image manipulation/ data prep on the CPU before pushing to the GPUs so you want as much on the connected die as you can, not up to 8 cores on EPYC.

I have no doubt people will try it, but as I talk to some of the big names doing this stuff, I am not hearing that it is close and that a small server vendor coming out with a box will change it.
Click to expand...
That's a good point, but also maybe they don't use EPYC just because there is no EPYC GPU server with PCIe extender on the market yet
 
You must log in or register to reply here.
Top Bottom