[microtonal]

But the large margins are in Xeons for enterprise and HPC computing. We use high-end Xeon machines for research and would think twice before purchasing AMD. Besides a generally good reputation and good performance, Intel has a great ecosystem with Intel MKL, ICC, etc.

Besides that, Spectre also affects AMD and many ARM CPUs.

[tomxor]

> We use high-end Xeon machines for research and would think twice before purchasing AMD

Ryzen supposedly have a price-performance benefit over Xeons for multi-threaded workloads.

What would you say the main reasons for staying with Xeons are for your type of research? Is single threaded performance intrinsic to the type of computation your research requires? or do those optimised math libraries you mentioned bring so much practical benefit as to outweight any current raw price-performance differences?

Soz, lots of questions, was just interested in how it applies to HPC based research in reality.

[microtonal]

What would you say the main reasons for staying with Xeons are for your type of research?

We use Intel MKL in some applications. We just use the BLAS interface, but Intel MKL is generally the fastest BLAS implementation. Obviously, it is optimized for Intel CPUs.

Also, with many vendors it is easier to get >= 768GB RAM 64 core machines using Intel Xeon CPUs.

[earenndil]

You could use mir glas[1], which has similar performance[2], but works on (pretty much) any architecture.

1: https://github.com/libmir/mir-glas

2: http://blog.mir.dlang.io/glas/benchmark/openblas/2016/09/23/...

[marmaduke]

At least my case, our compute servers are 4 dual Xeon servers in a 2U format, sold by Dell. I haven’t yet found any vendor with the same core density based on AMD.

[floatboth]

There are 2U dual EPYC servers e.g. from Supermicro

https://www.servethehome.com/dual-amd-epyc-7601-processor-pe...

[ihsw2]

The Dell EMC PowerEdge R7425 is a dual Epyc server in 2U format, sold by Dell. The top end Epyc chip gets you 64 cores (128 threads). The core density is competitive, along with its supported IO device (eg: NVMe drive) density.

[theevilsharpie]

Supermicro blows AMD out of the water when it comes to density.

Here's equivalent 4 node in 2U AMD systems (they differ based on disk configuration -- SAS, NVME, or a mix of both). They also support Supermicro SIOM, so you can choose what integrated networking you want:

http://www.supermicro.com/Aplus/system/2U/2123/AS-2123BT-HNC...

http://www.supermicro.com/Aplus/system/2U/2123/AS-2123BT-HNR...

http://www.supermicro.com/Aplus/system/2U/2123/AS-2123BT-HTR...

If you wanted even more density, there's the MicroCloud and MicroBlade series (although they use Intel processors):

https://www.supermicro.com/products/MicroBlade/

https://www.supermicro.com/products/nfo/MicroCloud.cfm

That being said, these don't make a lot of sense unless you're very space-constrained, or your racks have a ton of power available. Given a 30A 208V circuit, I could only safely run three of the above AMD systems at full power. I'd rather just get dedicated 1U or 2U servers, and not have to make compromises about expandability or serviceability.

[marmaduke]

thanks for the links. I hadn’t found any vendors on those 4 node systems.