AMD Ryzen 7000 Series Make For Compelling Budget Servers, Leading Performance & Value Over Xeon E

[phoronix]

Phoronix: AMD Ryzen 7000 Series Make For Compelling Budget Servers, Leading Performance & Value Over Xeon E

While this summer has been a busy season of benchmarking with the new AMD EPYC Bergamo processors providing up to 128-cores / 256-threads per socket and the new EPYC Genoa-X parts providing up to 1.1GB of L3 cache with 3D V-Cache to provide for excellent HPC performance, not everyone needs such levels of performance nor having the budget for such platforms. It's always fun talking about the high-end server platforms, but at the opposite end AMD and their platform partners have been rolling out an equally interesting assortment of AMD Ryzen 7000 series based server products. With the Ryzen 9 7950X/7950X3D providing up to 16-cores / 32-threads, a growing number of Ryzen server motherboards supporting DDR5 ECC UDIMM, and a number of innovative Ryzen server platforms coming to market, it's an interesting time to be after a budget-friendly server platform or other robust rackmount systems where looking for power efficient 16 cores or less configurations.

https://www.phoronix.com/review/ryzen-7000-servers

[ernstp]

Also nice for a (CPU-focused) workstation now with the integrated GPU.

[S.Pam]

How much is the total cost for the servers in these tests?

[skeevy420]

Michael While I'm assuming you used the default settings here, have you ran tests on the 7900x3d and 7950x3d with and without the cores on the 2nd CCX disabled to see how that affects things on Linux? A lot of Windows game reviewers and benchmarkers have the slower, lower core count 7800x3d performing better than the faster, higher core count x3d models due to how the v-cache is laid out.

[pegasus]

What is the availability of ECC UDIMMs? In my past experience these have always been a major pain point with such "low end" servers. Very poor choice and those you could get caused your mobo to refuse to boot. Hope this improves with DDR5 gen.

[skeevy420]

How much is the total cost for the servers in these tests?

A quick search shows ~1100+tax and shipping for the barebones server. Ram is ~120 a stick. Article has MSRP processor prices and not good deals like how yesterday y'all could have found 7800x3d's on the Bay for as low as $305 before tax and shipping.

So about $2000 or so for a higher spec'd one.

[Michael]

Michael While I'm assuming you used the default settings here, have you ran tests on the 7900x3d and 7950x3d with and without the cores on the 2nd CCX disabled to see how that affects things on Linux? A lot of Windows game reviewers and benchmarkers have the slower, lower core count 7800x3d performing better than the faster, higher core count x3d models due to how the v-cache is laid out.

Defaults. Not too worthwhile for multi-threaded workloads to segregate them since making use of all the cores will pan out better for said tests than limiting it to just a subset anyhow.

[Michael]

What is the availability of ECC UDIMMs? In my past experience these have always been a major pain point with such "low end" servers. Very poor choice and those you could get caused your mobo to refuse to boot. Hope this improves with DDR5 gen.

Seems to be working out pretty reliably so far with this ASRock Rack server.

[grigi]

There something seriously wrong with the joules/run calculations:

For "Timed Godot Game Engine Compilation" example:
Xeon E-2388G took 472.30 seconds, and used 92.86W avg.
Ryzen 5 5600X took 425.74 seconds, and used 73.73W avg.

Somehow that calculates to:
Xeon E-2388G: 8729 Joules/run
Ryzen 5 5600X: 31728 Joules/run

So the Ryzen 5 does it faster and uses less power during the task, but uses nearly 4x as much power for the task?

[S.Pam]

Michael While I'm assuming you used the default settings here, have you ran tests on the 7900x3d and 7950x3d with and without the cores on the 2nd CCX disabled to see how that affects things on Linux? A lot of Windows game reviewers and benchmarkers have the slower, lower core count 7800x3d performing better than the faster, higher core count x3d models due to how the v-cache is laid out.

Also, enabling different numa settings in bios, if available can make a huge difference too. On my EPYC server it makes cinebench 20% faster!