AMD Ryzen Threadripper 3990X Offers Incredible Linux Performance

[mppix]

Yeah, LTT calls out AMD on this one too. The lack of RDIMM/LR-DIMM support really limits the appeal of this CPU. AMD apparently hopes to sell more Epyc CPUs through these market segmentation games.

In principle, I'd understand that strategy. However, I don't see many 1P EPYC workstation motherboards that directly compete with Xeon W.

64 Zen1 cores is now "cheap" to build. Epyc 7551 costs around 1300 EUR here, a dual socket Supermicro H11DSi costs around 600 EUR. Add cooler, case, PSU, SSD and you will be at around 3600 EUR. For Zen2 with Epyc 7452 add another 2000 EUR.

Thank you for the suggestion. Unfortunatley, not all of our workloads are NUMA aware and can properly leverage a second socket.

[mppix]

Samsung has a product that fits, but it's not available yet in volume. In the meantime there's the 2666MHz version with the same status, but it's available for purchase. I've had success with running Samsung UDIMM ECC 2400 at 2933 or even 3200MHz on Zen+/2, but it requires a bit of tuning just like Threadripper 3

Thank you for the Samsung 3200 memory suggestion - sounds like something worth waiting for.

[mppix]

Related question, does anyone know how well intel optane performs as \swap?
They did persistent RAM, but how is the nvme performance?

[Linuxxx]

On a 3970x you can have idle=poll with no difference in power usage so something is up with acpi_idle.

So what You are saying is that these Threadrippers NEVER enter their sleep-states?

That alone would already put them at a huge advantage versus the Intel counter-parts! (With greatly increased power-draw, of course...)

Again, bridgman, any explanation for that behavior?

[Michael]

So what You are saying is that these Threadrippers NEVER enter their sleep-states?

That alone would already put them at a huge advantage versus the Intel counter-parts! (With greatly increased power-draw, of course...)

Again, bridgman, any explanation for that behavior?

When measuring AC power draw with Ryzen/Threadripper/EPYC is clearly a big difference between idle and load on all my systems tested.

[Linuxxx]

When measuring AC power draw with Ryzen/Threadripper/EPYC is clearly a big difference between idle and load on all my systems tested.

OK then, dispat0r, care to chime in & try to find out what's wrong with Your setup?

Also, I'd still be interested if AMD plans to ever finish & mainline their special CPU driver for the Zen architecture?
(Build around schedutil, I assume, since that what Linux developers requested from AMD.)

[dispat0r]

So what You are saying is that these Threadrippers NEVER enter their sleep-states?

That alone would already put them at a huge advantage versus the Intel counter-parts! (With greatly increased power-draw, of course...)

Again, bridgman, any explanation for that behavior?

Not exactly I mean the idle power draw is the same with idle=poll as without and a can see with perf top hat the acpi_idle functions are not active.

[vb_linux]

It may be this or their roadmap vision, but you seem to be jumping through hoops to somehow justify your bias.

[tildearrow]

"Fortunately" you can still take hours to compile Chromium

Hmmmm...

My 6700K took 8 hours to compile Chromium...

...which means: 64/4=16

(8/16)*4.0/2.9=approx. 0.69 hours on a 3990X (not considering architecture differences)

[xinorom]

It may be this or their roadmap vision, but you seem to be jumping through hoops to somehow justify your bias.

Last edited by tildearrow; 02-09-2020, 07:44 PM.

Way to make your censorship obvious, tildearrow . If you're going to delete my original post and his quoting of it, you may just as well hide your tracks and delete his whole post too.

Looks like we got a new Reddit-style power mod over here.