AMD Ryzen 7 7800X3D Linux Performance

[phoronix]

Phoronix: AMD Ryzen 7 7800X3D Linux Performance

While the AMD Ryzen 9 7950X3D and Ryzen 9 7900X3D processors went on sale at the end of February as the first Zen 4 3D V-Cache processors, today marks the availability of the Ryzen 7 7800X3D processor. I've recently been putting the 7800X3D through its paces under Linux and have a plethora of benchmark data to share for launch day.

https://www.phoronix.com/review/amd-ryzen-7-7800x3d-linux

[milkylainen]

Super impressive numbers under load. Idle numbers are not so pretty however.

Say you have this typical workstation, you work 8h, say you peak the machine about 1h out of those 8 on a work day.
The machine stays always on at work and sits at 20-30W more (or whatever) during the rest of the hours.
That's going to add up. Significantly so.

Real world numbers matter.

[middy]

With regards to gaming, if amd wants to continue the multiple ccds with 3d cache, they are going to need to do some work to improve scheduling on linux to pin games to the ccd with the cache like they did on windows with their chipset drivers and using xbox game bar to help feed it info. If you want to game on linux and want amd, the 7800x3d is the only oblivious choice. No scheduling woes.

[avis]

Super impressive numbers under load. Idle numbers are not so pretty however.

Say you have this typical workstation, you work 8h, say you peak the machine about 1h out of those 8 on a work day.
The machine stays always on at work and sits at 20-30W more (or whatever) during the rest of the hours.
That's going to add up. Significantly so.

Real world numbers matter.

For some reasons AMD has been unable to power gate the idling IO die however they've long solved the issue for their mobile CPUs. I don't know how and why. I was expecting desktop Zen 4 CPUs to fair better in this regard vs the previous four generations of Zen CPUs but it didn't happen.

Overall these extra ~10-14 watts don't matter much (the 13900K idles around 7W, the 7950X around 20W).

[avis]

With regards to gaming, if amd wants to continue the multiple ccds with 3d cache, they are going to need to do some work to improve scheduling on linux to pin games to the ccd with the cache like they did on windows with their chipset drivers and using xbox game bar to help feed it info. If you want to game on linux and want amd, the 7800x3d is the only oblivious choice. No scheduling woes.

taskset has got you covered.

Truth to be told, neither Intel, nor AMD care too much about Linux. The Linux kernel still doesn't properly support ADL which will soon turn two years of age. I guess the X3D parts will meet the same fate.

[schmidtbag]

Very impressive performance-per-watt results. Makes me wonder how much better it could be when undervolted. Pretty enticing to me since I want to prioritize efficiency for my next build, but I don't know how much I care for that price tag.

[V1tol]

Super impressive numbers under load. Idle numbers are not so pretty however.

Say you have this typical workstation, you work 8h, say you peak the machine about 1h out of those 8 on a work day.
The machine stays always on at work and sits at 20-30W more (or whatever) during the rest of the hours.
That's going to add up. Significantly so.

Real world numbers matter.

For some reasons AMD has been unable to power gate the idling IO die however they've long solved the issue for their mobile CPUs. I don't know how and why. I was expecting desktop Zen 4 CPUs to fair better in this regard vs the previous four generations of Zen CPUs but it didn't happen.

Overall these extra ~10-14 watts don't matter much (the 13900K idles around 7W, the 7950X around 20W).

On acpi_cpufreq AMD does not hit its lowest clocks on idle. For example my 5950X idles at 2.2GHz though it can do 800MHz or even 400MHz per core (if I remember Windows numbers correctly). We are just waiting for amd_pstate driver

[joshx1]

With regards to gaming, if amd wants to continue the multiple ccds with 3d cache, they are going to need to do some work to improve scheduling on linux to pin games to the ccd with the cache like they did on windows with their chipset drivers and using xbox game bar to help feed it info. If you want to game on linux and want amd, the 7800x3d is the only oblivious choice. No scheduling woes.

AMD makes fantastic high-core CPUs.. if you don't game or use Linux .

Their scheduling for the 64 core Threadripper 2990WX was (and still is?) absolutely hot garbage on Windows, with tools like Process Lasso etc. Works great under Linux though.

[Anux]

Don't expect pstate to change anything in idle, when you idle the frequency is irelevant because the core hopefully is in some C state. Atleast in arch I can easily activate pstate and it clocks down to 400 MHz but I didn't notice any performance difference or temp changes in idle. Where you might see differences is under full/light load, look at https://www.phoronix.com/review/linux-63-amd-epyc-epp

BTW does anybody have a Ryzen 5XXXG or the 5XXXX mobile variant? What is the lowest C state you can get, I can't get anything better than C3, but maybe that's normal?

[numacross]

On acpi_cpufreq AMD does not hit its lowest clocks on idle. For example my 5950X idles at 2.2GHz though it can do 800MHz or even 400MHz per core (if I remember Windows numbers correctly). We are just waiting for amd_pstate driver

I'm using amd_pstate with ondemand on a 6.2 kernel. With a Ryzen 3700X, 4x16GB 3200MHz RAM, and manual setting of power management features in BIOS [1] I am able to get it to idle around 12.5W with 5.5W going to RAM if the SMU reporting is to believed:



I am pretty sure that the core clocks in sleep are not being correctly reported via the usual mechanisms. Both this utility and HWiNFO under Windows are able to show "real" clocks as seen by the SMU.

For some reasons AMD has been unable to power gate the idling IO die however they've long solved the issue for their mobile CPUs. I don't know how and why. I was expecting desktop Zen 4 CPUs to fair better in this regard vs the previous four generations of Zen CPUs but it didn't happen.

I found that most power management options are disabled by default on Zen desktop platforms (hence [1] above).
The most important one is Infinity Fabric "SoC/Uncore OC Mode" which is enabled by default on every AM4 board I've seen and it keeps IF at maximum performance mode at all times. I am pretty sure this is available as "AMD CBS -> SMU Common Options -> DF Cstates" on most boards.
PCIe L1 modes are disabled, most likely due to possible compatibility issues.
"Power Supply Idle Control" is not set to "Low Current Idle" because it supposedly causes instability, but is fine in my setup.

I haven't had the chance to play around with any AM5 platforms, so I can't really judge if the situation has improved. All in all AMD's default settings are quite performance-oriented, but can be adjusted for power efficiency even on Zen 2.