AMD Ryzen 9 7950X3D Linux Performance

[phoronix]

Phoronix: AMD Ryzen 9 7950X3D Linux Performance

Ahead of tomorrow's launch of the AMD Ryzen 7800X3D / 7900X3D / 7950X3D processors, today marks the embargo expiry on the flagship Ryzen 9 7950X3D 3D V-Cache processor. Today I can share with you the initial performance around the performance of this $699 USD processor that features a 144MB cache.

https://www.phoronix.com/review/amd-ryzen9-7950x3d-linux

[Volta]

Awesome Ryzen performance. Not even a single reason to consider Intel as an alternative.

[L_A_G]

Seeing the comments from Intel fanboys on some tech sites on the leaked benchmarks; This result is going to trigger a good few people.

Still kind of annoying that AM5 motherboards are hugely overpriced (paid way too much for my B650 Prime Plus with it's crap VRM configuration), but at least AMD has the top end of the market sorted out once these things hit retail.

[joshx1]

I only care about code compiling etc, and workstation use, gaming is ¯\_(ツ)_/¯ for me .

Slightly slower in some benchmarks but MUCH lower thermals... interesting. Probably some half-decent overclocking to come for even better performance?

[GruenSein]

I wonder what the results would look like if both dies had some V-Cache. I mean, there has got to be some overhead/non-optimal use due to this heterogeneous setup.

[uid313]

What makes this CPU good is it that it is energy efficient for multi-threaded workloads.

Total energy consumption.
Rendering of the scene "BMW Car Demo" in Blender 2.93, graphics card excluded.

[zxy_thf]

Can we have numbers on CPU temperatures?

[intelfx]

This added complexity has the potential of being more troublesome on Linux than Windows, to which AMD and Microsoft have been working to prepare Windows and the AMD drivers for properly determining proper core selection between cache and frequency preferences.

<...>

AMD's Windows drivers with the AMD 3D V-Cache Performance Optimizer Service evaluate performance in real-time and modifies the preferred cores handling based on workload for desiring frequency or cache.

Yeah, naturally.

[hax0r]

Somebody tell me there's way to unlock TDP to at least 170W? Imagine shelling out $700 shekels just to get capped at 120W. Sucks that both CCDs are not equal, for proper performance it would require programs to be V-Cache & and CCD aware so kernel scheduler puts them on more performant cores, knowing the crappy state of Linux scheduler that's not happening. I can see people disabling CCD0 to get better performance (and losing the cores that paid for)

[unwind-protect]

Deciding whether to run any given thread on the more-cache die or the higher-frequency die is highly non-trivial. Even if the Windows "driver" uses performance counters to measure cache miss rates and then advise the OS scheduler to put high cache miss threads on the cores with more cache that is only part of the equation. You still don't know whether in absolute terms the high cache miss thread might still be faster on the turbo cores.

You also face a decision problem on a highly loaded system. If the turbo cores wouldn't be able to go into high turbo anyway due to thermal load then what? The equation changes and recommendations for the high-turbo cores might not work out. Since you are fully loaded you don't have enough high-cache cores to run everything.

Also, from a power consumption standpoint you will see better numbers by biasing toward the high-cache counts instead of constantly turboing. How is the scheduler going to know what the user wants? Except by setting things manually.

I am not too worried that Windows is getting a noticeable advantage out of having this.