AMD Ryzen CPU Core Scaling Performance

[M@GOid]

Thank you Michael for this test. It confirms a suspicion I have, that the majority of the games out there do not scale beyond 4 cores. Rysen R3/5, here I come.

[indepe]

3-0 means that there are three out of four cores from the first CCX active, while there are none active from the second CCX.

Why isn't it possible to enable 3 cores on the first CCX, as well as 3 cores on the second CCX? Is that in the BIOS and depends on the motherboard, and/or could even be changed with a BIOS update?

Does 1+1 use SMT and 2-0 does not, as one would think? If so, then SMT seems to be working surprisingly well.

[juno]

Why isn't it possible to enable 3 cores on the first CCX, as well as 3 cores on the second CCX? Is that in the BIOS and depends on the motherboard, and/or could even be changed with a BIOS update?

Oh, but it is, and Michael actually did that. See SIX (3 + 3)
I think any motherboard/bios should have that feature. On Windows, you can even do that with an application (called Ryzen Master).

Does 1+1 use SMT and 2-0 does not, as one would think?

Noone with a little knowledge of Zen would expect that and that question has been answered for multiple times already.

[VikingGe]

Even with this not testing SMT but the core arrangement, it is still a very interesting test setup because it does show where the interface between the two CCX is a bottleneck and where it isn't. Windows sites should test this as well, especially for gaming.

[juno]

Even with this not testing SMT but the core arrangement, it is still a very interesting test setup because it does show where the interface between the two CCX is a bottleneck and where it isn't. Windows sites should test this as well, especially for gaming.

The reviewers from a German site actually tested that for a few games: http://www.pcgameshardware.de/Ryzen-...ew-1222033/#a5
If you don't speak the language, the pictures should speak for themselves. Watch out for the gallery right above the conclusion ("AMD Ryzen R7 1800X: Fazit") 4+0 is better than 2+2 in general.

[indepe]

Noone with a little knowledge of Zen would expect that and that question has been answered for multiple times already.

I just understood it the way the article seemed to be suggesting, and your answers lacked the verbosity for me to understand exactly what you were saying. I don't think you can expect anyone here to be well informed regarding the Zen's inner architecture.

So... I guess one might think that in most test cases here there isn't a lot of shared data across all threads, since otherwise 2+2 would be more often slower than 4-0.

Dota2/Vulkan still seems to be an outlier with its own problems.

[VikingGe]

If you don't speak the language

Well actually I am German I even read that review, kind of embarrassing that I already forgot that they did in fact test that setup. Anyway, thanks for the reminder.

There's still one problem though, they only tested well-threaded games that do reasonably well on Ryzen anyway, not the ones where it is even slower than the FX-9590. What I'd like to see is whether 4+0 is better than 4+4 in those scenarios.

[indepe]

The relatively small difference between 2+2 and 4+0 is expected.

A general rule: If a scalable benchmark (such as: Timed Linux Kernel Compilation, C-Ray) has high IPC then 2-cores-4-threads should perform measurably slower than 4-cores.

IPC = Instructions Per Clock as measured by the Linux perf tool. A high IPC - in year-2017 terms - equals to about 1.5 or more instructions per clock.

As I just learned, 2 + 2 doesn't mean 2 cores 4 threads, it means 4 cores, where 2 cores are in the first so-called CCX group, and the other 2 cores in the second CCX group. Each CCX group has its own L3 cache, where each core in that CCX group has equal access to this L3 cache.

[xpander]

Does simple "taskset -c 0,2,4,6" or similar help with the Dota2 case? so you don't have to go to UEFI to turn off cores for gaming.

[indepe]

Interesting, very interesting. The "real" computing went as expected but holy cow at the Vulkan results with Dota. That issue should be inspected closely be developers of AMD firmware, Kernel, GPU driver and the game developers as well.

I don't think this has anything to do with AMD, as there are also strange results with i7 vs i5 when running Dota2/Vulkan. On the plus side, it may mean that Vulkan performance is much better than current test results (with Talos Principle currently out of the picture).