AMD Ryzen 7 1800X vs. Intel Core i7 7700K Linux Gaming Performance

[nuetzel]

Michael
Please test with SMT disabled as fast as you can.
Here are some numbers for starting (very promising):

https://i.redd.it/d5r8sd9qt0jy.jpg

[Michael]

Michael
Please test with SMT disabled as fast as you can.
Here are some numbers for starting (very promising):

https://i.redd.it/d5r8sd9qt0jy.jpg

As I mentioned in this very article, the BIOS of this motherboard doesn't give me the option to disable SMT.

[Michael_S]

Michael
Please test with SMT disabled as fast as you can.
Here are some numbers for starting (very promising):

https://i.redd.it/d5r8sd9qt0jy.jpg

That screenshot doesn't show the display resolution and graphics settings of the benchmarks. In benchmarks with 4K resolution, 1440p resolution, or even some of the benchmarks with 1080p resolution but the graphics settings on the highest levels, the difference between Ryzen and the recent Core i7s is tiny or even zero.

But that's because the GPU is doing all of the work. To compare the processors, the benchmarks should be 1920x1080 with low graphics settings or even lower resolutions. And in those, from what I've seen disabling SMT gives a 3-5% performance boost to Ryzen but not enough to catch Intel.

I'm an AMD fanboy. I plan to get an R7 1700 or R5 later this year. But it looks like the Core i7-7700 is still a better buy for gaming enthusiasts. No matter how much I wish differently, that won't change the reality.

[nuetzel]

As I mentioned in this very article, the BIOS of this motherboard doesn't give me the option to disable SMT.

Sorry, Michael I read it, but.
Sad. You didn't get a newer BIOS in the meantime...
Couldn't we do something with Linux 'self tools'?
Limit core number?
thread groups/siblings

[indepe]

Sorry, Michael I read it, but.
Sad. You didn't get a newer BIOS in the meantime...
Couldn't we do something with Linux 'self tools'?
Limit core number?
thread groups/siblings

The "isolcpus" kernel boot parameter? List or don't list the logical cores corresponding to SMT. Or maybe taskset is sufficient.

[nuetzel]

It seems to be a board and/or BIOS problem.
Golem saw mean 17% increase with MSI Board and newer BIOS.
With Asus even a little more.
Sorry German:

https://www.golem.de/news/ryzen-7-1800x-im-test-amd-ist-endlich-zurueck-1703-125996-4.html

[atmartens]

TDP == Thermal Design Power. That's the heat the processor can emit. But the actual power consumption is higher.

If the power consumption goes up then so too does the heat. You can't have one without the other.

[LeJimster]

I was hoping zen would overclock a little better. But maybe the 6 and 4 core will go higher. More wait and see. I didn't want to beta test the new chips anyway.

[zboson]

As I mentioned in this very article, the BIOS of this motherboard doesn't give me the option to disable SMT.

In that case you can try something like

numactl -C 0,2,4,6,8,10,12,14 ./a.out

Then the executable only sees 8 cores and the topology is scattered.

You could also try

numactl -C 0,2,4,6,8,10,12,14,1,3,5,7,9,11,13,15 ./a.out

This I think will ensure that the first eight threads run on eight cores and any addition threads uses SMT.

[zboson]

That's a great point about the thread splitting. But it would be bizarre for AMD not to have addressed this right out of the gate, if it's a common issue. "We're going to spend billions of dollars designing Bulldozer and later Ryzen, but not spend a few million getting the necessary patches into the Linux kernel and Windows kernel for the multi-threading to work efficiently." ??

I believe it's possible that in some rare cases if a single application is splitting threads across a single core it might be better than splitting across multiple cores. If there is frequent intra-thread communication, the shared L2 cache could allow that communication to run much faster. But that's my wild speculation.

AMD never addressed this with the bulldozer based systems which have been out for several years now. For example if you use OpenMP with a bulldozer based system with 4 modules, and eight threads and you set the number of threads to 4 you will find that the 4 threads run on only two modules and if you're doing floating point work that means only two FPUs are being used when you could have used four. This can depend on the OS and distro. You really have to check the thread topology of your system, they ultimately have control on the default toplogy. I just know what I have seen on every system I have checked. I would also not be surprised if AMD uses the same CPUID tags for SMT and cluster multi-threading mapping.