Retesting AMD Ryzen Threadripper’s Game Mode: Halving Cores for More Performance | Hacker News Reader

Retesting AMD Ryzen Threadripper’s Game Mode: Halving Cores for More Performance

anandtech.com

59 pointsbauta-steen17 comments

I wonder if someone has tested it for Ryzen SIGSEGVs and hangs. I know AMD said it's not affected, but it'd be good to see some third party confirmation to that.

Yep, would be good to know for sure whether it's affected or not.

Yes its been tested by a bunch of folks at phoronix. The script that causes issues on Ryzen has so far not caused any issues on either Epyc or ThreadRipper.

That's good to know. I've been itching to put in an order.

cool! any link to the results/discussions? many thanks!

I thought that it disabled SMT, not the actual cores. That would make it 16c/16t not 8c/16t.

No, I'm an idiot. It turns off one of the Zen modules, to kill the intermodule latency.

That seems like it could less destructively be done by setting CPU affinities appropriately. And, by extension, automatically through better kernel code.

Or is there more to it than that?

I wonder if NT considers Threadripper a NUMA.

Yes, Threadripper presents two NUMA nodes.

So this one I actually know - Threadripper can be configured (to present) in NUMA or UMA modes.

Sounds like "content" mode is UMA and all cores/threads active. "Game" mode is NUMA, but with half the CPUs disabled.

Seems like most folks with a decent OS would want NUMA and all cores/threads active. Is that possible?

Yes, but the software actually needs to be NUMA aware. A lot of current server software works that way - groups of tightly coupled threads target the same node, which prevents major performance issues in multi-CPU systems.

In general such systems are very good at multitasking, but not so good at running a single demanding application (if it is not NUMA aware).

The same thing applies with Threadripper - albeit with lower internode latency.

I guess the next step is to make game engines and web browsers NUMA aware.

I'd put money that automatically handling the affinity per CCX would give you a good chunk of performance boost, but that it would require a substantial amount of work to the OS that wouldn't be completed before benchmarks and reviews came out.

Given that Ryzen 3/5/7 performance has improved since they were released, but only a few of the reviews have been updated, it's likely that AMD wanted to avoid that and put in Game Mode as a stopgap.

If it is in their power to do so, they should do it. It helps users. If it happens in the OS or it happens in hardware, it really doesn't matter.

From the last page: http://www.anandtech.com/show/11726/retesting-amd-ryzen-thre...

(I think this answers your question?)

> Robert Hallock on AMD's Threadripper webcast has stated that Windows Scheduler is not capable of specifically zeroing out a full die to have the same effect. The UMA/NUMA implementation can be managed with Windows Scheduler to assign threads to where the data is (or assign data to where the threads are), but as far as fully disabling a die in the OS requires a restart.

Reply on news.ycombinator.com