i've discovered assigning a gayme's affinity to all odd/even cores improves performance a bit
i'm guessing it's because smt/hyperthreading is shit or is windows at fault here?
disabling smt negates this but prevents the machine from sleeping for some reason
r5 1600 win7 btw, any difference on win10/intel?
is smt/hyperthreading really worth the 20 or so % multithread increase with a potential decrease in gaymen fps?
> r5 1600 win7
There are 3 cores per ccx.. You want to run your gaymen w/ 3 cores? be your own guest.
As for why performance increases depending on core assignment, this is due to Ryzen's architecture :
> i.imgur.com
If you go across the ccx to the other 3 cores here is the latency penalty from ram :
> pcper.com
You will go from 40ns to 140ns latency to access DRAM
Also, it doesn't help that you're running windows 7 as they cut off updates before ryzen cpu support was added. So, i'm pretty sure the scheduler is thrashing the fuck out of the cores w.r.t to thread assignment.
> pcper.com
correction : this is the communication latency between cores
Stop fucking with this, it's not made for retards like you
>it's AMD loser realizes his mistake episode
SMT is always a double-edged sword. Some programs will benefit tremendously from it, others will see small losses. Very few programs will lose a lot from it, which is why it's turned on by default.
>pcper.com
I still haven't seen any explanation what they actually mean by "ping time", more precisely, since there's no sequence of instructions that is commonly thought of as a "ping" between cores.
If you're a gamer, you should find a cpu with incredible single core performance. Or dual core performance. Because most games aren't optimized to large amounts of cores.
Large core amounts are best for assigning virtual machines to them and running server stuff.
>So, i'm pretty sure the scheduler is thrashing the fuck out of the cores w.r.t to thread assignment.
Windows 7 should already be capable of reading the CPU topology from the ACPI tables, doe.
i want to thank you for providing an on topic post instead of babbling about random shit
what if I want to play multiple copies of a game at the same time? I can play 6 atm but would like to scale up to 16.
i hate you please shut the fuck up
stop playing video games
multiboxers pay for performance, so why are there no metrics for us?
I welcome the idiot to the party...
Magically, everyone forgot the Intel Stutterfest(tm) even with their "high-end" 7700k 4corelets.
>Windows 7
It does, but it doesn't know how to treat an unknown CPU because the scheduler hasn't got patches for that μArch.
the scheduler doesn't know where to assign the main threads and where the children(I know that winblows don't have child processes, but I am loaning the term from POSIX)
'Ping' is reflective of the best case latency you're going to see between a thread running on two separate physical cores. There's tons of communication between cores related to multi-threading, etc. A game engine is going to have tons of core to core communication thus why you're seeing a 20% bump if you locate the processes on the same physical CCX. What you're looking at is a more than double increase in latency if they aren't which is why you see a 20% performance drop when you just let windows' retarded scheduler fuck around w/ the threads.
Windows 7 doesn't properly support Ryzen because its scheduler doesn't take this into consideration and threads get thrashed about. So, you have two things to keep in mind :
You're running windows 7 and ryzen really is a numa architecture. Think about it as though you have two sockets w/ 3 cores each. You want to due cross socket communication as less as possible due to the added latency.
Make sense?
multi-what?
> mfw poorfags
Did you even read the article you linked? It's all about Windows 7 lacking USB XHCI support before having drivers installed and the motherboard only supporting XHCI rather than EHCI. It doesn't mention a word about CPU topology.
>It does, but it doesn't know how to treat an unknown CPU because the scheduler hasn't got patches for that μArch.
The point being that the ACPI tables describe the CPU topology and so the OS shouldn't require a driver for each and every specific CPU model just to understand its layout. There may be other issues, like using an optimal idle loop and such things, but that's another thing.
??????????????? excuse me
this is Jow Forums not /v/ piss off
>'Ping' is reflective of the best case latency you're going to see between a thread running on two separate physical cores. There's tons of communication between cores related to multi-threading, etc.
Thanks, Cap'n Obvious.
What I meant was what it means specifically. Taking a cache line from another core? Pulling a cache line from core A to core B and back again? APIC latency? Something else entirely?
>this is Jow Forums not /v/ piss off
Discussing the factors that affect gayming performance is safely in Jow Forums territory.
you're also gay, so you should head over to /lbgt/
Maybe, but that's an orthogonal question.
the topology cannot tell the scheduler how to treat each core(s) with every load.
Just remember the infamous windows 7 scheduler patch that gave Bulldozer's CMT implementation a 5% performance jump.
Don't forget that SMT and CMT have shared resources with their physical partners, placing the wrong load on the "weak" thread, it can cause many more hazards and nops especially with FP instructions.
>What I meant was what it means specifically.
noone knows, those metrics are not valid unless AMD verifies them.
There are several cases where you should check the latencies,
cross-CCX cache r/w
cross-CCX context switching
data forwarding from one core to another core on the other CCX.
...and even then you won't have a clear picture, because all those latencies are eliminated with a simple scheduler patch.
>excuse me
you are excused.
>the topology cannot tell the scheduler how to treat each core(s) with every load.
That's exactly what it does. It's the whole point of it.
>Just remember the infamous windows 7 scheduler patch that gave Bulldozer's CMT implementation a 5% performance jump.
That's because AMD's topology description of Bulldozer presented each "core" as an independent core, which is kind of understandable since the standard description format has (or had) no conception of CMT. Zen however has no such exotic attributes so it fits nicely into the standard topology descriptions.
>the "weak" thread
There is no weak thread. Both threads of a core are symmetric.
>noone knows, those metrics are not valid unless AMD verifies them.
It's not AMD's graph.
>dual core shill
t. 2008