Hi all,
Just wandering if anyone knows what exactly Legacy mode does on the new 3rd gen threadripper CPUs ? (i thought it disabled CCDs)
Doesnt seem to disable cores but makes windows identify the cpu as 2 Nodes instead of 1 cpu ? - as far as I can tell anyway !!
Anyone have a more detailed explanation or more knowledge they would care to share ?
Regards
Dan
I don´t know for sure either.
But i don’t really think that it would disable anything.
I’m not sure what you really mean by running legacy mode,
Legacy mode in sense like running a legacy Os (like windows7),
on the new Threadripper 3000 series cpu’s.
Or do you mean certain applications like games etc?
Hi Mistery !
In ryzen master there is a legacy mode meant for use in games / applications that have big issues with the # of cores (and maybe latency?) (firestrike Farcry 5 / new dawn for example).
Game / Legacy mode used to disable half the cores on the 19XX and 29XX threadrippers
Ah now i understand what you are aiming at.
Yes the previous gen Threadripper cpu’s,
did had a performance hit on the higher core count cpu’s,
for things like gaming and such.
The Game mode sorta kinda helped a little bit with that i believe,
but it still wasn´t that great from what i have seen,
at least not when it comes to the first gen Threadrippers.
However from what i have seen on the new 3000 series Threadripper cpu’s and gaming benchmarks,
there barely seems to a performance hit anymore,
at least not when it comes to gaming.
Sure in comparison to a intel 9900KS that boosts up to 5Ghz,
that cpu will still win in certain games and single threaded workloads.
But yeah from what i generally see in terms of benchmarks,
is that that even the 3960X / 3970X perform pretty well,
to their Ryzen 3000 counter parts in terms of gaming.
But absolutely shine when it comes to multi threaded productiviy work loads.
But what the gaming mode in the Ryzen master software pretty much did from what i understand,
is that it kinda tricks windows into looking at just one of the two 8 core dies when it comes to first gen Threadripper cpu´s.
And only use the said memory controller of the said die.
This of course improves latency significantly.
But i don´t really think that this issue still exists for the new Threadripper 3000 cpu´s.
At least i have not really heard or read much about it.
And gaming benchmarks don’t particulary showing it either to me.
With the Threadripper 1000 series there was trully a performance hit,
from what i said above.
But of course that performance hit is also relatively,
because it also highly depends on which resolution you game on,
and of course which gpu you use etc.
Maybe @wendell could shine some more detailed light on this.
But i don´t really think that there is much of a compromise,
with the new 3000 series Threadrippers in gaming.
Interesting thing is that not many reviewers seem to,
have talked that much about the gaming mode for TR.
Almost all games work great with new TR. Farcry 5 acts weird but process lasso can feel it in. It’s just fc5 being weird.
The gaming experience on TR was shockingly pleasant
I think “legacy” in this case means that it works like a TR1/2 CPU. Maybe someone has been using TR like a multi socket system with NUMA nodes and uses/is testing an application developed for NUMA access and needs it on TR3 aswell.
Thanks guys,
I have had all generations of TR so far and still have a 1950x…
On first and second gen, there is a legacy mode & a memory mode, memory switches between local & distributed memory access
There was a legacy mode which disabled 1 CCX on the 1900x, 2 on the 1920/1950 2920/2950 & 4 on the 2970wx / 2990wx
What im noticing with 3rd gen is there is only legacy mode, and its not disabling cores but I believe it is doing something with memory access.
In the case of farcry 5, believe if you will, disabling 2 CCDs (turning the 3960x into a 12core) has yeilded some amazing gaming results comapred to my very tweaked 3900x (roughly +20% fps to the min & average & 15% to the max)
Assassins creed origin & rise of the tomb raider have shown exceptional performance in 24c mode both in legacy & normal (yet to test in 12c mode)
It looks like with some proper software implementation these chips could definately be the future !
This might be interesting to investigate further indeed.
Farcry5 is a bit of a weird game in that regards.
But if the performance gains you expierenced here are consistant.
Then that is definitely a significant improvement in performance for Farcry 5.
Interesting that they changed the naming scheme of the said modes by the way.
Indeed it has been interesting, I will keep experimenting, and start working on ram / infinity fabric timings (currently 3600c14 1:1)
Interestingly… aida64 memory latency in 24c mode reports aorund 72ns, drop it down to 12c its around 63ns
pretty sure i could get it very low 60ns in 12c mode with proepr work on the fabric / ram timings
Well I have here some benchmark results…
3900x was at stock CPU 3733 C14
1950x was at 3466 C14
3960x was at 3600 C14
GPU used is a liquid cooled & shunt modded 2080ti at 2160/+1088 & run at 1080p for maximum CPU load
Thanks @wendell @MisteryAngel
Have done more testing, Legacy mode appears to turn thee CPU into 2 nodes, im guessing this does something with memory access to reduce latencies,
it jsut appears that some games still dont like being spread out over all 4 nodes, as you guys will see in my graphs above there is definately an issue when certain games are apread over all CCD’s, this is something I hope they will fix (developers / amd / windows ?)
For now im going to play around mroe with some infinty fabric tweaking but then look at what program i could use or run script to set games to only use XX cores to prevent the need for 12c mode or legacy node (though im not seeing any real negative impacts of legacy mode yet)
Will keep the data & info coming !
!This is pretty interesting, i´m going to follow this closely.
Best 24core latency test ive been able to achieve so far.
This is in legacy mode hoever !
Normal mode puts it back up to around 71ns
What i need to do now, is find a way to test cross CCX latency, and the individual core to memory latency.
I suppose that this might be an windows schedular issue.
I’m not sure if you could run that aida64 memory test on linux,
just for a comparison.
I believe that @wendell also addressed this topic with the TR1 review back in the day.
It looks that there actually is an improvement for some games.
which is kinda interesting.
I also checked the core useage of the games to see which were using all CCDs vs which were not !
F1 2018 uses all the CCDs are runs really well on all 24, 7 Days to Die is using all 4 CCDs but basically doubles its perfroemance if you restrict it to 1 ccd !
Also noted… PBO doesnt work in “legacy” mod !
Have done some tweaking, im seeing about 4211mhz all 24 core with PBO on under repeated Cinebench R20, and around 509 single core and single core spikes of 4617mhz
am wandering if they will allow PBO with legacy mode as legacy does seem to help general performance, atleast in my situation when ram timings are tweaked !
