acidx has quit [Remote host closed the connection]
chiselfuse has quit [Ping timeout: 272 seconds]
fedaykin has quit [Quit: leaving]
da5id has quit [Quit: WeeChat 4.7.0]
fedaykin has joined #osdev
edr has quit [Quit: Leaving]
eau has quit [Server closed connection]
eau has joined #osdev
skipwich has quit [Remote host closed the connection]
skipwich has joined #osdev
karenw has quit [Ping timeout: 248 seconds]
<klys>
anyone here have a radeon rx 9070 or 9060?
<klys>
right so that's a pcie5 card
<klys>
anyone here have a pcie5 mobo?
goliath has joined #osdev
<klys>
I see an amd 9755 (64/128) going for less than 2K at ebay, and a gigabyte rx 9070 xt for ~400. mobo? suggesting the gigabyte mz73-lm2, specs at https://www.gigabyte.com/Enterprise/Server-Motherboard/MZ73-LM2-rev-3x why pcie 5.0? gfx12-generic offers conditional branching it seems.
<geist>
i mean it doesn't relaly have different latency, but i guess the idea is by virtually doing this, one L3 will tedn to hold addresses for this chunk of ram and another will tend to hold this
<heat>
yeah
<geist>
and from a scheduling point of view it'll keep the amount of cross L3 down a bit
<geist>
since the L3 will be physically tagged, etc
<zid>
you basically need a threadripper for there to be ACTUAL numa, though
<geist>
and even that, post zen 1 designs threadrippers ceased being actual numa
<zid>
because otherwise there's enough space around the memory controller die for them all to just be the same distance :p
<geist>
they're basically a much larger desktop thing: bunch of CCXes connected to a large io die with a lot of memory controllers
<geist>
this fake numa thing may help for things where you're running large separate large processes that are basically idenepdent. ie, a VM host
<geist>
oh side note: i ran some benchmarks with my new 9950x3d which has two CCDs one with more L3 and one without
<geist>
so i ran some compiles (qemu) pinned to one or the other CCD
<geist>
and somewhat like i had heard, the extra L3 doesn't really help something large like a bunch of compiles
<geist>
the difference in compilation was like 2 seconds over the course of like 300 seconds
<geist>
like 305 vs 307
<zid>
yea it's mainly for something that has a large singluar dataset
<zid>
like factorio
<geist>
yah
<zid>
x3d = factorio accelerator
<geist>
i haven't really locally run any benchmarks to see any gaming diffs. the windows scheduler when dealing with game mode is kinda cute: when a program it thinks is a game is running and focused it seems to just hard park CCD1
<geist>
likle instead of just setting the affinity for the game to CCD0 with the x3d, it just globally shuts down the second CCD
<zid>
'game mode'
<geist>
essentially temporarily turns it into a 9800x3d
<zid>
There was actually a bios toggle in earlier bioses to do just that
<zid>
before windows got a bit better at it
<geist>
but that's also why the 9950x3d is about the same gaming benchmark wise (vs the earlyer gen x3ds where CCD0 was clocked substantially lower)
<geist>
yah
<zid>
or people ran core parking systray apps
<geist>
i guess the idea is there are a bunch of helper proceses that are in the rendering path or whatnot, so to be clear instead f tryiong to figure out how much of windows you need to set the affinity for dynamically, just functioally do it for everything
Lucretia has quit [Remote host closed the connection]
Turn_Left has quit [Ping timeout: 256 seconds]
guideX has quit [Ping timeout: 248 seconds]
Turn_Left has joined #osdev
Turn_Left has quit [Read error: Connection reset by peer]