Overclock.net banner
13,221 - 13,240 of 31,499 Posts
Have you tried using CPUsets to move all OS interrupts and Windows apps over to the non-vcache CCD then setting the app to test to use only the vcache CCD on the 7950X3D? I don't know what it's like for PyPrime, but it stopped the "7800X3D is just as good as a 7950X3D for games" debate in its tracks. Would assume it'll be less of an effect. Appreciate restarting to apply all interrupt affinities and the CPUset is a pain though.
yes I did that, found that utility on github but it's a pain to set up everything
that's why 7800x3d is better, u don't have to do too much trickery
I'll just stick to manual affinities and xbox app for core parking when I use specific apps, it works but I know I'm not having the best scores this way
Affinities are set all on specific cores, but I avoid to use frequency cores for lan/gpu/mouse
 
For others interested to configure CPUsets (without Processor Lasso) this is a useful tool:
no work will be scheduled to the cores at all, so you need to set processor affinity to them manually (or using Processor Lasso, script, etc) to put the workload on the cores you've freed.

Then for interrupt affinity this is is the easiest tool to use:

Then use PowerShell for network adapters RSS affinity:

It's a lot of work, but gives you a big advantage for gaming workloads on the 7950X3D over the 7800X3D, especially with NVIDIA graphics cards. I don't know if there's much difference with PyPrime though.

EDIT: if you're on the ASUS X670E Extreme with the 10gbe NIC, download driver version 3.7.0 or higher (not available from the ASUS site) to enable 8 RSS queues, as it'll only work properly with 4 and below on the driver they provide.
 
Weird better results with SCL's at 8/8 vs. 5/3..- ( 59.0 to 58.5 & 7.653 to 7.618 )
VDD
: 1.48v VDDQ/IO: 1.30v
Nitro: 1-2-0
View attachment 2643782
Nice

have you tested twrrd = 2

and your trc is really low, I haven’t tested it that low and run 58. I would test trc quite a bit higher and see if there is any auto correction by running it that low
 
For others interested to configure CPUsets (without Processor Lasso) this is a useful tool:
no work will be scheduled to the cores at all, so you need to set processor affinity to them manually (or using Processor Lasso, script, etc) to put the workload on the cores you've freed.

Then for interrupt affinity this is is the easiest tool to use:

Then use PowerShell for network adapters RSS affinity:

It's a lot of work, but gives you a big advantage for gaming workloads on the 7950X3D over the 7800X3D, especially with NVIDIA graphics cards. I don't know if there's much difference with PyPrime though.
Using CPUsets is like shooting yourself in the foot
whenever you need to assign an app all 32 cores you will have to use manual affinity, or a script, or a secondary app (process lasso)

for example it's a not a problem when you test pyprime because it will use only 1 core, but if you start to bench ycruncher or cinebench it will run on half the cores, unless you do additional and more complicate steps.
The same apply on every program that makes use of all the available cores in your system.


Affinity is another matter of discussion, that is usefull, you configure it 1 time and you can forget it, but the way you assign interrupts is different
gpu is simple, open GoInterruptPolicy, find the gpu and set your preferred core, ez
mouse almost the same, use the same app but don't assign the affinity to the device or it won't work: affinity must be set to the entire xHCI controller where it's connected (better have "only" the mouse connected on that entire controller, must play a bit with mb usb ports to find the optimal setup)
network adapter will ignore entirely what you set with GoInterruptPolicy, you need to use Set-NetAdapterRss commands to spread the RSS queues as you wish (number of RSS threads, base core, ecc), check the indirection table and double check with latency mon while you stress test the lan to see which cores are getting used.
 
Using CPUsets is like shooting yourself in the foot
whenever you need to assign an app all 32 cores you will have to use manual affinity, or a script, or a secondary app (process lasso)

for example it's a not a problem when you test pyprime because it will use only 1 core, but if you start to bench ycruncher or cinebench it will run on half the cores, unless you do additional and more complicate steps.
The same apply on every program that makes use of all the available cores in your system.


Affinity is another matter of discussion, and the way you assign interrupts is different
gpu is simple, open GoInterruptPolicy, find the gpu and set your preferred core, ez
mouse almost the same, use the same app but don't assign the affinity to the device or it won't work: affinity must be set to the entire xHCI controller where it's connected (better have "only" the mouse connected on that entire controller, must play a bit with mb usb ports to find the optimal setup)
network adapter will ignore entirely what you set with GoInterruptPolicy, you need to use Set-NetAdapterRss commands to spread the RSS queues as you wish (number of RSS threads, base core, ecc), check the indirection table and double check with latency mon while you stress test the lan to see which cores are getting used.
Yeah but if you're comparing PyPrime on a 7800X3D vs a 7950X3D, having a tranquil otherwise unused CCD on the 7950X3D should nets you better performance.

RSS queues are separate from interrupt affinity and not to be confused (think that's what you're trying to say but just to be sure).
 
RSS queues are separate from interrupt affinity and not to be confused (think that's what you're trying to say but just to be sure).
yeah, the main purpose is the same: spread the workload and move it away from core 0
You can't change msi affinity for the network adapter, but you can modify the indirection table and assign RSS queues to your preferred cores like I said
For example I like to move ndis.sys (4 rss queues) on core 2/4/6/8, and mouse+gpu on core 10


Weird better results with SCL's at 8/8 vs. 5/3..- ( 59.0 to 58.5 & 7.653 to 7.618 )
well, maybe I'm joining the "8-8" club

got my best result in pyprime with both scl on 8, like many others :ROFLMAO:

It's not a "GREAT" result overall because I'm limited to fclk 2133/2167, can't do more, and this bench likes a lot fclk as I said before

Image


PS: while testing 2200 I also corrupted my windows installation: it became slower and slower, and couldn't boot anymore XD
it's one of the first samples of 7950x3d, maybe newest ones have better luck
Strange because 2167 seems fine, tested the stability with ycruncher, I also get more gigaflops than 2133 so there shouldn't be any EC in action at 2167
 
I beg you kind people for a little help... first AMD build ever, just put 7800x3D with B650M Aorus Elite AX. I've got 3 A-Die kits and another M-Die.

I'm literally pulling my hair out and sobbing like a little girl every restart for every voltage or single timing in the BIOS takes like 5 minutes not even exaggerating.
All I wanted to do was test the memory controller capabilities for 1:1 and 1:2 but I'm essentially just stuck waiting for reboots.
Thanks in advance.
 
Since DDR5 has a very fast clock speed, there may be problems with signal quality or stability, so I have seen experts say that a CKD like a server RCD is needed.
View attachment 2643900

But it appears that something is rapidly changing.
Essentially moving from DDR to QDR. This’ll make things even more interesting for those who confuse memory MHz with MT/s, BIOS with UEFI, etc
 
Nice

have you tested twrrd = 2

and your trc is really low, I haven’t tested it that low and run 58. I would test trc quite a bit higher and see if there is any auto correction by running it that low
I've been running tRC low ~34-36 for months now with no issues w/ stability (no errors via many stress test). ASUS has pre set profiles with tRC = tRP ( they have profiles with tRC = 36 )
Running tRC = tRAS + tRP bring slightly higher latency ~ .4 - .5 ns or 59.0 vs. 58.5ns.

If running tRC = tRP was an "issue" I would think ASUS wouldn't have pre set profiles like this.. This is the reason I tried it.

No, haven't tried tWRRD at 2 as 4 works and I doubt a 2 would give much better results. Don't want to re run all stress test for this. Maybe I'll try it just to see if it improves anything.

tRC on DDR5 doesn't behave the same way it did on DDR4. There's probably a point at which tRC can be "too low" but at 34-36 at least it is okay from my experience at 6000-6200MT/s. If it is somehow "too low" I would wonder what ASUS is thinking.. Anybody with ASUS Bios can check for themselves.
 
I've been running tRC low ~34-36 for months now with no issues w/ stability (no errors via many stress test). ASUS has pre set profiles with tRC = tRP ( they have profiles with tRC = 36 )
Running tRC = tRAS + tRP bring slightly higher latency ~ .4 - .5 ns or 59.0 vs. 58.5ns.

If running tRC = tRP was an "issue" I would think ASUS wouldn't have pre set profiles like this.. This is the reason I tried it.

No, haven't tried tWRRD at 2 as 4 works and I doubt a 2 would give much better results. Don't want to re run all stress test for this. Maybe I'll try it just to see if it improves anything.

tRC on DDR5 doesn't behave the same way it did on DDR4. There's probably a point at which tRC can be "too low" but at 34-36 at least it is okay from my experience at 6000-6200MT/s. If it is somehow "too low" I would wonder what ASUS is thinking.. Anybody with ASUS Bios can check for themselves.
Ok, cool (y)

I’ll test trc = trp
 
Rectangle Font Screenshot Parallel Technology

Decided to settle for this for a moment. Memory didn't error so far, but I'll test the stability over night. FCLK 2133 appears to be the sweetspot for my CPU, at least according to GFlops in linpack. 2133 gets 383-385 while 2166 381-383. 2200 is crawl mode. Y-cruncher VT3 didn't error but I guess I'll run that some more too.
 
The information i'm asking for is physical written on the IHS, so you need a picture from the CPU itself. (this cannot be read from software/windows)
For an example, this is one my my earlier 7950X
Circuit component Electronic component Passive circuit component Electronic engineering Computer hardware
Production date is year22 week39
Hi @domdtxdissar , I finally got around to doing some much needed maintenance and got my CPU's info.

Font Material property Number Signage Circle


Looks like a early production run, year 22, week 32?. What can you tell me about this, any reference to what you might have come across in your arsenal?
 
Got my B650M-HDV. I've noticed a weird thing when running 1.4v VDDIO/VDD/VDDQ. For some reason VDDIO is running at 1.35v instead of 1.4v which was set in BIOS. Should it be like that? There's also a PHYRDL mismatch but I'll deal with it later (or not at all).
View attachment 2643882
Mine does the same. I think it's just bugged readout. Sometimes it does read 1.400 but not always. It's stable at 1.350 for me tho.

This boards an absolute beast for RAM OC. 6400CL26 1T GDM Off easily done and 8000CL34 1T GDM One can run at seriously low voltages. For 2200FCLK 8000CL34 tPHYRDL match and everything I only need 1.20 vSOC, 1.050 VDDP, 1.350 VDDIO, stock auto VDDG.
 
  • Rep+
Reactions: hsn and chispy
Temp delta here seems lower than I'm experiencing, but am using 1.65v and 24GB DIMMs. Getting the heat the whole 34mm (from bottom of memory IC to top of DIMM for waterblock contact) being most crucial.
So, I ordered Direct Touch RAM Block during the ICEMAN Winter Sale.
I inquired because delivery seemed to be delayed, and they said they were currently producing it and would ship it around next Monday.


Image
 
Bclk1 = 101MHz
6464/2133 1:1 CL28

Nice latency

Font Screenshot Software Technology Terrestrial plant


Stable TM5 25cycles 1usmus_v3 and y-cruncher 1-7-0 several iterations.
have to test Karhu now and corecycler, as cores boost higher.
Will post also Cinebench23 if full stable.
 
13,221 - 13,240 of 31,499 Posts