1950x with Titan Pascal (or Xp) terrbile gaming performance

G

groebuck

2[H]4U
Joined
Mar 9, 2000
Messages
2,636
I have a fresh install of windows 11 on 1950x Threadripper - all drivers up to date. Put in my Titan Xp which should run just as well as a t080ti - but all I get is choppy graphics and frame drops. I am not pushing 4k but rather 3240x1440 on a gsync enabled momnitor - game I use is Outriders. I serached youtube and found a video og a guy running it at 4k on ultra settiings and it was butter smooth???

So can't be the card - something screwy between windows and threadripper?
 
IIRC, the old TR multi CCD chips were terrible at gaming until you turned on Ryzen Master's gaming mode which disabled half the chip. The latency between the chiplets was the cause of the terrible FPS drops, and gaming mode was the "fix."
 
arestavo said:
IIRC, the old TR multi CCD chips were terrible at gaming until you turned on Ryzen Master's gaming mode which disabled half the chip. The latency between the chiplets was the cause of the terrible FPS drops, and gaming mode was the "fix."
Click to expand...

At one point I had "upgraded" to a Threadripper 2920X from a Core i7 5960X. At the same time I switched from a pair of GTX 1080 Ti's to a single RTX 2080 Ti. I found my gaming experience at 4K to be worse most of the time in the new configuration. Naturally, I did a lot of digging and testing. As a reviewer, I had a lot more access to hardware and had collected a lot of data to being with. The results of all my testing lead me to one inescapable conclusion. The 1st and 2nd Generation Threadripper CPU's were horrible at gaming. The thing is, if you had looked at most of the data out there you'd have seen average FPS ranges that were close if not better than their Ryzen counterparts and not horrible compared to Intel's then champion, the 9900K. However, average FPS does not tell the whole story. (It never does and never did.)

Conventional wisdom at the time suggested that at 4K you were entirely GPU bound and the CPU you used was irrelevant. My testing showed that this was not the case. Average frame rates were broadly similar across the 2920X, 9900K and I think a Ryzen 2700X or whatever Ryzen chip I had tested for that at the time. On the surface, the Threadripper seemed like it would be fine. However, when I'd play any games at 4K (some more so than others) the games weren't smooth. I'd have issues with what felt like hitching or stuttering occasionally. When I started gathering the data and parsing it out I saw that the Threadripper system would spike a lot higher than the 9900K, but that it also dropped into frame rate ranges well below anything else I tested. With Destiny 2, I got frame rate spikes into the 200's or more but it would often drop to 26FPS. With PBO it would do 36FPS instead. Meanwhile the 9900K never dropped below 56FPS in the game test. I never felt that drop with the Intel system but it happened a lot with the Threadripper system. Essentially, the Intel only dropped once to 56FPS and the rest of the time was solidly over 60FPS. The AMD TR system sometimes averaged higher FPS, but it dropped into that 26-36FPS range quite often.

I essentially concluded that the AMD system was all over the place and the reason the averages were so high in tests came down to the fact that the frame rates spiked super high, skewing the results. If you just took a random snapshot of the FPS at a some random time you'd generally see something similar ot the 9900K system. Those drops into the 26-36FPS range occurred relatively frequently but only for a second. Just long enough to feel it in the game. You had higher highs and lower lows with it but far less consistency. This comes down to the fact that there was a lot of latency crossing CCX complexes as well as CCD's. In the 3000 series, AMD compensated by doubling their L3 cache size. Even that didn't work all the time. This is why single CCD chips were slightly better at gaming than two CCD chips were. The infinity fabric helped as it basically made everything a bit more uniform but again, there were latencies internal to those CPU's that held the chips back.

It's not really until the 5000 series that AMD corrected this problem.
 
Dan_D said:
It's not really until the 5000 series that AMD corrected this problem.
Click to expand...
And as you probably know, the 7k series to have noticeable chiplet to chiplet latency again. Maybe not as bad as 1000 and 2000 series but it's there.
 
Oh yeah. I upgraded to the 7900x on my gaming system from the 5900x. I am regretting that. Games crash, or just disappear in the middle of matches. I thought it was psu so I upgraded that and nope still happening. From what I am reading AMD+Nvidia+win11 = fail.
 
groebuck said:
Oh yeah. I upgraded to the 7900x on my gaming system from the 5900x. I am regretting that. Games crash, or just disappear in the middle of matches. I thought it was psu so I upgraded that and nope still happening. From what I am reading AMD+Nvidia+win11 = fail.
Click to expand...
I plan on going to the 77(8?)00x3D when its released from my 5800x. One chiplet, and loads of cache.
 
I never had an issue with the 5900x so I thought going 7900 was a no brainer since all the review said OMG it screams. I knew I should wait...re the 1950x and gaming. It was my main rig a while back and never had an issue gaming. I don't think I ran Outriders on it but like I said I have seen it on youtube in 4k on a pascal run butter smooth. I was going to put it in my intel system but hey guess what the alienware piece of shit won't boot. i hate computers.
 
OFaceSIG said:
And as you probably know, the 7k series to have noticeable chiplet to chiplet latency again. Maybe not as bad as 1000 and 2000 series but it's there.
Click to expand...
noticable in benchmarks, or causing problems with games?
 
OFaceSIG said:
It probably depends on the game. If the .1% lows are bad enough, then you'll start noticing in game.
Click to expand...

Essentially that's the short version of what I said above. Even at resolutions where you are primarily GPU bound, the 19xx and 29xx series can't provide a smooth gaming experience in some cases. Destiny 2 was certainly an outlier in that it was an extreme example, but it wasn't the only one. Just the most extreme case. It's worth noting that at the time, you needed a 2080 Ti and an overclocked 9900K to stay above 60FPS 100% of the time.

In other games like Ghost Recon Breakpoint, the difference wasn't substantial if I recall correctly. Still, issues that I personally experienced with my 2920X are why I would go so far as to say those CPUs aren't that good at gaming.
 
Agreed with Dan. I've got both generations of Threadripper (x399 and sTRX40), and the x399 is... not great for gaming. Really hit or miss - and I'm at 4k60 locked. Some games work fine. A couple of others, don't. Others are in the middle. 2080ti.

Now my 3960X I've only found one game that really didn't play nicely, and that may be the AMD GPU (Beyond Human), as some of what they do on the DX12 side is apparently weird. (6800XT @ 1440P).
 
Dan_D said:
Essentially that's the short version of what I said above. Even at resolutions where you are primarily GPU bound, the 19xx and 29xx series can't provide a smooth gaming experience in some cases. Destiny 2 was certainly an outlier in that it was an extreme example, but it wasn't the only one. Just the most extreme case. It's worth noting that at the time, you needed a 2080 Ti and an overclocked 9900K to stay above 60FPS 100% of the time.

In other games like Ghost Recon Breakpoint, the difference wasn't substantial if I recall correctly. Still, issues that I personally experienced with my 2920X are why I would go so far as to say those CPUs aren't that good at gaming.
Click to expand...

Yeah, I remember reading about the 19xx and 29xx parts being terrible for gaming years ago. OFaceSIG's comment was the first time I'd seen similar concerns about the higher end 7xxx parts. I'm concerned because I don't use my PC exclusively for gaming and the doubled core count of the 7950X vs a 7700X looked appealing for other compute heavy uses.
 
I run a 2950x as my main system with a 2080ti. My build is about as non-standard as they come, but I found that fixing high DPC latency solved a lot of issues with running the 2950x in normal (not gaming) mode. I suspect it's one of the those "correlation, not causation" things, but do what you can to fix the high DPC latency and you'll find a lot of the stutters will be gone. Lot's of work in 2018-19 trying to tackle that issue... modded bios (MSI Creation, there was an OCN thread about this ages ago), 1usmus memory tweaks, applying every latency fix under the sun in Windows. End result: I've not had any stutters since and I have played a lot of games released in the past 3 years at either 2160p/1440p.
 
NUMA. You'll need to set memory interleaving to channel in your bios to expose the processor topology to the OS.

Think of your machine as a having two 1800x cpus with a really fast interconnect between them. Each memory module, PCIe device, usb controller, etc. in the machine is connected to one of those CPUs.

Some applications are NUMA-aware, and can make smart decisions about memory access on their own. However, many applications are not NUMA-aware and some are particularly sensitive to the system topology.

Controlling process affinity is your friend here. On Windows 11, you can use start.exe and the /NODE and /AFFINITY options to pin a process to a specific NUMA node and help minimize cross-node access.
 
OFaceSIG said:
Anandtech might have said that but many other reviews are noticing much worse game performance with dual chiplet CPUs versus single. Many many instances where the 7700x is outperforming the 7900 and 7950 in gaming.
Click to expand...

Okay, link then? Here's one I found with the 7700X and 7950X performing essentially identically in games, every game they tested really: https://www.techpowerup.com/review/amd-ryzen-7-7700x/19.html

Either way, I don't think you can look at better objective measurements of Zen 4 latencies vs Zen 3 and still claim that some reviews showing worse game performance has anything to do with inter-CCD communication vs the previous gen.
 
1667405276693.png

I know it's one game, but the 12 game average is still higher with the 7700X as well.

1667405470975.png


AMD's dual CCD CPUs also suffer from a performance related bug in this title, which is why the 7900X is so much slower. Disabling a CCD resolves this, but then you only have half as many cores, so that kind of defeats the purpose of that part.
Click to expand...
 
2% difference in the 12 game test isn’t really much of a difference.
 
CAD4466HK said:
View attachment 523639
I know it's one game, but the 12 game average is still higher with the 7700X as well.

View attachment 523640
Click to expand...

Thanks for the graphs. Still, the fact is the latencies are better vs. Zen 3. It seems that the new CPUs are fast enough that even better latency numbers are now a slight bottleneck in one game at 500 fps. That is a far cry from "noticeable chiplet to chiplet latency" as if there has been a regression when there really hasn't.
 
repoman0 said:
Okay, link then? Here's one I found with the 7700X and 7950X performing essentially identically in games, every game they tested really: https://www.techpowerup.com/review/amd-ryzen-7-7700x/19.html

Either way, I don't think you can look at better objective measurements of Zen 4 latencies vs Zen 3 and still claim that some reviews showing worse game performance has anything to do with inter-CCD communication vs the previous gen.
Click to expand...
Techpowerup tested those CPU's with a 3080, Techspot used a 4090. ;)
 
CAD4466HK said:
5.7GHz boost vs 5.4GHz boost.
170w TDP vs 105w TDP.
Those differences shouldn't exist. But because of the CCD latency, they do.
Sorry, not trying to derail the thread.
Click to expand...
Still insignificant- we’re well into the rounding errors level at that point. Sure it exists - but it doesn’t make something unplayable like the OP has. And wouldn’t stop me from buying a 7950 if I needed the cores AND gamed.
 
lopoetve said:
Still insignificant- we’re well into the rounding errors level at that point. Sure it exists - but it doesn’t make something unplayable like the OP has. And wouldn’t stop me from buying a 7950 if I needed the cores AND gamed.
Click to expand...
Agreed.
Op's first gen TR might not be the only issue.
 
CAD4466HK said:
Techpowerup tested those CPU's with a 3080, Techspot used a 4090. ;)
Click to expand...

It’s all pretty beside the point. You can game on a 7700X or a 7950X now and have a great, essentially identical experience. If your PC is also a workstation you have eight more very nearly compromise-free cores with the higher end SKU.

That obviously wasn’t the case back in the Zen 1 days.
 
I will probably just go back to making my 1950x system a hypervisor and run my labs on it. Shits all over the NUC work gave me lol. I have a 2700x and my old 5900x, might just slap together a second gaming system and give it to my nephew.
 
repoman0 said:
It’s all pretty beside the point. You can game on a 7700X or a 7950X now and have a great, essentially identical experience. If your PC is also a workstation you have eight more very nearly compromise-free cores with the higher end SKU.

That obviously wasn’t the case back in the Zen 1 days.
Click to expand...
Should it be an identical experience? More frequency, double the cores, and double the cache? For me, that's the issue. Also, that comes from someone who leans AMD and will probably be buying more AMD in the future.
 
OFaceSIG said:
Should it be an identical experience? More frequency, double the cores, and double the cache? For me, that's the issue. Also, that comes from someone who leans AMD and will probably be buying more AMD in the future.
Click to expand...
Frequency matters a bit, but 300mhz on 5ghz is... not that significant. Cache is per-core (this isn't x3d after all), and the cores don't matter (yet) for gaming, unless you're doing ashes of the singularity or a couple of other games that really love the threads.
groebuck said:
I will probably just go back to making my 1950x system a hypervisor and run my labs on it. Shits all over the NUC work gave me lol. I have a 2700x and my old 5900x, might just slap together a second gaming system and give it to my nephew.
Click to expand...
There's definitely a reason I have a "gaming box" with nothing else on it but games and entertainment software. But I have that luxury. My workstation can game, but if I hit an odd issue like this in something, I got no issue bumping it over to the gaming box and saying "fuck it" to fighting with it too.
 
I did game with it before i made it my esxi box and don't remember any issues. I thought maybe it was outriders was just to intensive for Xp but I saw someone playing it at 4k butter smooth on Xp so that can't be it. it might be some tuning but not sure I care enough to get it working as I have a few other systems I can drop it in. I may just go back to making it a workhorse lol.
 
SamuelL421 said:
I run a 2950x as my main system with a 2080ti. My build is about as non-standard as they come, but I found that fixing high DPC latency solved a lot of issues with running the 2950x in normal (not gaming) mode. I suspect it's one of the those "correlation, not causation" things, but do what you can to fix the high DPC latency and you'll find a lot of the stutters will be gone. Lot's of work in 2018-19 trying to tackle that issue... modded bios (MSI Creation, there was an OCN thread about this ages ago), 1usmus memory tweaks, applying every latency fix under the sun in Windows. End result: I've not had any stutters since and I have played a lot of games released in the past 3 years at either 2160p/1440p.
Click to expand...
Posting this for posterity, searched my old bookmarks and found this link: https://audiosex.pro/threads/dpc-latency-guide-possible-solutions.53253/
I can't speak to all the steps outlined there, but I remember doing similar when I was working on the latency and stuttering issues with X399. These steps, along with dialing in bios and memory should solve a lot of X399 stutters.

groebuck said:
I did game with it before i made it my esxi box and don't remember any issues. I thought maybe it was outriders was just to intensive for Xp but I saw someone playing it at 4k butter smooth on Xp so that can't be it. it might be some tuning but not sure I care enough to get it working as I have a few other systems I can drop it in. I may just go back to making it a workhorse lol.
Click to expand...
X399 is better left as a workhorse in my experience. Much effort involved to get a 2950x performing at/beyond 2700x in gaming - that's a big ask in 2022 when 2700x gaming performance is just middle of the road.
 
lopoetve said:
2% difference in the 12 game test isn’t really much of a difference.
Click to expand...
Agreed. And if 2% was enough bother me, I'd probably go Intel since the 13900 has 5% on the 7700.

The 7950 is a few percent worse at gaming but roughly twice as good at wide compute heavy stuff vs the 7700. For a multi purpose system that's a tradeoff I'm willing to make. Just like I'm going AMD for the potential of a 4 year platform instead of 2, and letting other people deal with all the edge cases from 2 different core types.
 
groebuck said:
I will probably just go back to making my 1950x system a hypervisor and run my labs on it. Shits all over the NUC work gave me lol. I have a 2700x and my old 5900x, might just slap together a second gaming system and give it to my nephew.
Click to expand...
Under a hypervisor cant you set the affinity of CPU cores on the vm allowing you to isolate the vm on one chiplet?

My main desktop is on esxi and there is almost no performance loss gaming in a vm.
 
cdabc123 said:
Under a hypervisor cant you set the affinity of CPU cores on the vm allowing you to isolate the vm on one chiplet?

My main desktop is on esxi and there is almost no performance loss gaming in a vm.
Click to expand...
Not so much - Type-1 hypervisors like ESXi don't understand different cores - because that has never, and likely never will, appear in the enterprise space. Even ARM tends to go with unified cores for their server setups. Running ESXi requires telling it to ignore the core differences (or it PSODs), and the little bit of study I've done says that performance can be inconsistent - it sees all the cores the same, so you may get something heavy hitting an E core or the like. For Type-2, you're relying on the OS (windows/linux/etc) interpreting the command to schedule it on the right core, which... I haven't actually tested much. It should work, but I have no idea how well - I do too much nested virtualization which would probably confuse the shit out of the schedulers. Also why I tend to use HEDT for my workstations.
 
I use it exclusivly for work running ovas for demos, encryption, tokenization etc. So I just need horsepower and memory---and disk space but the 18TB spinner takes care of that I put all my highly transactional stuff on nvme.
 
lopoetve said:
Not so much - Type-1 hypervisors like ESXi don't understand different cores - because that has never, and likely never will, appear in the enterprise space. Even ARM tends to go with unified cores for their server setups. Running ESXi requires telling it to ignore the core differences (or it PSODs), and the little bit of study I've done says that performance can be inconsistent - it sees all the cores the same, so you may get something heavy hitting an E core or the like. For Type-2, you're relying on the OS (windows/linux/etc) interpreting the command to schedule it on the right core, which... I haven't actually tested much. It should work, but I have no idea how well - I do too much nested virtualization which would probably confuse the shit out of the schedulers. Also why I tend to use HEDT for my workstations.
Click to expand...
I see ability to dedicate spacific CPU threads on multiprocessor systems to individual vms in this documentation.
https://docs.vmware.com/en/VMware-v...UID-F40F901D-C1A7-43E2-90AF-E6F98C960E4B.html
I'm assuming that would allow for grabbing spacific parts of a chip including grabbing only cores vs hyperthreaded cores. Provided the local resources and dependancies of the physical CPU are not strained (like grabbing just threads for a vm). I think the AMD chips are segregated in such a way to allow this. But I've never messed with it on my 3900x and the 'fabric' based chips I've used (phi chips) had oddities with dedicating different parts of the chip.
 
cdabc123 said:
I see ability to dedicate spacific CPU threads on multiprocessor systems to individual vms in this documentation.
https://docs.vmware.com/en/VMware-v...UID-F40F901D-C1A7-43E2-90AF-E6F98C960E4B.html
I'm assuming that would allow for grabbing spacific parts of a chip including grabbing only cores vs hyperthreaded cores. Provided the local resources and dependancies of the physical CPU are not strained (like grabbing just threads for a vm). I think the AMD chips are segregated in such a way to allow this. But I've never messed with it on my 3900x and the 'fabric' based chips I've used (phi chips) had oddities with dedicating different parts of the chip.
Click to expand...
You’re going to do that in an environment with DRS? Not sure how that would work unless you’re intentionally building multiple alder lake boxes…. And if you’re doing that, just buy an amd box and skip it and let the scheduler use every core. At normal consolidation rates you’ll run out of ram before cores.
 
lopoetve said:
You’re going to do that in an environment with DRS? Not sure how that would work unless you’re intentionally building multiple alder lake boxes…. And if you’re doing that, just buy an amd box and skip it and let the scheduler use every core. At normal consolidation rates you’ll run out of ram before cores.
Click to expand...
No? I don't think any clusters are involved. Just one local host where as I read it you could assign a single chiplet to a vm and leave the rest of the CPU to do other vms.
 
cdabc123 said:
No? I don't think any clusters are involved. Just one local host where as I read it you could assign a single chiplet to a vm and leave the rest of the CPU to do other vms.
Click to expand...
I mean sure, but that's not what I run :p

And also is a pain in the ass - you ever move to new hardware without that issue you have to remember you did that. Also no guarantee settings like those are still valid through updates.
 
You must log in or register to reply here.
Back
Top