Threadripper benchmarks breakdown

[grimreeper132]

11 hours ago, leadeater said:

This is a gaming focused forum after all, people usually talk about areas that are most relatable to themselves. I did make a comment a day or so ago though that AMD could in the next revision of Zen/TR implement a more sophisticated power management feature that could solely boost one die when the workload permits, you would still need NUMA mode on though to push the load on to a single die more reliably unless game engines become TR aware (HA not likely).

 

Zen+ TR with 8 core 4.5GHz boost would be awesome, dreams are free though.

depends what you mean by "TR aware" if by that you mean scale with cores better, then that might happen, we might get games which can use multiple threads, maybe even up to 32 (doubt it though, and if it does I doubt it'll be worth it) but I can see games slowly getting more benefit from more threads, I can definitely see them using up to maybe 12-16 threads soonish (in a year or two, not with the AAA games from this production cycle but probably with the one after, maybe) so there might be a point with it in the future, but yea your right at the moment there's not

[LAwLz]

16 minutes ago, ravenshrike said:

Except if you're not WORKING, as in the thing that people do to make money, why would you need to have it in all core mode? So you switch it, get up to pee, and then start playing games. The vast, vast majority of people don't switch between working and playing games willy nilly.

Yeah... That will not happen.

Have you seen how small the difference in performance is? I seriously doubt many people will actually do all the work of closing down everything, changing the setting, restarting and then starting all their programs plus the game again just to get like 80 FPS instead of 70 FPS. Not worth all the hassle.

And if you are going to buy a 16 core processor just to keep it in 8 core mode for most of the time then you might want to reconsider your purchasing decision.

 

 

 

 

13 minutes ago, XenosTech said:

You do know an EPYC single socket solution is about $3000 here and some other places just for the cheapest cpu after taxes and shit. Not all of us are going to spend that kind of money on just the cpu alone. 1900x would be plenty enough for me or the 10 core (if it ever does come to market), I just want a ton of lanes for cheap as possible. Cheap and intel don't go in the same sentence past the consumer platform

What are you going to use it for?

You might be among the 0.1% my post does not apply to.

 

 

 

17 minutes ago, leadeater said:

The 7351P is $750 so less than the 1950X 

 

Edit:

Interesting the 7251 is $475 and 7281 is $650.

Yep, EPYC CPUs are cheaper than ThreadRipper.

There is a 1GHz difference though (between the 7351P and the 1950X).

[leadeater]

1 minute ago, grimreeper132 said:

depends what you mean by "TR aware" if by that you mean scale with cores better, then that might happen, we might get games which can use multiple threads

More that they can see NUMA nodes natively and assign threads properly within the dies to avoid cross die communication when it's not necessary. Scaling to more cores isn't TR problem it's cores talking to others on a different die or pulling from memory using the other die's IMC.

 

If a game engine can only effectively use 8 cores for example it should only use 8 on the same die on the 1950X. For the 1920X it should put threads that communicate with each other the most on the same die so you might have 4 +4 or 6 + 2 etc.

[grimreeper132]

1 minute ago, leadeater said:

More that they can see NUMA nodes natively and assign threads properly within the dies to avoid cross die communication when it's not necessary. Scaling to more cores isn't TR problem it's cores talking to others on a different die or pulling from memory using the other die's IMC.

 

If a game engine can only effectively use 8 cores for example it should only use 8 on the same die on the 1950X. For the 1920X it should put threads that communicate with each other the most on the same die so you might have 4 +4 or 6 + 2 etc.

oh right that makes sense, as I said I wasn't particularly sure what you meant by "TR awar" but that makes more sense now you've said that

[leadeater]

3 minutes ago, LAwLz said:

Yep, EPYC CPUs are cheaper than ThreadRipper.

There is a 1GHz difference though (between the 7351P and the 1950X).

Not a problem for my use case though. I only run VMs and storage nodes and don't stress my CPUs at all. For me I just need cores/threads and I/O connectivity options. Being a lab I don't need super high end performance but I have needed to run over 100 VMs simultaneously before which is a reasonable strain on storage I/O, RAM and pCPU:vCPU ratio or contention.

[XenosTech]

43 minutes ago, LAwLz said:

What are you going to use it for?

You might be among the 0.1% my post does not apply to.

For the most part in my country the enterprise area is mostly virtualized so it's strictly going to be for VM's and I have a box of drives somewhere in my room or in storage that I'd like to make use of. I can't run much VM on my current PC at a time since I'm limited by what I can fit on my 1tb before I have to delete and add more for whatever I plan to test

[ApolloFury]

7 hours ago, LAwLz said:

Let's be real here, nobody is going to constantly go into the BIOS and change between gaming and creator mode, nor is anyone going to pay  a ~600 dollar price premium for a ~100MHz increase. That's just something they did for marketing purposes.

Only a really small amount of games actually need game mode. Codemaster games are the main offenders but not sure what else.

I think people are over-exaggerating about some games being incompatible with high-core counts.  

Don't get me wrong, it is a nuisance having to do that.

[leadeater]

5 minutes ago, ApolloFury said:

Only a really small amount of games actually need game mode. Codemaster games are the main offenders but not sure what else.

I think people are over-exaggerating about some games being incompatible with high-core counts.

Yea it's more like won't utilize rather than incompatible. If you're mostly gaming it's a bit like hiring a truck and trailer to move a box of books, a car will do that just fine but if you want to move a library on the other hand... 

[Bouzoo]

15 hours ago, leadeater said:

It's not even just old games, news games don't understand NUMA at all since that is normally a two socket system which in gaming is even more rare than dual GPU. Game engines don't have any thread management logic to keep cross communication within the same NUMA node so will often pick a non optimal core to place a thread on causing a latency hit for inter-core communication.

 

AMD is counting on Windows thread management which is NUMA aware to place threads in the same NUMA boundary, game mode makes the CPU report as 2 NUMA nodes and creator mode makes the CPU report as single NUMA node.

 

Edit:

Also high end workstation applications are usually NUMA aware like Solid Works or Adobe Suite as those are used in dual socket Xeon workstations.

I wasn't talking about NUMA, I'm talking about simple PCs with 1 CPU. But while we're at NUMA, all that I say below pretty much applies in even far, far more rare scenario. Did I Imply I was talking about NUMA? My bad if I did. 

 

14 hours ago, Misanthrope said:

4.5 would yield useful increases in IPC for when people might want to do 144hz gaming yet you can reboot and have your 12/24 for non-action games and productivity, just a minute or two.

 

See this computer devices can actually do many different things and people also might have varied interests with them like playing Counterstrike then editing counterstrike video footage.

Um... IPC has nothing to do with clock speeds. What you are referring is instructions per seconds which is IPC x GHz (if you want). You either mixed it up or worded it badly?

 

See, I'm pretty sure no one would be bothered to save your settings, projects, something, and restart your PC and manage settings so you can game. Every. Single. Time.

 

 

 

Sorry to both about late replies, I... forgot you quoted me. 

[Misanthrope]

8 minutes ago, Bouzoo said:

I wasn't talking about NUMA, I'm talking about simple PCs with 1 CPU. But while we're at NUMA, all that I say below pretty much applies in even far, far more rare scenario. Did I Imply I was talking about NUMA? My bad if I did. 

 

Um... IPC has nothing to do with clock speeds. What you are referring is instructions per seconds which is IPC x GHz (if you want). You either mixed it up or worded it badly?

 

See, I'm pretty sure no one would be bothered to save your settings, projects, something, and restart your PC and manage settings so you can game. Every. Single. Time.

 

 

 

Sorry to both about late replies, I... forgot you quoted me. 

Ok single core performance then. Misused the term but you know what I mean.

[leadeater]

1 hour ago, Bouzoo said:

I wasn't talking about NUMA, I'm talking about simple PCs with 1 CPU. But while we're at NUMA, all that I say below pretty much applies in even far, far more rare scenario. Did I Imply I was talking about NUMA? My bad if I did. 

Threadripper has NUMA and UMA modes so when talking about NUMA here it's not meaning dual socket like it normally would be.

 

NUMA and UMA are referring to memory sub system nodes or essentially memory controllers, since Threadripper has two IMCs it has two nodes. AMD has a switchable mode for the CPU in how it reports the memory nodes to the OS and applications: UMA or NUMA.

 

In UMA mode it's reporting the CPU as having only a single node memory sub system and every core is local to every memory module when in reality they are not, some RAM is remote but while operating in UMA mode it is not possible to know which. Using remote memory has a higher latency which is why you see in some games minimum frame rates are lower.

 

In NUMA mode it's reporting the CPU as having two memory sub system nodes which is more accurate to the actual hardware which will allow the OS and NUMA aware applications to more accurately place threads on cores to ensure performance consistency. However the downside there is that the effective memory bandwidth is halved so in very parallel workloads the performance drops. The between node latency is much lower than in standard dual socket systems and the bandwidth is also much higher.

 

Creator mode is UMA and SMT on

Game mode is NUMA and SMT off

 

Edit:

Updated post with info

[goodtofufriday]

6 hours ago, leadeater said:

Not a problem for my use case though. I only run VMs and storage nodes and don't stress my CPUs at all. For me I just need cores/threads and I/O connectivity options. Being a lab I don't need super high end performance but I have needed to run over 100 VMs simultaneously before which is a reasonable strain on storage I/O, RAM and pCPU:vCPU ratio or contention.

And here i am just running a pfsense vm off my nas.