7800X3D, 7900X3D, 7950X3D..

[some1]

2 hours ago, Rix said:

Wouldn't utilizing CCX1 cores be useful to prevent the non-vr applications from using the CCX0 cache? 

If other apps need them, the cores are unparked, it's not like they shut down completely:

 

[BitMaster]

It's all about bandwidth, no one would care about any CL value if the MHz were there to compensate it. Actually, the way it is handled shows that increasing CL values are accepted over generations of RAM, always increasing. The actual latency presented to the system is not the RAM's CL value but the outcome of MHz / CL.

When more RAM is needed, nothing can solve it but more RAM. As things are, you need the "mo RAM" in the same time as you did the with a lesser amount of RAM before, i.e. a texture got 4x as large as before because you upped the LOD, then bandwidth is your friend when you look aside and load that texture first time in game. The higher the fps the faster the new texture has to be processed/loaded etc..   a rat's tail performance wise. Be it system RAM, VRAM, SSD's.

 As long as the outcome of MHz/CL is acceptable, more bandwidth is the way forward.

My rule of thumb over the years for a responsive gaming rig is to stay below a value of 1 for actual used RAM/Bandwidth. My DCS needs usually anywhere 30-40GB mostly and I have ~60GBsec bandwidth at a good latency of 7.77ns ( 3600 CL14-14-14-34 )  If DCS would suddenly need 100GB of RAM, likely DDR4 was too slow to handle that while maintaining acceptable fps and fluid flying. Once it grows outside it's 40-50GB max usage size DDR4 bandwidth may start to struggle and DDR5 systems with twice the bandwidth will likely say "no, I don't have those stutters on my rig". Tune down your DDR4 to 1333MHz early DDR3 speed and see what happens, despite you could run now insane low CL values. 

I hope that with the next coming platforms, after 13th gen Intel and 7000 AMD, DDR5 will have more choices and better pricing, coupled with general availabilty. 

If I had to build one now that should last 4-5 years I'd pick a DDR5 system meanwhile, the prices have fallen enough to consider it and 64GB are still ok for DCS.

 

With newer, more detailed and feature rich modules and maps, hopefully Multi-Threading and Vulkan in the coming years, missions will grow in size and RAM demand,

etc etc. at some point, 64GB will be in any new system and 128GB the standard gamer, 196 or 256GB then the top end rigs.

 

Hey, my 1st one had 2MB RAM. I thought 16MB was heaven for eternity, now I am at 64GB but have understood, this is not gonna stop growing.

 

 

 

Edited March 3, 2023 by BitMaster

[EightyDuce]

1 hour ago, ironhard said:

That's the thing about it. The CL value is not an absolute number. It's comparing apples and oranges.

RL = CL *2000 / DR

Where:

RL = Real latency in nanoseconds

DR = Data rate in Mhz

So, for a 3000Mhz CL16 vs 6000Mhz CL30:

3000Mhz CL16 = 10.66667 nanoseconds

6000Mhz CL30 = 10 nanoseconds

 

https://www.teamgroupinc.com/en/blogs/memory-latency-en

 

I've given up trying to explain this to some people mostly just reply to their nonsense to correct misinformation and keep it from being propagated.

[Hoirtel]

Finally some VR benchmarking. Interesting...

[Greekbull]

I think that is not a very good comparison video for anything we care about.  None of those benchmarks have anything to do with simming.  Almost done with my build but I've already seen a few comparisons with IL2 in VR that show a massive jump from 7950X to 7950X3D so hopefully the same will be realized in DCS...bet it will...

[Thinder]

Quote

You are right there is no CL16 DDR5. 

What you can't seem to understand that doesn't have as much an impact as you think. Because zen4 X3D and non X3D on DDR5 6000 CL 32/30 are spanking their Zen3 counterparts on DDR4 with CL16/14 RAM due to sheer bandwith. 

What is equivent to B. Die anyway? Hynix A/M dies run circles around Samsung Bdie DDR4 in bandwith. 

That's perhaps why nobody is using Hynix A/M die with Ryzen3 or 3D, bandwidth is not what matters with the Ryzen3, low latency is and if you ask AMD support, they will recommend B.die, 4 X 1 rank Cl14 kits, 3200 or 3600 MHz, not anything with higher latency because those CPUs are not designed for it...

It's not you, your perception of "reality" or bias which dictate their controllers limits and the only reason why RAM manufacturers haven't lowered the latency of DDR5 is because their ships become unstable.

AMD knew that long ago, that's why their CPUs architecture is now designed for larger bandwidth and higher frequencies, it doesn't change the fact that the day they can design a Cl14 DDR5, AMD CPUs will be running faster.

So no, there isn't the equivalent of B.die for DDR5 yet and it certainly holds the CPU performances, especially AMD with their cache which benefit from lower latencies for an equivalent speed, it doesn't take a degree in physics to comprehend that.

And something else, it's not "what I think" it's what it is, those 9 series clock faster, have larger caches but AMD made it clear that caches were designed for one purpose and it's data access time, otherwise said latency.

Edited March 3, 2023 by Thinder

[Hoirtel]

50 minutes ago, Greekbull said:

I think that is not a very good comparison video for anything we care about.  None of those benchmarks have anything to do with simming.  Almost done with my build but I've already seen a few comparisons with IL2 in VR that show a massive jump from 7950X to 7950X3D so hopefully the same will be realized in DCS...bet it will...

It's VR and seems to be the only one, so not the worst way to spend 11 minutes while considering a large purchase. Also did show a scenario where vcache gave a huge help to VR performance, probably similar to what you have seen.

 

Edited March 3, 2023 by Hoirtel

[Rix]

Appreciate the spirit of what you're saying, though its not enough to help in that decision.  Usually find comparison videos for DCS sooner have to wait 

[skypickle]

5 hours ago, Greekbull said:

I think that is not a very good comparison video for anything we care about.  None of those benchmarks have anything to do with simming.  Almost done with my build but I've already seen a few comparisons with IL2 in VR that show a massive jump from 7950X to 7950X3D so hopefully the same will be realized in DCS...bet it will...

bravo palikari

[Hoirtel]

6 hours ago, Rix said:

Appreciate the spirit of what you're saying, though its not enough to help in that decision.  Usually find comparison videos for DCS sooner have to wait 

Fair enough, I'm interested in learning about the technology, not just DCS performance. 

[EightyDuce]

13 hours ago, Thinder said:

That's perhaps why nobody is using Hynix A/M die with Ryzen3 or 3D, bandwidth is not what matters with the Ryzen3, low latency is and if you ask AMD support, they will recommend B.die, 4 X 1 rank Cl14 kits, 3200 or 3600 MHz, not anything with higher latency because those CPUs are not designed for it...

It's not you, your perception of "reality" or bias which dictate their controllers limits and the only reason why RAM manufacturers haven't lowered the latency of DDR5 is because their ships become unstable.

AMD knew that long ago, that's why their CPUs architecture is now designed for larger bandwidth and higher frequencies, it doesn't change the fact that the day they can design a Cl14 DDR5, AMD CPUs will be running faster.

So no, there isn't the equivalent of B.die for DDR5 yet and it certainly holds the CPU performances, especially AMD with their cache which benefit from lower latencies for an equivalent speed, it doesn't take a degree in physics to comprehend that.

And something else, it's not "what I think" it's what it is, those 9 series clock faster, have larger caches but AMD made it clear that caches were designed for one purpose and it's data access time, otherwise said latency.

 

 

Honestly, I'm not entirely sure what point you were trying to get across. There was a lot of writing, but nothing of apparent substance that's not already commonly understood by most. 

Quote

That's perhaps why nobody is using Hynix A/M die with Ryzen3 or 3D, bandwidth is not what matters with the Ryzen3, low latency is and if you ask AMD support, they will recommend B.die, 4 X 1 rank Cl14 kits, 3200 or 3600 MHz, not anything with higher latency because those CPUs are not designed for it...

A, M, B die are all just revisions and not some alphabet magic that gives memory speed. Contrary to your belief, people do use Hynix A (AFR) die based DDR4 with Zen3 as it is able to hit 3600, but the timings aren't as tight as Samsung. Samsung's revision (B in this case) is the best chip for DDR4 (Zen3), followed by Samsung E. Die, followed by Hynix CRJ and AFR. In DDR5, Samsung is inferior to Hynix. 

 

Quote

It's not you, your perception of "reality" or bias which dictate their controllers limits and the only reason why RAM manufacturers haven't lowered the latency of DDR5 is because their ships become unstable.

AMD knew that long ago, that's why their CPUs architecture is now designed for larger bandwidth and higher frequencies, it doesn't change the fact that the day they can design a Cl14 DDR5, AMD CPUs will be running faster.

I'm not privy to AMDs design decision making, neither are you,  not sure if either of us could speak to what drove the design of Zen4. Zen 4 was designed around the same time as Alder Lake and Raptor Lake, so the same RAM was available to both, yet Intel MC can handle higher speeds.... Food for thought. 

And yes, when lower latency DDR5 is available (not sure what makes you think it will be able to get down as low as CL14 as there's no guarantee), it will be better than what's available now... Think that's pretty obvious. 

 

At the end of the day, the existence (or lack there of) of what you call "B. Die equivalent" is irrelevant. Equivelant Zen4 with DDR5 6000 outperforms Zen3 on DDR4 3600 CL14.

DDR4 is old tech at this point. It's not going to get any better. Is it still relevant for those on DDR4 systems? Absolutely and one should get the best for those systems.

DDR4 is at the end of its life cycle...DDR5 Zen4 and RL performance is already (right now, today!) better than what's available on DDR4...Time to join us in 2023 and let DDR4 go. 

[EightyDuce]

Some good discussion on the scheduling "hurdles" of the dual CCD X3D CPUs. 

 

[Th3ChosenOn3]

Alright, I've gone ahead and benchmarked my current system with CapFrameX to use as a baseline comparison for the 7950X3D upgrade I'll be installing later this week. I should also be able to use the baseline to see how much multithreading improves FPS when it gets released.

Looking at the current results, I'm pretty much CPU bottlenecked throughout most of the runs on all resolutions, hench no change in FPS by increasing resolution:

Won't be able to test VR improvements sadly, as I don't own a VR headset.

[trevoC]

For those interested in VR I've started a thread where I've so far run a baseline on a 5900x/3090 system.

Just waiting for the cooler for a 7900x3D/4090 system and will re-run the benchmarks sometime this coming week.

I'm not running any non-VR benchmarks so these threads seem to complement each other.

 

 

I'm also CPU limited on the 5900x/3090 setup.

[okletsgo]

@Th3ChosenOn3

Nice. I will try this too. I'll and post my result.

I am currently running a 5800x3D/4080/64GB-3600 CL16.

Can you post your graphics settings?

I have a 7950x3d sitting in the box waiting for my water block to arrive, I also have a 4090 and 64GB DDR5 CL30 ready to go.

 

Edited March 6, 2023 by okletsgo

[Th3ChosenOn3]

50 minutes ago, okletsgo said:

@Th3ChosenOn3

Nice. I will try this too. I'll and post my result.

I am currently running a 5800x3D/4080/64GB-3600 CL16.

Can you post your graphics settings?

I have a 7950x3d sitting in the box waiting for my water block to arrive, I also have a 4090 and 64GB DDR5 CL30 ready to go.

 

 

Nice, would be interesting to see how a 5800X3D compares.

I'm running the default High Preset, with full screen enabled:

Here is the track file that I used. I start CapFrameX capture the moment I press fly with a 1 second delay, and then I let it run for 222 seconds to have consistency between all of the runs. 

Plazma Torture Map - Level Flight.trk

[okletsgo]

@Th3ChosenOn3

Thanks - I have it installed and it is working. How do you create your nice plots? I am currently on a 5800x/3080/32GB-3600CL16 but I can move to the 5800x3d PC after.

Attached is what I got at 4k, does that look correct? 

 

Edited March 6, 2023 by okletsgo

[Th3ChosenOn3]

38 minutes ago, okletsgo said:

@Th3ChosenOn3

Thanks - I have it installed and it is working. How do you create your nice plots? I am currently on a 5800x/3080/32GB-3600CL16 but I can move to the 5800x3d PC after.

Attached is what I got at 4k, does that look correct? 

 

 

The charts I shared here I made in Google Sheets. You can right click on the charts on CapFrameX and export the data to Excel or Sheets to make nicer charts. CapFrameX can also make the same charts with the comparison tab if you don't want to use sheets/excel.

 

Here is a couple I made in CapFrameX on the comparison tab, although it's a bit limiting and finicky to make it how you would want it:

 

As for the result you got, it looks about right. Might be worth checking that you have Tacview, Shadowplay/Adrenaline Replay, Reshade and any background running processes disabled, as I saw it can lower FPS by about 10%. 

Edited March 6, 2023 by Th3ChosenOn3
Formatting

[okletsgo]

@Th3ChosenOn3

Does anyone want to doubt the 5800x3D?    I am surprised myself with the results.

This is a 5800x3D/4080/64GB 2x32GB 3600 CL16 on a custom loop. The system is at stock settings, no curve optimizer and the 4080 is on the latest drivers and also at stock.

Test was done on the current OB release. 4k resolution at HIGH pre-set. Nothing running in the background. The GPU avg utilisation was 98% - I might be GPU limited.

 

Edited March 6, 2023 by okletsgo

[Th3ChosenOn3]

29 minutes ago, okletsgo said:

@Th3ChosenOn3

Does anyone want to doubt the 5800x3D?    I am surprised myself with the results.

This is a 5800x3D/4080/64GB 2x32GB 3600 CL16 on a custom loop. The system is at stock settings, no curve optimizer and the 4080 is on the latest drivers and also at stock.

Test was done on the current OB release. 4k resolution at HIGH pre-set. Nothing running in the background. The GPU avg utilisation was 98% - I might be GPU limited.

 

 

Very impressive.

Compared to my 5800X it's about 20-30% higher FPS, and the frame time graph average is significantly smoother below 10ms. Might be worth it to do a quick run at 1920x1080 to see the max the CPU can do.

[okletsgo]

I just tried it at 1080p.....I get the same numbers. No improvement. I guess I am still not GPU limited.

For anyone out there with a different CPU, ideally a Ryzen 7000 or an Intel 12th/13th gen, please run the benchmark and post your results, it would be very interesting.

 

 

[dureiken]

how did you do those awesome graphs ?

 

would like to try your torture test with 12600k thanks

[some1]

Here's instructions: https://forum.dcs.world/topic/318603-7800x3d-7900x3d-7950x3d/?do=findComment&comment=5166464

 

And here's my 5800X3D + 4090 + 64GB 3600MHz CL16 RAM, at 4k High preset. Not sure where this huge spike comes at the beginning, maybe because I have HAGS enabled, which normally I would disable for DCS.

 

[dureiken]

So nice tool thanks ! didn't know this one

12600k (+64GB DDR4 + 4090) HIGH 1920*1080 fullscreen on plazma

 

almost same in 3440x1400

far beyond 5800x3D

Edited March 6, 2023 by dureiken

[Thinder]

4 hours ago, Th3ChosenOn3 said:

Nice, would be interesting to see how a 5800X3D compares.

zma Torture Map - Level Flight.trk 623.2 kB · 9 downloads

If you run it on Cl16 RAM, it won't actually be as fast as it can especially at 4K, that's what the problem is with AMD Zen3, they are almost never bounded to the RAM that actually make them work properly.