Kaby Lake to Zen 3: Few Performance Observations

[Andrei]

Preface

They say now is a bad time to do an upgrade given the situation with prices and stock. But having spent dozens of hours reading reviews and analytics and found 5800X at a quite reasonable price, I decided to go for it. 

 

The main reason was that I fly quite a lot of dynamic campaigns (like DCS Liberation and DCE) which are notoriously heavy on CPU and my 7700K did not fare too well in those. It was very noticeable when putting mission on pause - the FPS grew dramatically which was a prime indication that CPU is a bottleneck. The other reason was that I didn't win a silicon lottery with Kaby Lake and while it did run initially at 4,9GHz, the temps were high and the crystal eventually degraded to the point where it was barely holding 4,8 if I didn't stress it with all core workout. 

 

Methodology

Methodology is a far fetched word in this case. My goal was to see what kind of improvement would I get with this upgrade. So this is not a comparison of stock to stock. I've spent about a day fine-tuning 5800X to get more juice out of it - same what I did with 7700K. The end-result is a bit of a WIP and the system is still to go though a thorough stability test. 

 

The memory and video card are the same and running with same configuration. Current RAM is not optimal for Zen 3, but I haven't got the new one yet. The drivers and DCS version are same as well. So we're comparing apples to apples in this case. Additionally there are some Cinebench and MSFS comparisons for a bit more perspective. 

 

The mission used for DCS testing is supposed to load the CPU. It has quite a few units and I'm using my VR graphics settings for both 2D and VR test. 

2D stats are captured using MSI afterburner. 3D via FCAT.

 

Results

- "Shut up already and show me the graphs!" - Sure.

 

Starting off with Cinebench R20 to check the generational leaps. Zen 3 is much faster in single core than 7700K. With added overclocking and undervolting, the results are pretty good. 

 

 

DCS in 2D follows and we can see a healthy increase across the board, considering I only changed CPU and motherboard. You get almost 15 extra FPS just by having a better CPU. For the reference: when the track is paused, the FPS goes to 120+ so we're definitely CPU bound here.

 

 

DCS VR shows slightly different picture. You see virtually no difference in either frametime or unconstrained FPS. However, this is expected because FCAT gets those from the video card. Given that it's the same RTX 2070S, no big wonder they are nearly identical. The Delivered FPS is what you see in game, however. And here we get a healthy ~30% increase. What's more important - the subjective smoothness in DCE missions has notably improved. What was barely playable is now quite OK. Not perfect, but totally flyable. 

 

 

Finally an example from the other side - MSFS. This just validates findings in DCS with the FPS improvement across the board. Though the gains may not be so great compared to DCS.

 

 

Conclusion

I am quite happy with this upgrade. I've been planning it since December 2020 and there were always issues either with stock or price or just waiting for something better (hello, Rocket Late!) to come by. The outcome was well in line with my expectations. I'll need to spend a bit more time tuning 5800X and wait for the faster RAM to be delivered - which should help with the performance even more. 

 

Thank you for taking your time to read this.

 

Test Rigs

Intel i7 7700K @ 4,8GHz

Asus Prime Z270-A

4x8GB DDR-3200 CL16

GeForce RTX 2070S @ +140 Core + 250 Memory

 

AMD R7 5800X @ ~5GHz Single Core max boost and 4.6GHz All Core

Aorus B550 Pro V2

4x8GB DDR-3200 CL16

GeForce RTX 2070S @ +140 Core + 250 Memory

[Mars Exulte]

Excellent testing amd presentation.

 

Also, that cinches it for me. I've been wishy washy about whether it was worth doing an upgrade for me (Skylake 6700k @ 4.6/4.7). Seems like it's apt to be significant even if everything else remains the same. Good bye old and busted, hello new hotness!

[Sn8ke_iis]

Finally someone has done a real benchmark! Thank you! This should be the standard from now on. I would have posted basic Excel graphs. 

 

Just one fine point. You definitely didn't win the SL with your 7700K. That's why I buy from Silicon Lottery, it's worth the premium. Speculating but if you could keep 5.0+ stable I don't think you would have seen the delta on your 2D data. 

 

The 7700Ks were actually really good overclockers, but you needed to delid and change out the stock thermal compound. Mine hit 5.1 all core easily. Single was at 5.3 Ghz. DCS will run just fine on 3 Skylake cores. I've tested it myself my restricting cores in BIOS. No difference in frame rate between 3 and 4+ cores. 

 

https://siliconlottery.com/pages/statistics

 

Now, if you can just buy a stock AMD CPU and it's IPC and clock speed are good enough. That speaks volumes. Most PC gamers don't overclock or delid. 

 

How did you tune your AMD CPU, memory overclocking? 

 

Either way great benchmark. Please post more data if you have it. 

 

From the data I've seen the 5900X seems to be binned with the best single core performance which matters most for games and sims. I'm only seeing it scalped currently. 

 

Oh yeah, my chip hits about 620 on the single thread score in Cinebench when properly OC'd. I was hoping Rocket Lake would be a significant jump to upgrade but alas, the data on 11900Ks is not very good. A binned and overclocked 10900K is probably the fastest for gaming right now per Gamer's Nexus benchmarks. 

Edited April 3, 2021 by Sn8ke_iis

[Andrei]

I was thinking about delidding 7700K and giving it more run for it's money, but I should have done yet years ago - now it was time to get a new hardware anyway. Your 9900K is a pretty good piece, probably the best thing in the whole *Lake family  

 

In terms of tuning I'm trying to play around with undervolting and PBO boots. The latest AMD CPUs are being overclocked the "wrong" way  I can get pretty good frequecies - upwards of 5GHz, but stability becomes an issue. Enough to boot and pass all Cinebench tests, but then it will crash on random things, especially when cold starting. So it's a moving objective. I don't care too much about all core loads, so I need to maximize it for DCS. I've got a really bad memory for this build it runs only 3200, while you need 3600 - 4000 for Zen 3 for best results. The new memory is ordered, but due to general stock issues, I've no clue when it will come. 

[Thinder]

On 4/3/2021 at 6:55 PM, Andrei said:

 I've got a really bad memory for this build it runs only 3200, while you need 3600 - 4000 for Zen 3 for best results. The new memory is ordered, but due to general stock issues, I've no clue when it will come. 

 

RAM frequency is not the issue, Latency is.

 

Looks like people doesn't get how Ryzen 3 works just yet despite the fact that AMD recommanded frequency is 3200GHz  and not 3600/4000 for a good reason, higher speed doesn't have the single bank/lower latency offered by the 3200GHz kits like this G.SKILL TridentZ RGB Series 32GB (4 x 8GB) 288-Pin DDR4 SDRAM DDR4 3200 (PC4 25600)  F4-3200C14Q 32GTZR.

 

Having said that, G-Skill offers very specific kits with increased frequency for the Ryzen 3 but they are specifically designed for those who wants higher frequencies, they do not specify if they actually offer more speed, which is not aquired at all, especially because the faster 3200GHz single bank kits are in short supply due to the limited production of their B-Die chips.

https://www.newegg.com/global/uk-en/g-skill-32gb-288-pin-ddr4-sdram/p/N82E16820232488?Item=N82E16820232488

 

If you want a fast RAM to CPU bus operation, you'll need to bound your RAM to the Ryzen properly, which means CL 14 and 4 X single bank which offers the best results simply because it makes use of the CPU low latency and Rank interleave, while a double bank is going to cause isssues to the Ryzen controller even more so at higher frequencies, reason for AMD 3200GHz recommandation.

 

In short, you use higher frequency than 3200GHz and/or dual rank, you loose low latency and bank interleave, it has to be done for a very specific reason, not for the stake of using high frequencies only.

 

I got some serious increase in speed just changing my RAM from a 2 X CL 16 3200 dual bank, to a 4 X CL 14 single bank and testing the first kit up to 3600 didn't improve things at all.

 

Just for your info, I submited my results to MSI, AMD and G-Skill, all were impressed by the results, because the combo never was tested with the Ryzen 3/5600X and is not listed as compatible RAM for my motherboard, I figured it out roaming AMD users forums and was suggested this RAM kit by one of the top users, it works wonder.

 

Additional info, with Ryzen controllers, a bad RAM bound makes the CPU throttle down significantly because it struggle to cope with the higher frequency and number of ranks, with this kit, I gained an increase of 6.04% in CPU speed.

Edited April 13, 2021 by Thinder