News Cyberpunk 2077 Memory Tested: How Much RAM Do You Need?

VforV

Respectable
BANNED
Oct 9, 2019
578
287
2,270
Are you guys aware that CP 77 does not properly use CPUs, RAM and Vram on PC?

There is a mod that fixes both CPU SMT and memory allocation issues, as well as other issues.

Take a look > https://www.nexusmods.com/cyberpunk2077/mods/107?tab=description

While it's nice to have this test of yours, it's pretty much usless at this point if the game has so many issues and is not working properly with the hardware it's given, not even on PC.
 
  • Like
Reactions: Geef
Are you guys aware that CP 77 does not properly use CPUs, RAM and Vram on PC?

There is a mod that fixes both CPU SMT and memory allocation issues, as well as other issues.

Take a look > https://www.nexusmods.com/cyberpunk2077/mods/107?tab=description

While it's nice to have this test of yours, it's pretty much usless at this point if the game has so many issues and is not working properly with the hardware it's given, not even on PC.
I think it's worth still keeping the testing with a more or less out of the box configuration. It's what most people are going to be running with anyway.

Otherwise this could snowball into something like we may as well invalidate every test that doesn't include highly tuned "performance" tweaks because the software isn't "running properly"
 

salgado18

Distinguished
Feb 12, 2007
928
373
19,370
It needs to be measured on a drive by through the city, where asset loading is more requested. The results here are still interesting, but there is another aspect of this game, which is real-time loading of stuff.

Also, are you guys planning on testing HDD vs SATA SSD vs mid-range NVME vs PCIe 4.0 NVME? If so, test it driving! That's the most demanding scenario. And if you have the time, merge this with RAM testing, can one compensate the other?

Oh, and I wish you all infinite budget and time to do all the testing ;)
 

San Pedro

Distinguished
Jul 16, 2007
1,286
12
19,295
Are you guys aware that CP 77 does not properly use CPUs, RAM and Vram on PC?

There is a mod that fixes both CPU SMT and memory allocation issues, as well as other issues.

Take a look > https://www.nexusmods.com/cyberpunk2077/mods/107?tab=description

While it's nice to have this test of yours, it's pretty much usless at this point if the game has so many issues and is not working properly with the hardware it's given, not even on PC.

They literally have an article that went up yesterday about it. It was on the front page.
 
Dec 14, 2020
3
7
15
Jesus Christ Jarred, you aren't really testing performance scaling with memory capacity here. All of the 32 GB configurations here are dual rank (4x8 GB and 2x16 GB). The rank interleaving on these is providing a massive boost in performance that is completely unrelated to capacity. The ONLY test on capacity here is the 2x8 GB kit vs the 2x4 GB kit at the same timings because they are both single rank and even that produced a miniscule performance difference. Then you go and say "Capacity clearly starts to matter more with ray tracing, and the 32GB DDR4-4000 kit is about 9 percent faster than the same kit with only 16GB (two DIMMs)." That's ridiculous. The performance CLEARLY comes from the rank interleaving. want more proof? Grab a 2x16 GB kit that's single rank and test with that. They exist, and I'm pretty sure they're quite common in Crucial products right now. You've effectively mislead anyone who ends up buying a single rank 32 GB kit with this false explanation of capacity being the cause of the performance increase. Terrible article at best. I'd suggest it gets pulled or heavily modified.
 
Jesus Christ Jarred, you aren't really testing performance scaling with memory capacity here. All of the 32 GB configurations here are dual rank (4x8 GB and 2x16 GB). The rank interleaving on these is providing a massive boost in performance that is completely unrelated to capacity. The ONLY test on capacity here is the 2x8 GB kit vs the 2x4 GB kit at the same timings because they are both single rank and even that produced a miniscule performance difference. Then you go and say "Capacity clearly starts to matter more with ray tracing, and the 32GB DDR4-4000 kit is about 9 percent faster than the same kit with only 16GB (two DIMMs)." That's ridiculous. The performance CLEARLY comes from the rank interleaving. want more proof? Grab a 2x16 GB kit that's single rank and test with that. They exist, and I'm pretty sure they're quite common in Crucial products right now. You've effectively mislead anyone who ends up buying a single rank 32 GB kit with this false explanation of capacity being the cause of the performance increase. Terrible article at best. I'd suggest it gets pulled or heavily modified.
Dual rank does not make that much of a difference, sorry. It's a few percent typically. And considering you generally don't get single rank 16GB DIMMs, unless you explicitly seek them out (and why would you?), you're the one jumping to conclusions. I said the memory was single rank (technically I said bank, but I've edited it now -- blame the lack of sleep) on the lower configs and dual rank on the 32GB. Not much more you can do about that.
 
Dec 14, 2020
3
7
15
Common misconception and other sources disagree. It depends on the game! But one thing for certain is that it will be a much larger difference than changing capacity.

View: https://www.youtube.com/watch?v=-UkGu6A-6sQ

View: https://www.youtube.com/watch?v=25NW8cHNrgA


I love how you completely disregard the miniscule difference between your 8 GB vs 16 GB setup, but chalk up a 9% difference to a change from 16 GB to 32 GB with other factors in play.

Edit: Getting a 404 on the article now, so maybe the message came through shrug
 

danlw

Distinguished
Feb 22, 2009
137
22
18,695
It needs to be measured on a drive by through the city, where asset loading is more requested. The results here are still interesting, but there is another aspect of this game, which is real-time loading of stuff.

Also, are you guys planning on testing HDD vs SATA SSD vs mid-range NVME vs PCIe 4.0 NVME? If so, test it driving! That's the most demanding scenario. And if you have the time, merge this with RAM testing, can one compensate the other?

Oh, and I wish you all infinite budget and time to do all the testing ;)

This is spot on.

Unfortunately, its hard to write an article where hard numbers are not present. This would be a measure of user experience.

I did a similar test to the one you recommend a while ago in Star Citizen and found that when running with an HDD or when running with less memory, there would be frequent pauses while the game streamed in more assets. These pauses disappeared with an SSD and sufficient RAM.

So definitely, there should be a metric to measure player experience. It is not all about FPS. I could be playing at 240fps, but if the game is constantly freezing to load assets, Im not having a good time.
 
This is spot on.

Unfortunately, its hard to write an article where hard numbers are not present. This would be a measure of user experience.

I did a similar test to the one you recommend a while ago in Star Citizen and found that when running with an HDD or when running with less memory, there would be frequent pauses while the game streamed in more assets. These pauses disappeared with an SSD and sufficient RAM.

So definitely, there should be a metric to measure player experience. It is not all about FPS. I could be playing at 240fps, but if the game is constantly freezing to load assets, Im not having a good time.
Two things to consider.

First, I've provided percentile frametimes from 50 percentile (half of frames hit this speed or higher) up through 99.9th percentile (the slowest 0.1% of frame rates). This is a sensible way to guage the whether there's a lot of stutter, particularly if you focus on the 97th, 98th, and 99th percentiles. A few 'bad' frames can impact the 99.9th percentile, but if you see lots of dips showing up in the 97th that would indicate clear stuttering. In other words, the charts already cover this aspect if you look at the percentiles.

Second, driving in Cyberpunk 2077 is not really a good measure of how the game plays. Yes, you can drive around. No, it's not really the heart of the game. Driving is purely to get from point A to point B most of the time. You can't shoot out the windows while driving (unless it's a scripted sequence, AFAIK), so your real interaction with the game world -- ie, when performance and latency matter most -- will be when you're not in a vehicle. This isn't a racing game, and the driving isn't even that good (IMO, naturally). It was added later on in development I'm pretty sure. But, I did do some testing of vehicle performance just to check with one of the GPUs. I got in a car and drove mostly in one direction (ie, I didn't drive in a loop). Performance from what I saw was actually about 10% higher than my walking test. Are there areas in the game that perform worse than what I showed? Almost certainly, just as there are areas that will perform better. What I've seen of Night City however suggests that a crowded city street during the day (which is what was used in our benchmark) is reasonably close to the lower end of the performance spectrum.

Would the game run worse off a hard drive, with less memory? Absolutely. But I ditched HDD storage about five years ago and never looked back. These days, 1TB of SSD storage costs less than $100 for SATA, and that's usually more than sufficient. Even fast M.2 drives only cost about $110-$130 for 1TB. If you want to play Cyberpunk 2077 on an older PC, with a hard drive, 8GB RAM, and a four year old GPU, I'm sure you can do so. It won't be as smooth, but that's the case with most open world games running on that sort of hardware.
 

VforV

Respectable
BANNED
Oct 9, 2019
578
287
2,270
I think it's worth still keeping the testing with a more or less out of the box configuration. It's what most people are going to be running with anyway.

Otherwise this could snowball into something like we may as well invalidate every test that doesn't include highly tuned "performance" tweaks because the software isn't "running properly"
You can bet that this test will be a lot different in 6 months time after some serious patches. Not to mention 3rd party Fixes.
They literally have an article that went up yesterday about it. It was on the front page.
Two things:
  1. I know about the article and it only shows the CPU fix, not the RAM and Vram fix.
  2. They don't use it in this test and they should, or they don't specify if they use it - which is as bad.
For most people that have 8GB or 16GB RAM is much easier to just use the mentioned FIX and get a big performance boost than buying a new kit of ram and marking an upgrade.

I would even go so far as, you don't even need more that 16GB RAM if/when (with the FIX) the game uses the correct spcs of your system.

Also yes like Im_A_Decoy said, the major difference in this test comes from dual rank vs single, not because of RAM size - at least not in the case of 16 vs 32.
Dual rank does not make that much of a difference, sorry. It's a few percent typically. And considering you generally don't get single rank 16GB DIMMs, unless you explicitly seek them out (and why would you?), you're the one jumping to conclusions. I said the memory was single rank (technically I said bank, but I've edited it now -- blame the lack of sleep) on the lower configs and dual rank on the 32GB. Not much more you can do about that.
Actually dual rank does make the difference here and is clear. Unless you show us that the RAM usage in game exceeds 16GB (game only, not 15 Chrome tabs open in the background), the 32GB kits only have the advantage of the dual rank and that it is shown.
 
Nov 23, 2020
1
0
10
Spot on. You can see the difference with the dual rank kits. Also, it factors in more when you realize that the motherboard (like most Intel chipsets) is dual channel.
 
Spot on. You can see the difference with the dual rank kits. Also, it factors in more when you realize that the motherboard (like most Intel chipsets) is dual channel.
Dual channel has nothing to do with dual rank. A memory rank is just a separate tier of data stored on the same module. Because of the way DRAM works, each rank gets refreshed individually, so having two ranks within a chip means you get a slight speedup -- data can come from the second rank while the first is refreshed, and vice versa. Putting two single rank chips on one channel mostly accomplishes the same thing, though populating two slots on a channel can impact performance negatively in other ways (very small deltas). Which is why the 2x16GB dual rank DIMMs basically perform the same as the 4x8GB single rank configuration.
 
  • Like
Reactions: Anayalator94

r7litepro

Honorable
Nov 5, 2016
35
4
10,535
i know this is an enthusiast site, but i would specifically like to see this game performance using the 1060 GTX 6gb *arguably the most popular GPU in the planet, (also in the official system reqs) under different configurations: with a 5600x, with stronger CPU, with 32000 CL 16 32GB SINGLE RANK kits, with 32gb DUAL RANK KITS, 3600 cl18 KITS, with mitigations disabled, enabled, i mean. go crazy. thanks! btw 32 gb single ranks kits are really common this days. i have one! ( 16GBx2 3600 CL18 single rank)
 
  • Like
Reactions: Im_A_Decoy

Makaveli

Splendid
Jesus Christ Jarred, you aren't really testing performance scaling with memory capacity here. All of the 32 GB configurations here are dual rank (4x8 GB and 2x16 GB). The rank interleaving on these is providing a massive boost in performance that is completely unrelated to capacity. The ONLY test on capacity here is the 2x8 GB kit vs the 2x4 GB kit at the same timings because they are both single rank and even that produced a miniscule performance difference. Then you go and say "Capacity clearly starts to matter more with ray tracing, and the 32GB DDR4-4000 kit is about 9 percent faster than the same kit with only 16GB (two DIMMs)." That's ridiculous. The performance CLEARLY comes from the rank interleaving. want more proof? Grab a 2x16 GB kit that's single rank and test with that. They exist, and I'm pretty sure they're quite common in Crucial products right now. You've effectively mislead anyone who ends up buying a single rank 32 GB kit with this false explanation of capacity being the cause of the performance increase. Terrible article at best. I'd suggest it gets pulled or heavily modified.

Have to agree with this.

The only thing this articles shows is you want a Dual Rank set to get every bit of performance out your memory. And there are articles on this site showing that as well as other websites.
 
Dec 14, 2020
3
7
15
Jarred, I just wanted to say thanks for appending that info into the article. I really appreciate it and would like to apologize for what likely came across as hostility.
 
  • Like
Reactions: CodgerFace
Appreciate the extra effort to update the article with the feedback here, Jarred. 👍
The best part is that the tweaked text was done four days ago, but people never actually read the whole thing. They looked at the headline and got offended that the SEO made it sound more interesting than it is. I originally wrote "Cyberpunk 2077 Memory Benchmarks" as the headline, but it would have likely done way less traffic that way. The point wasn't to prove whether dual rank vs. single rank vs. capacity alone are factors, which they obviously are, but to show how much of an impact various RAM configurations can have on performance in Cyberpunk 2077. As with all benchmarks, for every configuration tested there's a virtually unlimited number of configurations that weren't tested (due to time and hardware on hand). But, you know what they say about gift horses. 🐴
 

danlw

Distinguished
Feb 22, 2009
137
22
18,695
Two things to consider.

First, I've provided percentile frametimes from 50 percentile (half of frames hit this speed or higher) up through 99.9th percentile (the slowest 0.1% of frame rates). This is a sensible way to guage the whether there's a lot of stutter, particularly if you focus on the 97th, 98th, and 99th percentiles. A few 'bad' frames can impact the 99.9th percentile, but if you see lots of dips showing up in the 97th that would indicate clear stuttering. In other words, the charts already cover this aspect if you look at the percentiles.

Thanks, admittedly I wasn't sure what to do with the percentiles, so I did some reading... let me see if I understand this correctly now.

On the charts, there is the line graph up to 99.9th. Does that mean that 99.9% of the time the fps is higher than this number? And likewise at 50%, half the time the fps will be higher, half the time they will be lower?

So theoretically if I sampled fps in a game for 10 seconds and the game played at a solid 100fps for 9 seconds and then froze for 1 second, my 90th percentile would be 100fps but my 99th percentile would be 0fps. Therefore if a game is stuttering, there should be a very large dip in the highest percentiles.

So in this article:
https://www.tomshardware.com/news/amd-radeon-rx-6900-xt-review

Am I correct to infer that in the Boundary benchmark, the AMD cards suffered from significant stuttering in the 1920x1080 tests as displayed by the sharp dip at the 99.9th percentile?
 
Thanks, admittedly I wasn't sure what to do with the percentiles, so I did some reading... let me see if I understand this correctly now.

On the charts, there is the line graph up to 99.9th. Does that mean that 99.9% of the time the fps is higher than this number? And likewise at 50%, half the time the fps will be higher, half the time they will be lower?

So theoretically if I sampled fps in a game for 10 seconds and the game played at a solid 100fps for 9 seconds and then froze for 1 second, my 90th percentile would be 100fps but my 99th percentile would be 0fps. Therefore if a game is stuttering, there should be a very large dip in the highest percentiles.

So in this article:
https://www.tomshardware.com/news/amd-radeon-rx-6900-xt-review

Am I correct to infer that in the Boundary benchmark, the AMD cards suffered from significant stuttering in the 1920x1080 tests as displayed by the sharp dip at the 99.9th percentile?
So technically the percentiles are based on the number of frames that are above a threshold. If I do a test that renders at an average of 60 FPS for 60 seconds, that’s 3600 frames. The 99th percentile means 1 percent (36 frames) fell below that mark, and the 99.9th is 0.1 percent (4 frames).

By pure time, then, more than 1 percent of the sample period would be at less than the 99th percentile. How much more would depend on other factors but isn’t super critical.

The main point is that, if you’re seeing more than 1 percent of frames below 30 FPS, that will definitely show up as stuttering. I also find that games where the 99th percentile is less than about two-thirds of the average FPS tend to feel like there’s microstutter, though it’s important to look at the particular test. (The FFXIV test has a lot of scene transitions that drop the 99th percentile. In the actual game, it’s a bit less of a factor, though some of the spell effects can definitely tank performance.)

I’d have to run some actual calculations on some benchmarks to give you exact time amounts that a game spends under a specific FPS, but rough estimate: if a game has a 99th percentile of 60 FPS, probably 1 second out of every 60 seconds will run at less than 60 FPS. How much less could be estimated by the 99.9th percentile.
 
  • Like
Reactions: danlw
Nov 17, 2020
11
0
10
I played for about 10 hours and began to notice that the game does not under any circumstances want to take more than 8 Gb (8gb and the ceiling, or a megabyte more) of RAM from the system, although it can and should do so.