Hello, this is Soul Seeker. Today, we are going to take a look at the topic written on the screen. When you do a full bank in DDR5, it is known that the clock speed decreases. Many people now know that when you do a full bank, the clock speed decreases. In fact, even in DDR4, there was a decrease in clock speed when doing a full bank, and the clock speed was slightly reduced. However, in DDR4, the standard clock speed of around 3200 was maintained to some extent, so there wasn't much of an issue with full banks. But as we transition to DDR5, the clock speed has increased significantly. Although the clock speed has increased, if you do a full bank, it operates at a clock speed that cannot even keep up with the standard clock speed of around 4800. Therefore, people have come to realize that when you do a full bank, the performance decreases. That's why in DDR5, full banks are not recommended.

If we talk about it accurately in terms of FM, there would be a 16GB module and a 32GB module. There is also an 8GB, but let's put that aside. So, there would be 16GB and 32GB modules, and typically, using two 16GB modules provides the best clock speed. When we say the clock speed enters well, it means that in cases where the clock does not enter, there may be a problem with clock degradation. So, if you use two 32GB modules or four 16GB modules, you will have a 64GB setup. Which one allows the clock to enter better between two 32GB modules and four 16GB modules? Two 32GB modules allow the clock to enter better, followed by four 16GB modules, and then four 32GB modules. Recently, 24GB and 48GB modules have been included. If we list them in order, the clock speed entering would be like this: multiply by 2 here, and then full bank like this - four 24GB modules and four 48GB modules.

You should see it like this. In fact, the difference between the 16GB and 24GB is minimal. In terms of clock speed, however, the reason why the 16GB is expressed as being slightly higher is that the timing is tighter for the 16GB. As the capacity of the 24GB is relatively higher, it loosens slightly. Similarly, there is a difference between the 32GB and 48GB. So, when you split them into two, you should see it like this. There is such a difference. As you can see now, those marked as multiplied by 4 mean that they relatively have lower clock speeds. Therefore, manufacturers initially thought, 4800 won't work, because the standard clock speed when DDR5 was first introduced was 4800. They thought, 4800 won't work, so they initially set it to operate at a lower clock speed. Therefore, Intel operates at 4000, and AMD operates at 3600 when in full bank mode.

If we go into this theoretically, it seems like I've uploaded this video before, but this motherboard is called a Daisy Chain structure, so it's called Full Bank. There is a related video that I've uploaded. If you search well, you will find it. Anyway, when the clock drops like this, anyone can easily think that the performance will decrease. So when people ask if the clock drops when they do Full Bank, the usual answer is that the performance drops naturally. If you haven't tried how much the performance drops, I haven't either. So, I tried it this time. In my common sense, when you do Full Bank, if you have DDR5 4800, the standard value for CL is 40, and for 5600, the standard value is 46. And then, when you go to this standard, how is this standard value calculated?

When the clock is set at 4800, the speed per clock is this much. If it's CL 40, then it's 16.6666667ns. So, the speed of processing CL is like this. Now, let's see at 5600. At 5600, it's 16.428ns. There isn't much difference, so this is usually the standard. This is the standard. As you can see from the standard, when the clock decreases, the CL timing decreases. So, when you look at the standard values, if it's 4000 clocks, it becomes CL 32, and if it's 3600, it becomes CL 30. If you calculate this, it's 16, and then 16.6667 as the standard, so it roughly matches this. So, when you consider this, since it's 16ns, actually, CL 32 at 4000 is faster than CL 40 at 4800. Originally, that's how it should be. However, why does performance decrease when the clock speed decreases? In fact, the CL speed doesn't decrease; it gets faster. It gets faster. It's not just about the CL timing. For example, when a specific timing A is 8 at 4800, it's still 8 even at 4000.

Even at 5600, it is still 8. When this happens, there is a difference occurring here. So, at a specific timing A of 4800, if it is 8, then the speed is about 3.3333ns. When it is 8 at 4000 and also at 5600, if we calculate it in the same way, it becomes 4ns and 2.857ns, respectively. Speed differences like this occur. It's not just about CL timing; there are various timings in timings. Among these timings, there are timings that do not change significantly. When the clock changes, speed differences occur. Depending on how often these timings are processed, these speed differences can vary significantly. So, in common sense, when looking only at CL, the CL 32 of 4000 seems faster. But in reality, for AMD processors, many use CL 42 for 5200. When calculated, the CL 42 of 5200 is faster than the CL 40 of 4800, and even faster than 46 of 5600.

When you just look at CL, the performance improvement coming from this CL is not only from CL but also from other timings, especially these fixed timings. Considering that, you should take that into account. So, if you think logically, based on the standard values, the performance seems similar. However, considering how much difference there might be in these similar timing differences, you can think like this. I also thought that way. Thinking like that, I took the opportunity to measure it this time. I actually tested how much it drops. Let me show you the results once. I tested various clocks, and since the memory currently installed is 4800 CL 40, I tested it with Hynix M-die. First, let's look at the M-die test. I ran it about 5 times with AIDA64. The average of running it 5 times is in the yellow cell. When I set the average value of 4800 to 100 and tested it with full bank.

Then, the clock automatically shows 4533, not 4000. It seems like the latest BIOS standards are applied, probably due to the Auto Booster. The clock did not drop significantly to 4533, but decreased by about 266 clocks. What is this? I didn't expect this. The timing remains the same, right? It's the same above and below. Next, let's take a look at these detailed timings. RRDS, RRDL, FAW are all the same. Even RFC2 and RFCPB are the same, almost everything is the same. The difference is when you go further down, there are differences depending on the clock. For example, these timings are now influenced by the clock. These timings and these timings, you've probably never seen these timings before. There are timings like this, which are influenced by the clock, and timings that are not influenced by the clock, are all the same.

The timing isn't automatically tightened. Maybe it's because it's Intel, I'm not sure if it would be the same on AMD, but when I installed the latest BIOS on the 14th generation CPU from Intel, this issue occurred. So if the timing isn't automatically tightened, it will be suboptimal, right? I thought the CL was 32 at 4000 clock speed, but that's the standard, but the standard timing is not showing up. I tried lowering the clock speed to 4000, but the timing remains the same. If you look at this, it's almost the same. Even at 4000, the timing is almost the same. What this means is that if the timing isn't tightened, performance will decrease when the clock speed drops. Look at the latency, it came out to an average of 83.8. Focus mainly on latency. Anyway, the read, write, copy values ​​are highly dependent on the clock speed. If the clock speed is high, they are high, and if it's low, they are low. For example, these read, write, copy values are

When reading, copying, or writing small miscellaneous files continuously, it is efficient and actually latency-focused in terms of memory performance. So, if you look at it, even though the clock was only lowered to 266 with the Auto Booster, the performance decreased by about 6.3%, dropping from 100% to 93.7%. This resulted in a performance decrease of about 6.3%. In practical terms, if it drops down to 4000 clocks, the situation worsens. Performance dropped to 84.7%, a 15.3% decrease. Therefore, I tried adjusting to the standard value CL 32. Of course, I left other timings on Auto. Since only the 1st timing is specified in the standard value, when I adjusted only the 1st timing to 32, the performance almost rose back by nearly 10%, from 84.7% to 94.1%.
(The values calculated based on 4800, so with actual 4000 clocks, there is an improvement of over 10% in performance.)

So, even if you just keep an eye on the CL RCD RP timing standard values, it's not like there is no performance degradation at all. It's true that there is a performance drop of about 5.9% to 6% compared to 100%. Currently, AMD has not been verified, and it seems that Intel is maintaining the timings in the default state as they are. I haven't tested this with 5600 RAM, but with 4800 RAM, the CL RCD RP timing remains unchanged. So, this is quite critical. Therefore, the performance drop during full bank was larger than I expected. Even more than I expected. Next, although I haven't tested an AMD system, I did test it with Intel at 3600. If the standard values are properly applied at 3600, there is about a 11.6% performance drop compared to 4800 CL 40. I don't know how it will be when actually done on an AMD system. If these sub-timings are also applied on an AMD system.

If it is made to be adjusted in proportion to the clock, the performance drop may not be as significant as expected. The performance drop that occurs here, as I mentioned earlier, is not much affected by the CL timing itself. The RCD is the same. This 11.6% performance drop is due to the lack of variation in specific timings. Especially the RFC, it should decrease when the clock drops, but it doesn't. It's a problem caused by these timings not tightening. We may need to reconsider how AMD handles this, but based on Intel's testing, this is the result. For now, let's assume full bank, we often use 4800 RAM, then 5200, and even 5600 RAM. I tested all the way up to that. In the 14th generation of Intel, the reason I didn't expect the timings to be tightened is because...

When we increase the clock like this, these sub-timings automatically adjust a little as you can see. Even the RFC loosens proportionally to the clock when we raise it. It loosens when going up, but why doesn't it tighten when we lower the clock? I really didn't expect that. Anyway, at 5200 and 5600, based on this, I measured the latency five times each and found that on average, 5200 CL 42 was about 2.3% better than 4800 CL 40 RAM. And then, with 5600 RAM, there was about a 5.4% performance increase compared to 4800 CL 40. So, not only did the clock increase, but also these timings, as you can see, all the RRDS are the same at 8, right? This FAW is also the same at 32, and there are timings that do not change. Due to these timings that do not change, performance improvement occurred here when the clock increased.

So now, we have 5600 RAM available, but in the future, 6400 standard RAM will be released. I have tested the setting values once by sampling 6400 RAM. When set to 6400, the standard is CL 52. When only CL, RCD, and RP were set to CL 52, the other timings were set to Auto. This is how the timings were set to Auto, and in this way, the performance increased by as much as 13.2% compared to CL 40 at 4800. It's quite promising. If around CL 52 comes out at 6400, the performance is expected to improve by about 13% compared to CL 40 at 4800. Of course, the performance improvement may not be significant compared to 5600 RAM, but as you can see in terms of latency, it's almost in the mid-70s, right? So, I am looking forward to it. Anyway, I measured the performance changes according to the clock change like this.

One of the unexpected results was the Auto Booster, and then the timing didn't get tighter. In that aspect, a somewhat negative outcome emerged. Since Auto Booster is Hynix memory, it seems to have maintained the clock up to 4533. If it's not Hynix memory, Auto Boost may not occur like this, so it would be good to keep that in mind. Anyway, for those who are going for full bank, it's advisable to be aware that there may be a performance drop of about 6.3% around 93.7% when everything is set to Auto. When everything is set to Auto and full bank is installed, there was a performance drop of about 6.3%, so if you can tolerate that level of drop, you can go for full bank. If you can't tolerate it, there's nothing you can do. But in reality, most people who go for full bank do so because they lack capacity, as mentioned earlier. I will finish the video by mentioning this much.

Even with two 32GB sticks or up to four 16GB sticks, some level of clock speed can be achieved. The ones that can't achieve the clock speed are the issue. They really can't handle the clock speed. Some may not even reach 4800. Therefore, with these differences, since I'm overclocking anyway, I can improve performance to some extent through overclocking. I can boost the performance that has dropped to some extent. For example, even with just 4000 clock speed, as you can see, the CL is only 32. This is the standard value. The standard value is a very safe value. When I apply the standard values for CL and RCD, the performance increased by nearly 10%, right? For example, if I install full banks, it's 128GB or 196GB. So, the clock speed doesn't go up, and it's set at 3600. For example, because it's AMD, it's set at 3600.

Then you just need to adjust the timing. By tightening the timing, you can take care of the performance. Problems occur because the timing is loose, so by adjusting the timing, you can actually get better performance even if the clock is at 3600, exceeding the stock value of 4800. This applies to Intel as well. Of course, ultimately, this falls into the realm of overclocking. Adjusting the timing is also part of the realm of overclocking, so some additional work to ensure stability is necessary. We will go up to that point. In systems where stability is really important, there is no choice but to use it as stock. However, for those who want to try a slight overclock, if you learn about overclocking and adjust the timing, you can achieve a certain level of performance even without significantly increasing the clock. So, I will stop here. Thank you for watching.

[아는것만 알려드림]DDR5 풀뱅크 클럭 저하에 따른 성능 하락 과연 어느 정도 일까?(인텔 기준) / 구글시트 링크 설명탭에 공유

소울시커의 IT 하드웨어에 대한 취미생활을 담은 채널입니다.
비교적 최신 모델의 CPU, Memory에 대한 오버클럭 컨텐츠를 주로 다루고 있고
그외 BIOS 셋팅법, 하드웨어 팁, 각종 문제 해결 및 노하우 등도 소개해 드리고 있습니다.
재미있는 IT 이슈도 있으면 전달해 드리겠습니다.

제휴 및 협약 : soulseeker.officially@gmail.com


[아는것만 알려드림]DDR5 풀뱅크 클럭 저하에 따른 성능 하락 과연 어느 정도 일까?(인텔 기준) / 구글시트 링크 설명탭에 공유

[아는것만 알려드림]DDR5 풀뱅크 클럭 저하에 따른 성능 하락 과연 어느 정도 일까?(인텔 기준) / 구글시트 링크 설명탭에 공유

Scripts