The only reason to have 4 sticks of RAM is because you didn’t buy enough to start with and are trying to save money on an upgrade. There’s an easy solution to that! 32GB is enough, and 64GB is plenty more than enough (Bill Gates can help out here).
I have 64GB in 2 sticks of Corsair 6000MHz. Not fancy stuff, just rated. It works fine at 6000 using the default expo with my Asus Tuf Gaming X670E mobo. However it only works if I let it train every time, if I set it for fast memory restart then it’ll crash under use. The training adds half a minute or so to the boot which is annoying, but compared to the almost 10 minutes MSFS takes its not a problem. I don’t use the PC for anything else, or maybe it would be really annoying. I didn’t check any mobo compatibility list before buying, I just bought medium priced with reasonable CAS latency so YMMV.
I don’t care for Intel/AMD or NVidia/AMD. I just buy what is best at the time, this time it was AMD+NVidia.
Your motherboard is not the issue. The issue is that 4x sticks puts too much stress on the IMC and you’re limited to a lower RAM frequency. If you remove two sticks, you should be able to run at least 6800 and possibly even as high as 7200 MT/s on most 4-DIMM motherboards. Motherboards with 2-DIMM slots (like the Apex) can even hit as high as 9000 MT/s if you have a god-bin IMC.
Memory training is about establishing signal integrity between the memory controller and RAM sticks and has absolutely nothing to do with memory errors. You can boot, train just fine, then instantly BSOD when logging into Windows if your RAM is unstable.
I’ve seen people literally bragging before that their RAM overclock is “stable” because it trained and booted into Windows. Then they run Karhu or TM5 and it instantly errors.
The memory controller on Ryzen 700 isn’t necessarily better. The current Ryzen 7000s are bottlenecked by the infinity fabric running slower than the UCLK and MCLK at higher RAM frequencies, so there’s very limited benefit of running higher than 6400 MT/s with UCLK=MCLK/2 mode (Gear 2). For Zen 4, you would get the most benefit from running 6400 MT/s at UCLK=MCLK mode and tightening timings from there.
The monolithic die design of the 13th and 14th gen allows scaling of bandwidth with memory frequency, as there’s no infinity fabric bottleneck. I’m not saying this is necessarily better, but 13th and 14th gen read and write bandwidth will scale with memory frequency.
That is until 8000mt/s is reached in Gear 2 (People hitting 10K mt/s one after another on 8000 series). This is now quite easily done on AM5. The more recent BIOS for AM5 has allowed users to hit 2300+ on the Infinity Fabric. In the right hands, as of right now. AM5 does handle DDR5 better. Don’t get me wrong, AM5 is no holy grail platform. There is additional latency created by multiple chiplets on a 7950X3D, and using an OS to control chiplet usage is pure folly. The 7800x3d displays way better memory latency then the 7950X3D. It’s a hard fact, Intel above 7000mt/s there is no guarantee that memory will post no matter what you do or who you are. That doesn’t make Intel bad, In fact, If i could back to the Spring, I would’ve purchased a 13900K and kept my 64Gb of Samsung B-die DDR4 (Which DDR5 can’t touch) I will always prefer tight timings over high bandwidth personally
WHY you ask? Because the hype train for AMD has passed, and I’m left with a chip I can’t OC. Undervolts do absolutely nothing, except introduce poor stability in low wattage usage (Many undervolters PC’s crash when Idle, or running netflix) And it’s a feature that does… nothing. Woohoo, I can run less voltage on a chip that never gets hot anyways (Not since I removed MSFS that is) I found it actually runs better slamming it more voltages (Set PBO for “Motherboard Limits” and Maximum power plan) I’d rather be Overclocking my CPU, and see those benefits
The Infinity fabric itself is bandwidth limited and doesnt have anything to do with FCLK frequency. Zen 4 infinity fabric bandwidth to a CCD is 64 GB/s and 32 GB/s from a CCD. Single CCD Ryzens (like the 7800X3D or non3D) will have no benefit of running anything faster than 6400 MT/s because of the severe bandwidth bottleneck from the Infinity fabric with a single CCD. Dual CCD Ryzen chips can have a benefit from running higher RAM frequencies, but this doesn’t really overcome the latency penalty of running UCLK=MCLK/2 until 8000 MT/s or so, and even then can start hitting bandwidth limitations from the infinity fabric.
This isn’t an Intel vs AMD argument, but I am simply pointing out that Zen 4 is bandwidth-limited due to the infinity fabric, especially for single CCD. Intel will have the full memory bandwidth available without any bandwidth bottlenecks due to the monolithic die.
Yes, the higher read bandwidth and lower latency with higher RAM frequencies can help signifcantly improve the 1% lows when there’s a heavy CPU bottleneck like MSFS. If you want to maximize performance for MSFS, the way to go is large L3 cache or maximize RAM frequency for Intel and dual CCD Ryzens.
I also wouldn’t put too much trust in Jay’s videos. The 1% lows are conveniently absent from his benchmark results, and that would be where faster RAM would shine.
Not totally up to speed here on DDR 5 memory speeds and I wonder if anyone here could answer these (pretty simple I think) questions.
Taking the AM5 Asus Tuf Gaming x670 e plus wifi MB as an example which has a DDR 5 base rating of 5600 MHz and an (EXPO) OC rating of upto 6400 MHz and 2 DDR 5 32gb modules rated at 660O MHz (EXPO):
(1) without EXPO overclock, would these modules have a speed at boot up of 5600 MHz?
(2) if you don’t EXPO overclock, would there still be DDR memory training at boot up?
(3) would 5200 MHz DDR 5 memory still work in the MB or would it have to be rated to at least 5600 MHz?
(4) the maximum you could overclock this 6600 MHz DDR 5 memory is limited to the max allowed by the MB which is 6400?
(5) if you buy 6600 (XMP) MHz by mistake would these still work at all in this EXPO MB?
(6) if you buy XMP DDR 5 modules would this mean that it would not be possible to OC the DDR 5 modules at all?
(7) can you reduce boot up times by not overclocking your DDR 5 memory?
I think that’s probably the max automatic but there’s a lot of manual control if you want to go crazy.
Probably will work. May need to manual program timings?
I think you could.
Yes it’s a lot faster without the training, and for me I’ve not been able to use the option that remembers the training to speed up boot because when I enable that I get random crashes.
There has been a bug with AM5 for some time. To enable Memory Context Restore, you must also enable Gear Down Mode. Then It should work as advertised.
EDIT: The more recent AM5 BIOS’s the memory training is so fast, none of this is really required anymore though. It really takes a matter of seconds to train my 64GB now. Around 45 Seconds, used to take upwards of 15 minutes lol Also, my training is extended RX Burst length 8x, and TX Burst Length 8x with DDR5 Nitro. Even that extended… 45 seconds
No, You’re going to have to train your memory whether it’s OC’d or not. You can however extend the training in favour of added stability using DDR5 Nitro feature
I had memory context restore enabled, so not every one. But If I adjusted anything in the BIOS related to memory, I would go through another 15+ minute retrain. Even with Memory Context Restore enabled it will still retrain every so often. It made Overclocking memory very tedious, and even more time consuming
When I first built my machine, I thought I messed up when on first boot it was black screen with a MB 15 code. I shut it down, and tried to boot again. Took me a few hours before I went on my phone, and learned about DDR5 memory training, and the painfully long first boot at that time
Not too knowledgeable on this (that’s quite obvious I think ) but with the memory controller being on the AMD CPU I wonder if these memory training times and boot up times could potentially be improved significantly with the new 9000 CPUs and 800 MBs probably incoming in the 2nd half of the year.
I know nobody really knows yet (outside of AMD) but just wondering …?
The memory controller has been on the CPU for a while now, so nothing new there in terms of design. The latest AGESA versions have significantly improved training boot times, so this is no longer as big of an issue.
) but with the memory controller being on the AMD CPU I wonder if these memory training times and boot up times could potentially be improved significantly with the new 9000 CPUs and 800 MBs probably incoming in the 2nd half of the year.