A video dedicated to teach how to tune the memory for Ryzen would be incredibly appreciated! Also explaining what exactly is being changed and why it improves performance.
Best start with the Ryzen DRAM calculator and Typhoon Burner (careful, that's a tool that can damage your computer) to find out what chips your RAM has.
Yes please, its been a long time since we have had an expert explain and show how to properly tighten memory timings.. Please make content on that!!!!!
Hi Steve! Great video, and thanks for the investigation! I think 6% preformance uplift in corona is kind of a big deal. Those guys who are using corona/vray/cinerender are rendering multiple images for hours sometimes. With 10 1 hour long renders you can make up 36 minutes, when you can get back to your PC and do stuff. So I think it is worht it to talk about 5-10% performance uplifts with these types of applications. Keep up with the great work guys! Have nice days!
this is the most interesting testing i've seen in a while. I've never messed with fine tuning memory in 20+ years of building PCs. I had no idea it could make that much of a difference even on a ddr4 3000 kit. Thanks for the awesome work!
Yes please do a video on how to fix those sub timings! Thinking on upgrading to Zen 2 so great that this video showed the Flare X and Trident Neo, memory I have been looking at. But another video showing the calculator would also be great ;) Or even better.
Agreed. I bought C16 because fuck me if I was about to pay three times as much for C14 just to get a marginal framerate bump, especially when now I know you can just manually tune up the timings.
This is a bit late, but a lot of the kits that you're talking about like 3200 14-14-14 @ 1.35v bins are high end samsung b-die which can readily be clocked to tight primaries + subs at any frequency that you feel like. 3800c16? No problem. 4400c18? As long as the IMC and Motherboard can do it. Meanwhile, a lot of the other cheap 3200 kits - 16-18-18 at 1.35v is common - are bad bins and lower end IC's which aren't capable of tight primary or subtimings, struggle to scale frequency, might not respond well to additional voltage and so on. You can often get more performance out of them but they are only a shadow next to the capability of actually good kits. It's of questionable value to spend days or weeks of your time tuning a dud kit to be okay when you could use the time to buy something better, tune it for a fraction of the time and have far better performance across the board.
Actually the GTX 2080ti is more costly. You must have enough firepower to invade area 51, then steal the dimension traveling machine, go to an alternate reality where gtx2080ti is available and buy (or steal) from their. Not worth the risk
Steve, your a machine! Thanks for the info and good job as usual. Time to buy some cheap RAM and be done with it, I have been waiting for information like this for some time now. Thanks again.
3600 CL14 custom sub timings is the ticket with 1800 Mhz Fclk. This is the line where performance increase is noteworthy. Beyond this, cost can go up dramatically, stability changes or might not even be achievable and so on. This is not to mention in my experience with 3600 Cl14 and 3700x/3900X I am able to match or better my 9900K on 3600 CL16 XMP with 2080 Ti in 1080/1440 gaming. Bottom line however, with a good GPU and 1440p res or less memory tweaking is just as important as ever on Ryzen and now you can push new boundaries.
@@JayPTek Steve posted a while ago that overclocking the infinity fabric on Ryzen 3000 performed differently based on the processor, like It was harder to achieve stability on a 1800 MHz IF with a Ryzen 5 3600. Do i have to buy a 3600X in order to get my memory sticks to 3600mhz on a 1:1 IF ratio? I would appreciate some further info about It.
@@jaxmoreira_81 You shouldn't have any issues running 1:1 with 3600 memory on R5 3600 and up. It could be possible to have issues with an older board that doesn't support that memory speed. Pretty much any X570 should be fine and many X470 boards. 1900 Fclk is in the realm uncertainty whereas every R5 3600 and up should do 1800 without issues.
RAM overclocking now has more effect on performance than CPU overclocking... As others said, an in depth memory overclocking guide for Ryzen would be nice. I’ve got my 3200-16 kit running at 3600-17 but I’m sure I get get more out of it.
I know I'm 3 years late, but just wanted to mention that I have done something similar, had 2* 3200-16 and now running them at 3533-16, 3600 will require big voltage increase, probably outside of safe limit unless I loosen timings, which I don't want to. Will try some BCLK overclock as well as DRAM calculator to squeeze out a bit more.
ASUS ROG Strix X470-I with 3700X and with latest bios and “XMP”/DOCP doesn’t work. Even manually setting voltage doesn’t work, BUT manually overclocking is super easy. Bought Corsair Vengeance RGP PRO 3466 CL16 (b die), and it easily clocks to 3800 16-18-18-36 @ 1.4V (FCLK 1900) or 3733 14-16-16-36 @ 1.4V (FCLK 1867).
And I've been reading that people saying that you can't overclock anything without a x570 if you're using 3700 etc etc .... Maybe they like to dump money
WoW!!!! I tried to Overclock and do a battery of testing like you did with my Ryzen 1700 and VEGA 64, and after 4 hours I was bored AF! I cannot imagine doing that with all those settings and different components, thank you guys, amazing vids, super professional as always.
For my new rig based on the 3600, I got the corsair Vengeance LPX DDR4 16GB Kit (2x8GB) 3200MHz CL16 kit which, for whatever reason, is said to perform really well with ryzen 5. I'll test it next Monday when I get my stuff
@@faycalassaad6066 now i think im feeling sad that almost just a week ago I just got a 3200mhz tforce delta rgb 16gb 2x8. :( I didnt knew that 3000mhz was better for my 3600
I don't know of a T topology layout for x570, maybe ASRock has one but all the others are daisy chained. This limits you to two sticks for optimal conditions, so I bought 2x16 GB Trident Z Neo CL16 and will manually tune it like Steve.
@@Jeon1958able to be fair unless your using a 2080ti it seems dont bother doing anything was the overall thing i got from the video ,steve basically said it at the end but mentioned a ryzen5 based system (which usually wouldnt have a $1000 gpu anyways)
Jeon1958able As Linus said, you can tweak both memory and infinity fabric separately so that they aren’t running at 1:1 (sync) all the time. Out of sync memory-to-fabric sometimes give you a performance boost as seen on 2133Mhz low latency memory kits running at 1900Mhz FCLK. You get the performance of out-of-the-box 3600Mhz memory kits, with 2/3rd of freq.
Thank you for you work and research. I am looking at parts today for a new build. Going Ryzen and looked for RAM info, I found this today. Thank you again, for helping me make an informed decision. Stay safe and keep up the great work. Just sub'd and looking forward to more.
I was waiting for Battlefield V . Cause It is one of the cpu bound games . Anyway thanks for the video . Waiting for tutorial video about manually changing memmory timings and overclocking infinity fabric
Damn, would have hoped that 3700x/3900x with RAM tweaking would have matched/beaten the 9900k , but looks like 9900k still destroys when it gets a similar treatment of RAM tweaks + uncore OC..
@uP-Andre normal people don't care if they get 450 fps or 500 fps in csgo. They want the best for the buck, like the 3600 or 3700x. Intels offerings are just not as good in the reasonable price range.
@@McCarthy880 i just used ryzen memory calculator, samsung b die, 3600 single rank and followed timings by taking picture with phone, and off i went, took less than 5 mins
Let’s get that Ryzen DRAM calculator video going! And explanations of what all the acronyms stand for, and why each setting impacts performance. Oh and maybe stuff like why you want to use normal over something like extreme etc. thanks for all your videos!
I'll add my voice to the chorus of requests for a memory tuning how-to. I've wanted to do that for some time but I'm baffled by the BIOS interface for memory timings on my board.
Great video, Steve! I also saw Buildzoid's video on memory timings and apparently he saw big performance gains by running dual rank memory. Would be interesting to see how high frequency single rank memory stacks up vs lower clocked dual or even quad rank memory.
Buy a 3200mhz CL16 Micron E die memory. It is really cheap and overclocks like crazy. You can overclock it to all the way to 3800 mhz CL16 somewhere @1.4-1.45V without problems.
@Yearning for Another World check DDR4 IC lists online. Many forums have threads where users accumulate that data. For Micro E Die specifically, a safe bet are all Crucial Ballistix kits with -AES* at the end of the model number.
I ran 3200 CL14 and 3600 CL16 B-die kits both at 3600 CL14 no problem on Asus X470 Hero. These kits really aren't that expensive. You can get a non RGB GSkill set for $130 (16GB) and 3600 Cl14 is bit better than 3800 CL16 from my own testing. Save money where you have to save money though I get it. Buy memory based on your use case not just because some of use swear by it. Want max frames for competitive play? You need to put more money in your memory.
@@JayPTek B die is ridiculously expensive compared to other ICs. 130$ for 16GB lol, Micron E die is like 80$ or so for a 16GB kit. And that E die is fine for Zen 2. Easily reaches 3600MHz (or 3400 for a bad batch) at CL16 or CL14 (if it's a really good batch) and is only slightly behind B die at that point. B die can more often get to 3800, 4000, etc, but for Zen 2 that's pointless because most chips don't make it much past 1800 FCLK. PS: really cheap ICs cost less than *half* of your B die price. True, they often top out at like 3000-3400 with mediocre timings, but still.
One of the things about your reviews I like so much is you detail performance based on a combination of GPU/CPU/Resolution (and now memory). It's beneficial to know what CPU is faster in games overall but it's also really really beneficial to know that at a particular resolution with X other hardware it won't matter. My next build I'm leaning towards a 3600 (was going to get a 3900x until I watched your reviews and Gamers Nexus) and a 5700XT to replace my 4670k and GTX 1070. However, since I'm running at 3440x1440/60hz I now know that the 3600 memory I was going to buy isn't really needed and I'll be fine with 3000 or 3200 which is cheaper. You have really helped steer me away from overpaying in my last build or two for parts that won't really help that much. Thank you.
@hardware unboxed you are almost there. There is one key thing that you seem not to have thought about yet. Infinity Fabric interconnects between each of the CCD modules to the Scalable Data Fabric (SDF) on the IO die and the interconnect from the SDF to the Dual Channel memory controller are both 32 byte/cycle connections. Dual channel DDR4 memory has the same theoretical maximum bandwidth and is restricted by the timings overhead. You can calculate the maxumum Infinity fabric interconnect bandwidth by multiplying 32 bytes x the actual Frequency of the FCLK which by default is 1:1 with the Ram frequency. Theoretical maximum bandwidth of the Infinity Fabric for different discussed Ram configurations are as follows: 3800 32bytes x1900,000,000 cycles (1900Mhz)=60.8GB/s 3600 32 X 1800,000,000 = 57.6GB/s 3200 32x1600,000,000 = 51.2GB/s 3000 32x1500,000,000 = 48GB/s Take off 5% or so for the timings overhead and you end up with an actual memory Bandwidth result that is pretty close to what you observed in the video for each of the memory kits. Intel by comparison uses Ringbus to perform the same connectivity functions as Infinity Fabric. With a cache multiplier of x45 you have a total theoretical maximum bandwidth of: 32x4500,000,000=144GB/s. The Dual Channel memory will still perform at roughly the same level of bandwidth as observed with Ryzen. L3 Caches on both Ryzen and Intel operate at 500+ GB/s and do not impose any limitations on performance. The interconnects between Ryzen SDF and CCDs and Intel Ringbus to teh memory and PCIe controllers are both shared resources that carry both CPU memory read/write requests to the RAM and GPU DMA requests between GPU/PCIe controller and the L3 Cache. A 2080Ti can just about saturate PCIe 3. For the sake of the discussion lets say that it needs 15GB/s of instructions/data from the L3 Cache to reach the limits of its performance. In reality that number is variable depending on the game, how many instructions the CPU can actually process etc. The GPU relies on the CPU to process and output the graphics instructions to the GPU and is making memory requests concurrent with the GPU rendering images to the monitor. In reality On a 3900x there are two interconnects that can spread the load if the application also spreads its processing across threads on both Chiplets. not all games do. On a 3700x, there is only a single interconnect to the L3 Caches because of the single chiplet. Ringbus bandwidth exceeds the total memory read/write bandwidth by a significant amount and allows a 9900K, while concurrently servicing the 15GB/s demands of the GPU. The CPU can still access the RAM at the same levels to continuously access memory at the rate the bandwidth benchmarks indicate. On the other hand, your Ryzen memory bandwidth tests are tested with the CPU having exclusive use of the interconnect and while valid for blender, handbrake etc, are not valid during gaming. With an FCLK of 1900Mhz, Assuming that all teh processing is done on a single die, the 15GB/s requirement of GPU memory access running in competition with the CPU memory read requests, you only have an absolute maximum read bandwidth of about 45GB/s less the timings overhead a game. Ryzen receives about 10GB/s less data per second than a 9900K to process all the games instructions. FCLK at 1600Mhz only leaves 36-40GB/s of bandwidth available to the CPU. Less memory data being delivered to the CPU = less instructions being processed and given to the GPU = lower frame rates than a comparable Intel CPU. The one area that you did not look at that is important is the fact that Ryzen 3000 allows you to decouple the Infinity Fabric Clock from memory frequency for the first time. The higher FCLK increases the Infinity Fabric total bandwidth to levels higher than the max theoretical bandwidth of the DDR4 dual channel ram at the cost of slightly higher latency because the ratios are not at 1:1. at 1900Mhz FCLK, The 15GB/s GPU load will not compromise the CPUs memory read bandwidth as much so more data = more FPS In games, a 3800Mhz kit is only giving you about 80-85% of its total performance with the rest being left unused because of the Infinity Fabric bottleneck. It will give you the best performance at very high costs. If you run a 3000 or 3200 Ram kit with the tightest timings possible and the FCLK set to 1900Mhz, the tight timings help offset the latency penalty caused bu the higher FCLK setting. You should see frame rates with a 2080Ti close to or matching the results that are shown in the video for a 3800Mhz Ram kit. The reason is that with the 60.8GB/s of Fabric bandwidth, those kits are running at or near 100% of their performance capacity instead of 85% as with a 3800kit. Wile higher latencies do have a small impact on performance they are well outweighed by the reduction in the bandwidth bottleneck. 2133MT/s Ram us well underperforming with regards the Infinity fabric bandwidth limits and that is th ereason why that doesnt perform very well
#HUN any chance to do a "standard build" aka around 2070 level/5700XT and fitting CPU and then test it raw stock out of the box vs. fully optimized/overclocked/optimized on CPU/RAM/GPU etc. Seems like 5-10% here and there and everywhere could lead to a fair boost total with everything optimized. Nice video, looks like manually tuning your RAM is the best way to go and if it is as simple as it sounds with the memory calculator then that is nice enough
In before people start to ask Steve to compare 3900x with 9900k at ddr4 3800 optimized timing fclk
5 лет назад+3
It's certainly easier to use Intel. This timings stuff is beyond what i've ever done. I prefer xmp and a simple OC on my cpu so i can get back to gaming.
@ you don't have to do this and just do typical Intel drop and forget about it thing and Ryzen will still work fine. Fine tuning is just part of the fun for us hands on enthusiasts. Tuning memory can increase Intel performance too btw. This is not like omfg I don't know how to tune memory so Ryzen just won't work, gotta get myself an Intel.
@ I'm not 100% sure but I think if you change the clock speed it will use auto timings, which are very loose and will hurt performance even if the clock speed is higher, think about it as increasing the clock speed a bit but losing a lot of IPC, resulting in worse performance. You could check the timings in the BIOS, with CPU-Z or with HWinfo64. If you're going to overclock the RAM, watch a guide, use Ryzen DRAM calculator and stress test it properly, you also need to know which memory chips you have and other things. If you don't want to bother just use XMP. Anyways, this isn't really about being necessary or not being necessary, Ryzen 3000 is already plenty fast at stock, the point of doing this is to squeeze all the performance possible.
Buildzoid has some great vids on this topic also. 3800 CL14 with infinity fabric manually tuned to 1900mhz seems to be the absolute best OC settings possible. And you are prrooooobably only getting that with TridentZ Neo or other B die sticks. Edit: killer vid btw. I wish my B450 Steel Legend had more up to date AGESA so i could mess with things without crashes and had boosts closer to the rated 4.2ghz instead of it being impossible to get past 4.1 to 4.125 on my 3600, let alone messing with memory much Til then i guess lol
XMP vs. Manual Timings..... I think a better test case might be XMP vs manual memory overclocking. Though, understandably, memory performance differs greatly comparing die to die/within same die, how much voltage are you willing to put in, and quality of mobo chokes.... etc
I have been eyeing that particular DDR4 3000 and am very happily surprised by what you have found. I will purchase some quite soon. Add my vote for a memory tuning tutorial, please and thank you.
THANK YOU for the hard work you put into making this video!! I plan on surprising my friend with a PC build in November so he can enjoy Forza Horizon 3 and 4 (among other non-PS4-exclusive games) in 4K at 60fps on his 60Hz 4K tv. Xbox One X would be a waste of money for him even on Black Friday. I'm getting the 3700X for the build. Now I have an even better understanding of what RAM to get for Ryzen 3000.
Excellent video. When Linus starting talking about FCLK and sub-timings I hopped right on and started messing around, and basically came to the same conclusion that you just did. FCLK can do something if you are running at 2133, but once you get to 3000 (1500MHz FCLK) then pumping it up more generally doesn't do a whole lot. Changing the sub-timings has a bigger effect and it seems like a lot of memory can handle that quite handily. Even normal 2133 memory can usually be overclocked to at least 2666CL15 just by pumping the voltage up to 1.35V. In fact, most memory kits are just binned 2133 memory running at a higher voltage. Its not as though the chip makers are actually producing these higher-spec chips. Running faster and tighter is 80% voltage and 20% silicon lottery. I do have to add a note of caution, though. Messing with sub-timings might seem stable but it IS playing with fire to some degree. It isn't as guaranteed stable as people think, and it also tends to make systems a bit more vulnerable to memory-based attacks like RowHammer, particularly if you don't give the RAS cycle enough time to fully recharge the row. -Matt
Can you point which timings had the biggest impact on performance when manually overclocking? That manually overclocked DDR4 3000 results are really impressive. Great vid btw!
Nice! So I made the right pick a couple of weeks ago, when my brother and me ordered 64gig DDR4 3000 CL16. I was a little nervous when I clicked on this video. ^^
Interesting because, yes, you're almost never CPU bound in most gaming scenarios--but all testing is in CPU bound scenarios. Very comprehensive testing!
Another great video! Thank you! I know you get a lot of request, but a video on how to best make use of that Ryzen memory calculator tool would be amazing. Either way, keep up the great work! Hope all else is going well for you guys 👍🏼😁
Great video Steve! Yes, I would love to see a video explaining how to setup manual DD4 memory timings (Ryzen or otherwise). I haven't seen a good video on how to manually tune DDR4 memory timings and really want to try it!
I would have liked to see other games (namely hitman, given its odd nature) but 120 benchmark runs per game is just absolutely incredible and and you will hear no complaints from me. Well done guys, a masterclass on benchmarking and testing.
A great video, thank you. I particularly appreciate the fact that you included the RX 580 as we do not all have the latest, greatest GPUs. Liked and subscribed!
![AMD Ryzen 7000 Series [Zen 4] RAM/Memory Scaling feat. Tuned Memory](http://i.ytimg.com/vi/MOatIQuQo3s/mqdefault.jpg)
A video dedicated to teach how to tune the memory for Ryzen would be incredibly appreciated! Also explaining what exactly is being changed and why it improves performance.
Check out Actually hardcore overclocking's video on Ryzen ram stuff
true, i have to make cmos every time I use it. Maybe because i have d-die
@@imfllirymydude341 got a link to that video?
Yes it would be really neet
Best start with the Ryzen DRAM calculator and Typhoon Burner (careful, that's a tool that can damage your computer) to find out what chips your RAM has.
Wow, Steve. Great moves, keep it up. Proud of you!
Papa bless
Haha h3 reference
@@Hardwareunboxed lmao
Letters you can hear
@@Hardwareunboxed VAPE NATION
Yes please,
its been a long time since we have had an expert explain and show how to properly tighten memory timings.. Please make content on that!!!!!
Hi Steve! Great video, and thanks for the investigation! I think 6% preformance uplift in corona is kind of a big deal. Those guys who are using corona/vray/cinerender are rendering multiple images for hours sometimes. With 10 1 hour long renders you can make up 36 minutes, when you can get back to your PC and do stuff. So I think it is worht it to talk about 5-10% performance uplifts with these types of applications. Keep up with the great work guys! Have nice days!
Just what the doctor ordered!! Gotta love HU!
Only channel I actually use the Bell on lool You guys are such quality! Thanks for all the hard work!
C'mon man. Gamers Nexus!
this is the most interesting testing i've seen in a while. I've never messed with fine tuning memory in 20+ years of building PCs. I had no idea it could make that much of a difference even on a ddr4 3000 kit. Thanks for the awesome work!
Huh? It only makes a difference when you playing low res and low settings. Soon as you enable realistic used settings there is almost zero difference.
yes please make another video on how to use ram timings :), thanks
You can look this up on reddit or other places like that too presumably
github.com/integralfx/MemTestHelper/blob/master/DDR4%20OC%20Guide.md
Agreed. This is a topic I want to see more about since it can be inscrutable.
I have option to go for g.skill triden 3600 or hyperx predator 3600, is g.skill better for ryzen or it doesn't matter?
Thanks sooo much for including the Rx580 for context! I appreciate all the extra work it must have been.
Yes please do a video on how to fix those sub timings!
Thinking on upgrading to Zen 2 so great that this video showed the Flare X and Trident Neo, memory I have been looking at.
But another video showing the calculator would also be great ;) Or even better.
Some surprising results. Some expected results. Definitely worth a watch and I thank you for all the work!
I would like to see DDR4 3200 C16 vs C14. C16 is far more common, and typically far more affordable.
Yes please
Agreed. I bought C16 because fuck me if I was about to pay three times as much for C14 just to get a marginal framerate bump, especially when now I know you can just manually tune up the timings.
his 3000mhz is samsung s-die this mean that is cl 16 or more without improved tiing
This is a bit late, but a lot of the kits that you're talking about like 3200 14-14-14 @ 1.35v bins are high end samsung b-die which can readily be clocked to tight primaries + subs at any frequency that you feel like. 3800c16? No problem. 4400c18? As long as the IMC and Motherboard can do it.
Meanwhile, a lot of the other cheap 3200 kits - 16-18-18 at 1.35v is common - are bad bins and lower end IC's which aren't capable of tight primary or subtimings, struggle to scale frequency, might not respond well to additional voltage and so on. You can often get more performance out of them but they are only a shadow next to the capability of actually good kits.
It's of questionable value to spend days or weeks of your time tuning a dud kit to be okay when you could use the time to buy something better, tune it for a fraction of the time and have far better performance across the board.
@@AerynGaming Are you saying a higher quality RAM is better than a lower quality RAM? Damn...
This video single handily answered questions I've been trying to answer for weeks. Thank you so much! Gave this video a like and subscribed.
That GTX 2080ti does look like a good GPU 8:30 I hope it's not overpriced like the RTX 2080ti
And I guess we get no 3800 speed manual timings on the rx 580 on far cry new dawn
Actually the GTX 2080ti is more costly. You must have enough firepower to invade area 51, then steal the dimension traveling machine, go to an alternate reality where gtx2080ti is available and buy (or steal) from their. Not worth the risk
Actually,it should be named gtx 1680 ti.
lamo
Glad to know Im not the only one who "spot" that one.
Steve, your a machine! Thanks for the info and good job as usual. Time to buy some cheap RAM and be done with it, I have been waiting for information like this for some time now. Thanks again.
3600 CL14 custom sub timings is the ticket with 1800 Mhz Fclk. This is the line where performance increase is noteworthy. Beyond this, cost can go up dramatically, stability changes or might not even be achievable and so on. This is not to mention in my experience with 3600 Cl14 and 3700x/3900X I am able to match or better my 9900K on 3600 CL16 XMP with 2080 Ti in 1080/1440 gaming. Bottom line however, with a good GPU and 1440p res or less memory tweaking is just as important as ever on Ryzen and now you can push new boundaries.
Faycal Assaad Hey Fay!
@@JayPTek Steve posted a while ago that overclocking the infinity fabric on Ryzen 3000 performed differently based on the processor, like It was harder to achieve stability on a 1800 MHz IF with a Ryzen 5 3600.
Do i have to buy a 3600X in order to get my memory sticks to 3600mhz on a 1:1 IF ratio?
I would appreciate some further info about It.
@@jaxmoreira_81 You shouldn't have any issues running 1:1 with 3600 memory on R5 3600 and up. It could be possible to have issues with an older board that doesn't support that memory speed. Pretty much any X570 should be fine and many X470 boards. 1900 Fclk is in the realm uncertainty whereas every R5 3600 and up should do 1800 without issues.
@@JayPTek Thanks!
RAM overclocking now has more effect on performance than CPU overclocking... As others said, an in depth memory overclocking guide for Ryzen would be nice. I’ve got my 3200-16 kit running at 3600-17 but I’m sure I get get more out of it.
I know I'm 3 years late, but just wanted to mention that I have done something similar, had 2* 3200-16 and now running them at 3533-16, 3600 will require big voltage increase, probably outside of safe limit unless I loosen timings, which I don't want to. Will try some BCLK overclock as well as DRAM calculator to squeeze out a bit more.
Really appreciated! :) Very nice to see comparisons between different classes of GPU's too.
You just saved people a ton of money.
Thank you Sir.
I thought I made a bad call going for 3200 memory with my new build. Thanks for clarifying how much more I would gain going to 3600 and above.
Can’t wait for the g.skill neo 3800 CL14
@@erichall090909 Agreed although I'm thinking a good 3200 CL 14 kit is 130ish now.. that kit is probably gonna be 175 or so.
@@erichall090909 especially if you crank that voltage and try and drop the cl value. Wonder if it will lower anymore? :)
@@Seatux what's your cpu sir?
Appreciate you being so thorough. Excellent report 😎👍
Wow, that was some impressive amount of work. Well done, thank you Steve!
Dang dude. That was a lot of testing. Props to you. Left a like
Keep up the good work!
ASUS ROG Strix X470-I with 3700X and with latest bios and “XMP”/DOCP doesn’t work. Even manually setting voltage doesn’t work, BUT manually overclocking is super easy. Bought Corsair Vengeance RGP PRO 3466 CL16 (b die), and it easily clocks to 3800 16-18-18-36 @ 1.4V (FCLK 1900) or 3733 14-16-16-36 @ 1.4V (FCLK 1867).
And I've been reading that people saying that you can't overclock anything without a x570 if you're using 3700 etc etc .... Maybe they like to dump money
Hey can u clear one thing for me.
I am getting a mobo with 3466(oc) support can i run a 3600mhz ram on it?
@@cyano3d Yes, you can.
cyanogenmoded yes, more than likely. Will have to manually set speeds and voltages.
@@cyano3d yes, my mortar titanium running 3666mhz cl16 fclk 1833
WoW!!!! I tried to Overclock and do a battery of testing like you did with my Ryzen 1700 and VEGA 64, and after 4 hours I was bored AF! I cannot imagine doing that with all those settings and different components, thank you guys, amazing vids, super professional as always.
Literally reading about DDR4 memory options for a Ryzen 5 3600 now and find this posted half an hour ago. Perfect timing!
For my new rig based on the 3600, I got the corsair Vengeance LPX DDR4 16GB Kit (2x8GB) 3200MHz CL16 kit which, for whatever reason, is said to perform really well with ryzen 5. I'll test it next Monday when I get my stuff
@@faycalassaad6066 now i think im feeling sad that almost just a week ago I just got a 3200mhz tforce delta rgb 16gb 2x8. :( I didnt knew that 3000mhz was better for my 3600
Quality Content. Presentations are always professionally done. I now have 2 favorite channels in Hardware Unboxed and GN!
Awesome video Steve, always appreciate the time and effort you guys put into your content
Thanks for a great memory speed performance breakdown. Very interesting about the manual latency results there Steve.
As an acoustic engineer, I prefer more ram vs faster ram. My gf said the same thing, though I think she means different kind of ram.
She prefers more male sheep over faster male sheep? To each their own.
Depending on what you do, most x570 use daisy chained, making it worse with more than 2 sticks
I don't know of a T topology layout for x570, maybe ASRock has one but all the others are daisy chained. This limits you to two sticks for optimal conditions, so I bought 2x16 GB Trident Z Neo CL16 and will manually tune it like Steve.
Well, most people prefer more carrying capacity over speed in a truck, so your girlfriend isn't wrong.
@@praetorxyn How much fps do i really loose if i put more than 2 sticks in it? I need a video on that!
I swear Steve be answering my questions with these videos! Great work
So basically, tweak the RAM!.
I thought he meant overclock the infinity fabric.
@@Jeon1958able to be fair unless your using a 2080ti it seems dont bother doing anything was the overall thing i got from the video ,steve basically said it at the end but mentioned a ryzen5 based system (which usually wouldnt have a $1000 gpu anyways)
Jeon1958able As Linus said, you can tweak both memory and infinity fabric separately so that they aren’t running at 1:1 (sync) all the time. Out of sync memory-to-fabric sometimes give you a performance boost as seen on 2133Mhz low latency memory kits running at 1900Mhz FCLK. You get the performance of out-of-the-box 3600Mhz memory kits, with 2/3rd of freq.
@@Jeon1958able you have to do both
Thank you for you work and research. I am looking at parts today for a new build. Going Ryzen and looked for RAM info, I found this today. Thank you again, for helping me make an informed decision. Stay safe and keep up the great work. Just sub'd and looking forward to more.
Insanely informative video.
Now I just have to look for a decent 32 GB kit of memory with Hynix C-Die or Micro E-Die. Maybe even with RGB ;)
No RGB but Crucial Ballistix Sport LT 2x16GB 3200MHz works well BLS2K16G4D32AESB
Micron E-Die is the new thing so I would go for them, Crucial I use for years and years with absolutely no issues
Lots of channels cover this stuff, but your results and suggestions always seem much more consumable and directly helpful.
I was waiting for Battlefield V . Cause It is one of the cpu bound games . Anyway thanks for the video . Waiting for tutorial video about manually changing memmory timings and overclocking infinity fabric
Excellent video from all aspects! Top notch quality there Steve. Will help muchly many people in buying correct RAM kits and tuning Ryzen based PCs.
excellent video as always steve! can you do 9900k @ max oc, tweak ram vs 3900x @ max oc, with tweak ram?
Intel gains even less (less than 1% from 2400 -4000mt/s)
Not Steve but the results are predictable ruclips.net/video/LCdA-bLRAfM/видео.html
This, also with 9900k's Uncore overclocked please.
Damn, would have hoped that 3700x/3900x with RAM tweaking would have matched/beaten the 9900k , but looks like 9900k still destroys when it gets a similar treatment of RAM tweaks + uncore OC..
@uP-Andre normal people don't care if they get 450 fps or 500 fps in csgo. They want the best for the buck, like the 3600 or 3700x. Intels offerings are just not as good in the reasonable price range.
Appreciate the effort, Steve! Looking forward to the mem timings video. Thank you!
Benchmark king, trim bearded jesus has heard my question! and I listen to his wisdom!
Wonderful job! thanks for adding the GPU scaling, really shows the benefits on every range.
3:52 yes please, I could use a tutorial on ddr timings!
Dem gains though! Some really impressive jumps in 1% lows with the manually tuned DDR4-3800 kit.
Ryzen memory calculator ftw, got my teamgroup 4000 to 3600 cl 14 for max performance
and there is huge difference between 3600c14 and 3600 c16, there is a RUclips channel called "jayp tek" who made video on it
@@McCarthy880 i just used ryzen memory calculator, samsung b die, 3600 single rank and followed timings by taking picture with phone, and off i went, took less than 5 mins
@@Tallnerdyguy hello. How to make 3600mhz cl14? I watch a benchmark using that ram and they said that its the best ram for ryze 5 3600
@@islandboy9235 google ryzen memory calculator.
Let’s get that Ryzen DRAM calculator video going! And explanations of what all the acronyms stand for, and why each setting impacts performance. Oh and maybe stuff like why you want to use normal over something like extreme etc. thanks for all your videos!
I'll add my voice to the chorus of requests for a memory tuning how-to. I've wanted to do that for some time but I'm baffled by the BIOS interface for memory timings on my board.
Great video, Steve! I also saw Buildzoid's video on memory timings and apparently he saw big performance gains by running dual rank memory.
Would be interesting to see how high frequency single rank memory stacks up vs lower clocked dual or even quad rank memory.
Buy a 3200mhz CL16 Micron E die memory. It is really cheap and overclocks like crazy. You can overclock it to all the way to 3800 mhz CL16 somewhere @1.4-1.45V without problems.
@Yearning for Another World all new crucial ballistix 2x8gb 3200/c16 or 3000/c15 are micron e die
@Yearning for Another World check DDR4 IC lists online. Many forums have threads where users accumulate that data.
For Micro E Die specifically, a safe bet are all Crucial Ballistix kits with -AES* at the end of the model number.
I ran 3200 CL14 and 3600 CL16 B-die kits both at 3600 CL14 no problem on Asus X470 Hero. These kits really aren't that expensive. You can get a non RGB GSkill set for $130 (16GB) and 3600 Cl14 is bit better than 3800 CL16 from my own testing. Save money where you have to save money though I get it. Buy memory based on your use case not just because some of use swear by it. Want max frames for competitive play? You need to put more money in your memory.
@@JayPTek B die is ridiculously expensive compared to other ICs. 130$ for 16GB lol, Micron E die is like 80$ or so for a 16GB kit. And that E die is fine for Zen 2. Easily reaches 3600MHz (or 3400 for a bad batch) at CL16 or CL14 (if it's a really good batch) and is only slightly behind B die at that point.
B die can more often get to 3800, 4000, etc, but for Zen 2 that's pointless because most chips don't make it much past 1800 FCLK.
PS: really cheap ICs cost less than *half* of your B die price. True, they often top out at like 3000-3400 with mediocre timings, but still.
Isn't 1.45v to much? I just killed 1 of my 2 dimms just by tweaking tertiary memory timings (@ 1.35v).
Damn good content ! Those manually tuned 3000s are impresive, and half the cost of 3800 where I'm at.
And we know HWU is a completionist and does all the side quests.
One of the things about your reviews I like so much is you detail performance based on a combination of GPU/CPU/Resolution (and now memory). It's beneficial to know what CPU is faster in games overall but it's also really really beneficial to know that at a particular resolution with X other hardware it won't matter. My next build I'm leaning towards a 3600 (was going to get a 3900x until I watched your reviews and Gamers Nexus) and a 5700XT to replace my 4670k and GTX 1070. However, since I'm running at 3440x1440/60hz I now know that the 3600 memory I was going to buy isn't really needed and I'll be fine with 3000 or 3200 which is cheaper.
You have really helped steer me away from overpaying in my last build or two for parts that won't really help that much. Thank you.
Are we not going to talk about how sometimes the RX5700 had better 1% lows at 1080p than the freaking RTX 2080 TI?
Partially why I hope AMD surprise drops big Navi/5800 XT by year's end. I want a worthy upgrade for my 1080Ti at the $699 price point.
In CPU bound games lol
@hardware unboxed you are almost there. There is one key thing that you seem not to have thought about yet. Infinity Fabric interconnects between each of the CCD modules to the Scalable Data Fabric (SDF) on the IO die and the interconnect from the SDF to the Dual Channel memory controller are both 32 byte/cycle connections. Dual channel DDR4 memory has the same theoretical maximum bandwidth and is restricted by the timings overhead.
You can calculate the maxumum Infinity fabric interconnect bandwidth by multiplying 32 bytes x the actual Frequency of the FCLK which by default is 1:1 with the Ram frequency.
Theoretical maximum bandwidth of the Infinity Fabric for different discussed Ram configurations are as follows:
3800
32bytes x1900,000,000 cycles (1900Mhz)=60.8GB/s
3600
32 X 1800,000,000 = 57.6GB/s
3200
32x1600,000,000 = 51.2GB/s
3000
32x1500,000,000 = 48GB/s
Take off 5% or so for the timings overhead and you end up with an actual memory Bandwidth result that is pretty close to what you observed in the video for each of the memory kits.
Intel by comparison uses Ringbus to perform the same connectivity functions as Infinity Fabric. With a cache multiplier of x45 you have a total theoretical maximum bandwidth of:
32x4500,000,000=144GB/s.
The Dual Channel memory will still perform at roughly the same level of bandwidth as observed with Ryzen.
L3 Caches on both Ryzen and Intel operate at 500+ GB/s and do not impose any limitations on performance.
The interconnects between Ryzen SDF and CCDs and Intel Ringbus to teh memory and PCIe controllers are both shared resources that carry both CPU memory read/write requests to the RAM and GPU DMA requests between GPU/PCIe controller and the L3 Cache.
A 2080Ti can just about saturate PCIe 3. For the sake of the discussion lets say that it needs 15GB/s of instructions/data from the L3 Cache to reach the limits of its performance. In reality that number is variable depending on the game, how many instructions the CPU can actually process etc. The GPU relies on the CPU to process and output the graphics instructions to the GPU and is making memory requests concurrent with the GPU rendering images to the monitor. In reality
On a 3900x there are two interconnects that can spread the load if the application also spreads its processing across threads on both Chiplets. not all games do. On a 3700x, there is only a single interconnect to the L3 Caches because of the single chiplet.
Ringbus bandwidth exceeds the total memory read/write bandwidth by a significant amount and allows a 9900K, while concurrently servicing the 15GB/s demands of the GPU. The CPU can still access the RAM at the same levels to continuously access memory at the rate the bandwidth benchmarks indicate. On the other hand, your Ryzen memory bandwidth tests are tested with the CPU having exclusive use of the interconnect and while valid for blender, handbrake etc, are not valid during gaming.
With an FCLK of 1900Mhz, Assuming that all teh processing is done on a single die, the 15GB/s requirement of GPU memory access running in competition with the CPU memory read requests, you only have an absolute maximum read bandwidth of about 45GB/s less the timings overhead a game. Ryzen receives about 10GB/s less data per second than a 9900K to process all the games instructions. FCLK at 1600Mhz only leaves 36-40GB/s of bandwidth available to the CPU.
Less memory data being delivered to the CPU = less instructions being processed and given to the GPU = lower frame rates than a comparable Intel CPU.
The one area that you did not look at that is important is the fact that Ryzen 3000 allows you to decouple the Infinity Fabric Clock from memory frequency for the first time. The higher FCLK increases the Infinity Fabric total bandwidth to levels higher than the max theoretical bandwidth of the DDR4 dual channel ram at the cost of slightly higher latency because the ratios are not at 1:1. at 1900Mhz FCLK, The 15GB/s GPU load will not compromise the CPUs memory read bandwidth as much so more data = more FPS
In games, a 3800Mhz kit is only giving you about 80-85% of its total performance with the rest being left unused because of the Infinity Fabric bottleneck. It will give you the best performance at very high costs.
If you run a 3000 or 3200 Ram kit with the tightest timings possible and the FCLK set to 1900Mhz, the tight timings help offset the latency penalty caused bu the higher FCLK setting. You should see frame rates with a 2080Ti close to or matching the results that are shown in the video for a 3800Mhz Ram kit.
The reason is that with the 60.8GB/s of Fabric bandwidth, those kits are running at or near 100% of their performance capacity instead of 85% as with a 3800kit. Wile higher latencies do have a small impact on performance they are well outweighed by the reduction in the bandwidth bottleneck.
2133MT/s Ram us well underperforming with regards the Infinity fabric bandwidth limits and that is th ereason why that doesnt perform very well
579th
Very nice video. Was looking forward to this for a while.
#HUN any chance to do a "standard build" aka around 2070 level/5700XT and fitting CPU and then test it raw stock out of the box vs. fully optimized/overclocked/optimized on CPU/RAM/GPU etc. Seems like 5-10% here and there and everywhere could lead to a fair boost total with everything optimized.
Nice video, looks like manually tuning your RAM is the best way to go and if it is as simple as it sounds with the memory calculator then that is nice enough
The sheer amount of work doing all those benches... Thank you.
I have c16 3200mhz kit. Please make video for timings tuning like ur 3000mhz one..
have you tried Ryzen DRAM calculator "fast" preset? :)
@@Istvan_F didnt work for me, had to reset the cmos just to get the pc to reboot
Good to know that adjusting timings makes such a difference. Thank you very much for this video, incredible job!
Interesting would be a test with if-clock @1900mhz and the manual setting for ddr4 3000.
Ahoc has testet this already.
@@SuperUltimateLP Could you link the video
@@paulmurphy6907 here you go:m.ruclips.net/video/10pYf9wqFFY/видео.html
uits been done by actual hardcore overclocking
@@lordzed83thats what i said ?! Ahoc is short for actualy hardcore overclocking..
Thank you doing all that testing! Awesome stuff.
In before people start to ask Steve to compare 3900x with 9900k at ddr4 3800 optimized timing fclk
It's certainly easier to use Intel. This timings stuff is beyond what i've ever done. I prefer xmp and a simple OC on my cpu so i can get back to gaming.
@ you don't have to do this and just do typical Intel drop and forget about it thing and Ryzen will still work fine. Fine tuning is just part of the fun for us hands on enthusiasts.
Tuning memory can increase Intel performance too btw. This is not like omfg I don't know how to tune memory so Ryzen just won't work, gotta get myself an Intel.
Meh unnecessary. I already know how my 9900k works with optimized 4400mhz ram. I’ll run my 3950x at 3800mhz cl14 and sit happy with both systems.
@ You could do this with Intel as well to get more performance.
@ I'm not 100% sure but I think if you change the clock speed it will use auto timings, which are very loose and will hurt performance even if the clock speed is higher, think about it as increasing the clock speed a bit but losing a lot of IPC, resulting in worse performance. You could check the timings in the BIOS, with CPU-Z or with HWinfo64.
If you're going to overclock the RAM, watch a guide, use Ryzen DRAM calculator and stress test it properly, you also need to know which memory chips you have and other things. If you don't want to bother just use XMP.
Anyways, this isn't really about being necessary or not being necessary, Ryzen 3000 is already plenty fast at stock, the point of doing this is to squeeze all the performance possible.
Holy shit Steve! That is a ton of work...thank you for the hard work brotha!
Buildzoid has some great vids on this topic also. 3800 CL14 with infinity fabric manually tuned to 1900mhz seems to be the absolute best OC settings possible. And you are prrooooobably only getting that with TridentZ Neo or other B die sticks.
Edit: killer vid btw. I wish my B450 Steel Legend had more up to date AGESA so i could mess with things without crashes and had boosts closer to the rated 4.2ghz instead of it being impossible to get past 4.1 to 4.125 on my 3600, let alone messing with memory much Til then i guess lol
Cheap e-die sticks can also do that (Crucial Ballistix LT Sport with AES in the product name)
Thanks Steve, memory timing stuff is super appreciated.
XMP vs. Manual Timings..... I think a better test case might be XMP vs manual memory overclocking.
Though, understandably, memory performance differs greatly comparing die to die/within same die, how much voltage are you willing to put in, and quality of mobo chokes.... etc
XMP vs. Manual Timings and XMP vs manual memory OC is the same thing ...
@@janeksheppix Hmm.... I thought they only moved the timing, and not the clock, did I miss something in the video? It's early morning here.
you just saved me well over 120€ by going from a 3000MHz kit to a 3200 and try to manually tweak it myself. you got yourself a subscriber
RAMmstein
Comprehensive and informative man...thank you! Sitting here like..."ok...damn!"
"Manual tuning"
in this video "manual tuning" means just using the calculator instead of using the xmp profiles for the target MHz, i think.
Great work Steve! Answers many questions very conclusively :)
Nice summary at the end. My eyes crossed a bit during all the graphs. lol
I have been eyeing that particular DDR4 3000 and am very happily surprised by what you have found. I will purchase some quite soon. Add my vote for a memory tuning tutorial, please and thank you.
Thanks a lot for taking the time and making this video, I am planning to go for a ryzen build, and this would be quite useful.
Thanks for including the lower end graphics cards! I might tune up next time I upgrade my card
This is an amazing video and reporting, Steve. Thank you so much!
THANK YOU for the hard work you put into making this video!!
I plan on surprising my friend with a PC build in November so he can enjoy Forza Horizon 3 and 4 (among other non-PS4-exclusive games) in 4K at 60fps on his 60Hz 4K tv. Xbox One X would be a waste of money for him even on Black Friday.
I'm getting the 3700X for the build. Now I have an even better understanding of what RAM to get for Ryzen 3000.
Those are some impressive gains! Nice work!
Excellent video. When Linus starting talking about FCLK and sub-timings I hopped right on and started messing around, and basically came to the same conclusion that you just did. FCLK can do something if you are running at 2133, but once you get to 3000 (1500MHz FCLK) then pumping it up more generally doesn't do a whole lot. Changing the sub-timings has a bigger effect and it seems like a lot of memory can handle that quite handily.
Even normal 2133 memory can usually be overclocked to at least 2666CL15 just by pumping the voltage up to 1.35V. In fact, most memory kits are just binned 2133 memory running at a higher voltage. Its not as though the chip makers are actually producing these higher-spec chips. Running faster and tighter is 80% voltage and 20% silicon lottery.
I do have to add a note of caution, though. Messing with sub-timings might seem stable but it IS playing with fire to some degree. It isn't as guaranteed stable as people think, and it also tends to make systems a bit more vulnerable to memory-based attacks like RowHammer, particularly if you don't give the RAS cycle enough time to fully recharge the row.
-Matt
Can you point which timings had the biggest impact on performance when manually overclocking? That manually overclocked DDR4 3000 results are really impressive. Great vid btw!
Nice! So I made the right pick a couple of weeks ago, when my brother and me ordered 64gig DDR4 3000 CL16.
I was a little nervous when I clicked on this video. ^^
Interesting because, yes, you're almost never CPU bound in most gaming scenarios--but all testing is in CPU bound scenarios. Very comprehensive testing!
Holy moly, That's some serious benchmarking, Great Job Mr. Steve.
Awesome vid Steve!
Thanks Steve. You just saved me some money on my next build!
Great work Steve, i would love to see a video about Dram Calculator , keep up the good work.
Another great video! Thank you!
I know you get a lot of request, but a video on how to best make use of that Ryzen memory calculator tool would be amazing.
Either way, keep up the great work! Hope all else is going well for you guys 👍🏼😁
Fantastic work. Informative as always. Thanks dude.
Wow thanks for the effort. Good to hear I can settle for some budget ram when I get the 3600 :)
My favorite hardware channel. Scientific depth of GN combined with the editing and entertainment skills of LTT. Cheers
That is some impressive gains. Did not expect that at all.
Great work Steve. Tuning the timings is like putting turbos on a Golf
Great video Steve! Yes, I would love to see a video explaining how to setup manual DD4 memory timings (Ryzen or otherwise). I haven't seen a good video on how to manually tune DDR4 memory timings and really want to try it!
I would have liked to see other games (namely hitman, given its odd nature) but 120 benchmark runs per game is just absolutely incredible and and you will hear no complaints from me. Well done guys, a masterclass on benchmarking and testing.
A great video, thank you. I particularly appreciate the fact that you included the RX 580 as we do not all have the latest, greatest GPUs. Liked and subscribed!
Absolutely awesome! Exact content I wanted! More videos like this please,
120 benchmark runs per. Thats amazing. Thanks for all your hard work!
Thanks for the dedication.
Love your work, keep it coming !
Thank you so much for your time and the video!!
Excellent work :)
Good job mate! Always quality content on your channel