What is current sweetspot for DDR5

[nomdeplume]

There seems to be no end of confusion surrounding every aspect of PC gaming nowadays. With the unoffical hidden retail holiday "Christmas in July" fast approaching there is always a chance of prices slightly dipping as models with new stickers and a new marketing budget to match hit warehouses.

With how fast TeamGroup (among a few other non-notables) is bottoming out already it seems the least desirable option at any speed or latency.
Friends don't let friends buy Corsair RAM.
Crucial only appears to have adapted SKU for OEM business and workstation PC.
G-Skill don't seem to be all that popular, or low latency, or OC'able, or faster than DDR4.
24GB seems to be stagnating and rare enough to cost more than 32GB pairs.
Actually seeing a gaming mobo hit 7200 is a rarity by most accounts.


Considering DDR5 6000 CL 36 is not faster than pretty easy to find DDR4 3600 with almost a third lower latency. Considering DDR5 has yet to bring anything to gaming performance.
Is there a sweetspot for DDR5 currently?

 

[JWNoctis]

For AMD, I think it's less of a DDR5 sweetspot, than FCLK sweetspot and whatever memclk it would happen to imply for minimum added latency. Current IOD and datapath infrastructure is bottlenecking memory bandwidth, and that's one of the reasons why DDR5-4800 C40 on AM5 is inferior to DDR4-3200 C20 on AM4, for instance: Exact same effective maximum bandwidth with much more latency. There's a lot of room for improvement there.

No idea what goes on with Intel these days myself. I wonder whether it's in any way alike.

At least DDR5 6000 C30 kits are common, and need not cost an arm and a leg when compared to more exotic kits.

 

[oxrufiioxo]

6000-6200 low latency for Zen 4 and as high as possible on Intel 7200 shouldn't be difficult for a 14th gen K series sku with a half decent Z790 board. Maybe 6800 for 13th gen and 62-6400 on 12th gen.

I'd probably at least try to get 8000 working on 14th generation though.

So it just depends on what cpu/mobo you're looking at.

@ir_cow what say you lol. He'd be the best one to ask imho.

 

[Princess Garnet]

I hope I can sort of piggyback in on this as I'm sort of wondering the same things, given I might be jumping onto a DDR5 platform after the next generations of CPUs launch (likely on AM5, but I won't finalize any decisions until performance and pricing is known).

From what I know, 6000 MHz/CL30 is often the sweet spot for AMD's Zen 4 CPUs, and Intel's LGA 1700 platform can generally tolerate a bit more (?), especially with the latest chipset series that launched with Raptor Lake, but using some combination of four DIMMs over two, and dual rank over single, makes these sweet spot speeds harder to attain. These factors always made a memory configuration more difficult, but it seems like it's at the point with DDR5 you have to choose between higher capacity or speed more than ever before.

Contrast that to later DDR4 platforms, where I'm using 64 GB (4x 16 GB, all dual rank DIMMs at that, at 3,600 MHz and CL 16), and it's been fine. I've been wondering what to expect if I want to go up in capacity and still hit the sweet spots on a DDR5 platform. It's hard to guess since we don't know if Zen 5 will increase this sweet spot for the IF clock over Zen 4, and/or if future 800 series chipset platforms, new BIOS, and new RAM will make this "easier" to attain. Right now, it seems like going back to two DIMMs and trying to get 2x 48 GB at as close to 6000 MHz+ is my most practical option for a higher capacity/speed balance. But I'm also interesting in hearing from others on DDR5 (especially AM5, and especially those using higher capacities, like 64 GB+).

By the way, I wouldn't equate RAM performance just based on latency. Yes, that's very important, but bandwidth also matters. We've seen Alder Lake gain performance, even in gaming but more so outside it, with DDR5 over DDR4, and I imagine Zen 4 saw part of its performance increase over Zen 3 for the same reason. If only latency mattered, we could still be on DDR1/2 with much slower speeds and not lose performance, but that's not the case.

 

[ir_cow]

@ir_cow what say you lol. He'd be the best one to ask imho.

I wish it was a simple answer because the sweet spot can either be best for "plug n play" or best for performance within reason. Also the CPu generation changes this as well...

But I would assume plug n play without any additional tweaking is what 99% of people are after.

For Intel 12th gen plug n play is generally DDR5-6000 for Z690 Motherboards with 6400 MT for Z790

13/14th gen its 6400-7200 using a Z690 chipset and 6800-7200 for Z790.

For AMD the best performance, it's highest MCLK matching UCLK (1:1) ratio. Bonus points if you raise the FCLK as well.

Basically Ryzen 7000 series would be DDR5-6000 if you just want to rock out and 6400 (FLCK 2133) if you have lots of time on your hands to tweak stuff.

AMD is more sensitive to DRAM timings in general, Intel is about speed!... Only to a point though.

Look at one my lastest reviews and a eSports game like CS2 has a 1.9% gain with 8200 over 6400. This is one example of a bit of overkill. I would rather save $100 and spend that on a nicer video card over faster memory when restricted by a budget.



AMD has a distinct gap between 5600 and 6000 for a number of games. But this can be offset with lower Primary Timings. Like 5600 CL26 is pretty sweet. Working on including that in my next review.



Edit: just for clarification, everything said above is for single ranked memory. Dual rank has lower values.

 

[Braegnok]

I seem to have too much time on my hands.

 

[ir_cow]

I seem to have too much time on my hands.

View attachment 351087

That's because you are a memory overclocking god. Also higher voltages dont scare you. I don't have a spare CPU

 

[tabascosauz]

There seems to be no end of confusion surrounding every aspect of PC gaming nowadays. With the unoffical hidden retail holiday "Christmas in July" fast approaching there is always a chance of prices slightly dipping as models with new stickers and a new marketing budget to match hit warehouses.

With how fast TeamGroup (among a few other non-notables) is bottoming out already it seems the least desirable option at any speed or latency.
Friends don't let friends buy Corsair RAM.
Crucial only appears to have adapted SKU for OEM business and workstation PC.
G-Skill don't seem to be all that popular, or low latency, or OC'able, or faster than DDR4.
24GB seems to be stagnating and rare enough to cost more than 32GB pairs.
Actually seeing a gaming mobo hit 7200 is a rarity by most accounts.


Considering DDR5 6000 CL 36 is not faster than pretty easy to find DDR4 3600 with almost a third lower latency. Considering DDR5 has yet to bring anything to gaming performance.
Is there a sweetspot for DDR5 currently?

Any mention of "sweetspot" and anything non-Hynix goes straight out the window. It IS the sweet spot, and it's not expensive. There STILL isn't anything non-Hynix worth mentioning, so mentioning any of these brands is pretty pointless - ensuring Hynix is inside is literally all that matters if you want to actually do stuff with your DDR5.

• As far as I can tell Team has the same access to Hynix as everyone else.
• Corsair also has their Hynix bins, but in typical Corsair fashion also plenty of other garbage. They also have some high speed Micron above 7000...but it's Micron.
• G.skill has plenty of Hynix in all capacities.
• Crucial has not been relevant for years at this point since they retired Ballistix. Not that you'd want their DDR5 anyway, it's not Hynix.

Hitting 7200 is not a rarity, it's just pointless on AMD unless you know what you're doing with your Hynix kit and can make it to or past the 8000 mark. And Intel has a temperamental memory controller so you really better know what you're doing at those speeds.

• 2x16GB is single rank Hynix A or M 16Gb
• 2x24GB is single rank Hynix M 24Gb
• 2x32GB is dual rank Hynix A or M 16Gb
• 2x48GB is dual rank Hynix M 24Gb

Have not kept up in a little bit, but afaik the absolute cheapest Hynix is something like [some] 6000CL34 or 5600CL28 (maybe CL30?) bins. Anything cheaper or slower than that and I think you might be at risk of getting Micron. Samsung seems to be easy to avoid, just avoid flat timings (36-36-36)

 

[Princess Garnet]

AMD has a distinct gap between 5600 and 6000 for a number of games. But this can be offset with lower Primary Timings. Like 5600 CL26 is pretty sweet. Working on including that in my next review.

View attachment 351084

This is what I understood, similar to how 3,200 MHz to 3,600 MHz would show a nice difference on AM4, but not an end of the world if you had the slower speed.

The X3D CPUs, however, seem to be less intolerant of a slight loss of RAM speed (reasoning being, they are that much less reliant on fetching to RAM due to the larger cache). And your chart indeed shows a 7950X. I presume the 7000 series X3D would see a smaller difference, yeah?

Just trying to tell myself that if I have to settle for a single speed grade lower (with an X3D CPU), so to speak, that it won't be that bad I guess.

 

[nomdeplume]

I would posit the common sense exhibited by sales of B760 in combination with K processors on the Intel side is due to the lack of a single: PCIe 5 GPU, gen 5 NVMe offering any benefit whatsoever, and finally DDR5 even at rates unattainable on this chipset fall well within a rounding error of DDR4's highest speeds the D4 versions simply run without issue. Z boards are for show or eventually for benchmarking and OC'ing.

Getting slightly more technical than plug and play. As we should on this site where set it and forget it might be a better term for stability. All DDR5 being dual channel on the RAM side and single channel on the cpu side does induce the exact complication Princess Garnet mentions. Defining worth, much less a sweetspot, is incredibly taxing.

I know nothing of AMD. It's complicated is the only impression I've garnered on it's relationship with RAM.

Edit: The amount of responses focused on AMD is interesting. I fall into the cram hard for an exam and follow through by acing it category of setup more than the long term knowledge retention to continually provide support one. Going to spend a bit more time learning about the X3D and AMD side.

 

[ir_cow]

Apparently KLEVV is a division of SK Hynix. I'm ashamed I didn't know of this until recently. So all KLEVV has to be SK Hynix.

Te X3D CPUs, however, seem to be less intolerant of a slight loss of RAM speed (reasoning being, they are that much less reliant on fetching to RAM due to the larger cache). And your chart indeed shows a 7950X. I presume the 7000 series X3D would see a smaller difference, yeah?

Never tested this myself, but it seems the X3D ones are less affected by latency due to the CPU cache from other people's benchmarks. Less by how much exactly over the none X3D I do not have any data on this.

 

[JWNoctis]

Edit: The amount of responses focused on AMD is interesting. I fall into the cram hard for an exam and follow through by acing it category of setup more than the long term knowledge retention to continually provide support one. Going to spend a bit more time learning about the X3D and AMD side.

That's probably because AMD often marketed DDR5-6000 as the "sweet spot" of Zen 4, using the exact same words, and it is.

They tried to market DDR4-4000 using the same terms in the previous generation, but that did not quite work out when, apparently, very few non-APU Zen 3 processors could reach 2000MHz IF and stay stable.

 

[jarablue]

I tweaked my teamgroup UD5-6000 32GB DDR5 modules. I got as low as 61 for latency then just gave up all the tweaking. Flashed my bios to the newest version, enabled xmp and do not notice one single bit of difference. Unless you are benchmarking, really. No gains at all here for me. Going from 78-74 to 61 in latency, I just do not notice anything.

 

[ir_cow]

@jarablue mostly your 1% lows will change, but of course that is per game and the graphical settings you are using.

 

[#22]

How it is recently with more than popular 32GB? Few years from now popular will be upgrading owned 2x16 with another pair. How fast 4x16 tend to go and how it compares to 2x32?

 

[GerKNG]

Hynix A-Die 6000 CL30.

 

[jaszy]

I run an early 5600 CL46 JEDEC 10 layer Teamgroup kit (M-die 16GB) with manual timings on a z690 Tomahawk. (12900k)

Seems to bug out in sleep mode at CL30 with any frequency, but I'll just assume it's a MSI bios issue (something to do with my Kit base timings not going below 32?)...

6400 cl32 is unironically more stable. had it locked in for a year.. never BSOD'd.

Apparently KLEVV is a division of SK Hynix. I'm ashamed I didn't know of this until recently. So all KLEVV has to be SK Hynix.

Not saying that won't be true for the foreseeable future, but Crucial (Micron division) used to bin 8GB Samsung B-die If im not mistaken ("Sport LT" generation). Who knows what could happen.

 

[Bill_Bright]

Is there such a thing as a sweetspot for RAM (other than "amount" of RAM)? I just don't see how one can be determined since it will vary with every task you perform, not to mention every program or programs you run, and the OS too.

Plus other variables are at play here including the capability of the CPU, GPU, motherboard bus speeds, and drive I/O, again dependent on the task or tasks being performed.

I say the sweetspot, in general, is the most RAM your budget allows - with a wide variety of exceptions to that thrown in based on how you use your computer and what you run. With very few exceptions, I would rather have more RAM than faster RAM. And frankly, the only exception to that I can think of now is if my motherboard's amount capacity is already reached and it will not support more RAM. Then I'll take faster. Now whether I will "see" any performance increase is another mater. My guess is "no" - except on paper/benchmarks.

Clear as mud, right?

 

[sneekypeet]

I'm gonna be pretty blunt here with a few things.

If looking for DDR5, go Patriot. They tend to have class leading performance at whatever speed and density its sold at.
Also no revisions (Corsair) and whitewashing ICs like Klevv does and claims its how they keep bins seperated. However, Klevv started when they took the best binned chips they had in DDR3 (iirc), and then built kits with them. Not too sure that is the same today.

As to speed and timings, as noted, it depends greatly on CPU, board, asnd BIOS (ME/AGESA).
I like IR Cows defined speeds, as it is what I see across many users as well.

If you want to see what DDR5 can do with many less boundaries, try an APEX, Tachyon for Intel for 9000+, or maybe a Gene and an 8500G and swing for 10,000+ at C42.

 

[ShrimpBrime]

Is there such a thing as a sweetspot for RAM (other than "amount" of RAM)? I just don't see how one can be determined since it will vary with every task you perform, not to mention every program or programs you run, and the OS too.

Plus other variables are at play here including the capability of the CPU, GPU, motherboard bus speeds, and drive I/O, again dependent on the task or tasks being performed.

I say the sweetspot, in general, is the most RAM your budget allows - with a wide variety of exceptions to that thrown in based on how you use your computer and what you run. With very few exceptions, I would rather have more RAM than faster RAM. And frankly, the only exception to that I can think of now is if my motherboard's amount capacity is already reached and it will not support more RAM. Then I'll take faster. Now whether I will "see" any performance increase is another mater. My guess is "no" - except on paper/benchmarks.

Clear as mud, right?

Yes, totally a sweet spot for ram speed and timings, depending on cpu generation and capability of course.

In example, a 12400F has a locked system agent voltage. Same goes for 14100F and in between. You are limited to this voltage to find "sweet spot". On these systems, typically up to, but not often exceeding 6400mhz [3200mhz effective] with whatever timings you can get within that spec.

If we are talking DDR4 14th gen and lower, 4000mhz [2000mhz effective] at least CL16 on any system would be a solid sweet spot.

DDR5, now becoming mature is a little more difficult to produce a solid suggestion, but we see a lot of systems (Intel mostly??) Perfectly capable of 8000mhz [4000mhz effective] with a cas latency as right as 36 or lower.

Of course the kits for higher speeds cost more than the performance given at the rated speeds.

Sweet spot on DDR3 would be at least 2000mhz [1000mhz effective] or higher with CL8 or tighter.

DDR2 would be a minimum of 800mhz [400mhz effective] CL4.

DDR1 at 400mhz [200mhz effective] CL2 would be a sweet spot.

Based on averages I've seen through the years off my head.

These figures based off personal experience of running over 170 different processors and over 195 different video cards. I do not have links to provide at this time.

 

[Bill_Bright]

My bold added:

Yes, totally a sweet spot for ram speed and timings, depending on cpu generation and capability of course.

Which is exactly the point I made above when I said (and you quoted),

other variables are at play here including the capability of the CPU

The point is, the sweet spot is not something that can be answered in a forum post without a whole host of specific information about that specific computer, the specific tasks that will demanded from it, and how the user uses it. Every computer is unique.

 

[dgianstefani]

There seems to be no end of confusion surrounding every aspect of PC gaming nowadays. With the unoffical hidden retail holiday "Christmas in July" fast approaching there is always a chance of prices slightly dipping as models with new stickers and a new marketing budget to match hit warehouses.

With how fast TeamGroup (among a few other non-notables) is bottoming out already it seems the least desirable option at any speed or latency.
Friends don't let friends buy Corsair RAM.
Crucial only appears to have adapted SKU for OEM business and workstation PC.
G-Skill don't seem to be all that popular, or low latency, or OC'able, or faster than DDR4.
24GB seems to be stagnating and rare enough to cost more than 32GB pairs.
Actually seeing a gaming mobo hit 7200 is a rarity by most accounts.


Considering DDR5 6000 CL 36 is not faster than pretty easy to find DDR4 3600 with almost a third lower latency. Considering DDR5 has yet to bring anything to gaming performance.
Is there a sweetspot for DDR5 currently?

6000/36 is pretty bad RAM, it's better than 5600 JEDEC, but that's a low bar. 6400 is doable on Zen 4, 7200 for 13/14th gen K series.

Mostly depends on your motherboard for Intel, you're gonna need a Z790 dual DIMM board.

Almost any Raptor Lake K CPU with a Z790 board will hit 7200 with XMP voltages. 1DPC boards (OC or ITX) will hit 7600. With some effort you can stabilize 8000 MT or even 82-8400 MT on a two DIMM board with daily voltages, if your CPU memory controller can handle it. Going ITX is a good way to avoid spending stupid money on a two DIMM special edition ATX board.

As for Zen, 6400 MT is your best bet, 6200 will work on most CPUs and 6000 is safe, but it's ideal to get a 7200 MT + kit since it will be almost always be Micron A die, and manually set voltages/timings for best performance.

If you want more than 32 GB of RAM this will limit performance for both Intel/AMD.

@ir_cow

This likely won't change until CAMM2, Intel's DDR5 only MC, or Zen 6.

Once you get below 50 ns AIDA memory latency on current platforms, there doesn't seem to be much benefit for most games.

With X3D, you'll get more performance from a mobo with an external clock generator to OC the CPU without disrupting other components.

Without X3D on Zen, get X3D for gaming, for other memory intensive workloads get Intel.

If it's Karhu stable, it's stable.

For tuning I use the OCCT memory stability test to dial in what is likely to be stable and what isn't, since full memory tests take too long for that stage.

For a quick and dirty "OC", manually set voltages and tune tRFC and TREFI, keep other XMP values.

Here's what I run daily, the voltage readouts are wrong since I am using an ASUS board with a certain BIOS setting.

This is likely an M die kit.

BCLK OC, so MT is different from what you will achieve with 100 BCLK.

 

[ShrimpBrime]

My bold added:

Which is exactly the point I made above when I said (and you quoted),

The point is, the sweet spot is not something that can be answered in a forum post without a whole host of specific information about that specific computer, the specific tasks that will demanded from it, and how the user uses it. Every computer is unique.

You had asked if there was even a known sweet spot for RAM.

I generously gave the sweet spot figures since DDR1 without deceit and exclaimed DDR5 isn't fully matured yet to have this fortified conclusion.

For a specific system, other than a general question, I fully agree with you and see my mistake there. My apologies.

 

[nomdeplume]

Not ignoring or abandoning this thread. AMD was just something I was less considerate of after hearing of their issues with RAM and RAM training.

 

[ir_cow]

Not ignoring or abandoning this thread. AMD was just something I was less considerate of after hearing of their issues with RAM and RAM training.

Issues are like bugs found by the community. I think AMD has certain design and people expectations are different.