Wednesday, August 31st 2022
AMD Confirms DDR5-6000 as "Sweetspot" Memory OC Frequency for Ryzen 7000
AMD in its Discord AMA confirmed DDR5-6000 to be the "sweetspot" memory overclock for its upcoming Ryzen 7000 "Zen 4" processors. A sweetspot frequency in AMD jargon is an inflection of performance, stability, cost, and ease. For the very first Ryzen, this was DDR4-3200. For Ryzen 2000, it was DDR4-3400. For the Ryzen 3000 "Zen 2," it climbed to DDR4-3800, the Ryzen 5000 "Zen 3" it was DDR4-4000.
At the architectural-level, it is usually the highest frequency where FClk, UClk, and MClk maintain a 1:1:1 ratio, before having to engage dividers that impact performance, making it a point of diminishing returns for investing in faster memory modules. AMD's Robert Hallock, leading the Discord AMA, recommended that FClk be left untouched at "Auto" for the best results, and overclockers look for an Auto:1:1 ratio for the FClk, UClk, and MClk. As with both AMD and Intel now, the highest frequencies are possible only with one single-rank DIMM per memory channel (1DPC), and memory overclocking yield lower results with dual-rank DIMMs, or two DIMMs per memory channel. Among the AMD EXPO-certified DIMMs announced over the past few days, some do engage memory clocks beyond DDR5-6000. It would be interesting to see how they affect the "golden ratio" for Zen 4.
Sources:
Wccftech, VideoCardz
At the architectural-level, it is usually the highest frequency where FClk, UClk, and MClk maintain a 1:1:1 ratio, before having to engage dividers that impact performance, making it a point of diminishing returns for investing in faster memory modules. AMD's Robert Hallock, leading the Discord AMA, recommended that FClk be left untouched at "Auto" for the best results, and overclockers look for an Auto:1:1 ratio for the FClk, UClk, and MClk. As with both AMD and Intel now, the highest frequencies are possible only with one single-rank DIMM per memory channel (1DPC), and memory overclocking yield lower results with dual-rank DIMMs, or two DIMMs per memory channel. Among the AMD EXPO-certified DIMMs announced over the past few days, some do engage memory clocks beyond DDR5-6000. It would be interesting to see how they affect the "golden ratio" for Zen 4.
21 Comments on AMD Confirms DDR5-6000 as "Sweetspot" Memory OC Frequency for Ryzen 7000
1733MHz default Fabric sounds like a big step up over 1066MHz but that probably just from DDR5 JEDEC speeds being higher. Robert using 2000MHz FCLK as an example suggests not much has changed and it's still the same slow-ass Fabric links in there, minus generational improvements for slightly higher clock. And optimizations to lessen the unsync penalty since it's the norm now.
Although, Fabric isn't strictly bound to 1:1/2/4 ratios like Intel IMC is, so maybe that's what Robert meant by Auto FCLK - you just let Fabric run as high as it can. iirc on AM4 there was some performance improvement to keeping Fabric fast even if desynced; it just didn't get much attention because no one ran desynced daily.
wrong
for Ryzen, 4 ranks has always been best, the difference in performance between 2 vs 4 ranks can be almost 10% in games, plenty of benchmarks / reddit / famous German overclocker videos show that
now, finding dual rank DDR4 isn't easy (not sure about DDR5), both 8 and 16GB sticks are coming as single rank nowadays, so having 2x8 or 2x16GB will give you dual rank and you're gonna have LESS performance than a 4 stick 4 rank system. A 4 rank 2 sticks would be better yes, but hard to find.
so for the past 1-2 years on Ryzen 3000/5000, the best doable performant build on Ryzen has been 4x8GB single rank, running fclk 1:1, between 3600 and 4000 MHz with the tightest timings possible
edit :Um no, 2 DIMM are better than 4, dual ranks are better than single, so in perfect condition 2 DIMM dual ranks DDR4 3600Mhz PC4-28800 CL14 give highest performance.
That is highly misleading." Sweetspot" can't be something a lot of processors can't achieve, and these numbers aren't guaranteed or even commonly achevable. Ryzen 3000 with Infinity Fabric at 1900 MHz and Ryzen 5000 with Infinity Fabric at 2000 MHz were rare.
There was a company, now gone, that pre-binned the processor for you, that claimed that only 14% of 3900X could achieve IF 1900, and that was Zen 2 with highest success rate. I guess they could be exaggerating to justify the high price, but there were a lot of reports of people that couldn't achieve these "sweetspots" with top of the line processors and very expensive RAM.
Here is the Zen 2 slide
3733 is sweet spot not 3800 as stated in the article. Some could do 3800 but it was a lot harder.
For Zen 3 they didn't have such a chart but they did have this slide
Where they state 4000 for Zen 3 is about the same difficulty as 3800 for Zen 2. That is not at all like saying 4000 is the sweet spot. For Zen 3 3800 was considered the sweet spot.
Historically AMD have used the sweet spot to describe the highest speed that can be achieved by the majority of samples so if AMD are describing 6000 as the sweet spot then that can probably be done by the vast majority of samples rather than it being like 4000 on Zen 3 where only a few samples can do it.
"Performance sweet spot" just meant that higher DDR speeds surely lost the 1:1 IF sync for most of users, and lost quite a bit of performance.
Their sweet spot has always been an inflection point of cost, performance, % of samples that can hit that speed.
The likening of 4000 on Zen3 to 3800 on Zen2 is to illustrate that some samples can do it but you need a good sample. Otoh 3733 is far easier to achieve on zen2 and 3800 on zen3 hence why AMD picked 3733 as the sweet spot.
Have you not noticed over the last 5 years when it comes to Ryzen AMD have tried very hard to be realistic with the claims they make. They actively try not to mislead people. This sweet spot stuff is just another example, DDR5 6000 will probably be pretty easy to achieve unless you get really unlucky. Faster will get harder quicker so I expect 6400 to be like 3800/4000 was for Zen2/3, doable but you need a bit of luck.
EDIT. In any event AMD self described the Zen 2 sweet spot as 3733 in that slide so the article stating "For the Ryzen 3000 "Zen 2," it climbed to DDR4-3800" is simply not true and should really be corrected. Given that the claim for Zen 3 is also untrue because AMD likened 4000 to 3800 which would also mean that 4000 is not the Zen 3 sweet spot as claimed in the article.
Remember when Zen 3000 CPUs didn't achieve their boost frequencies? No reviewer even mentioned this, and it took a wide poll by an enthusiast (Der8auer) which showed that majority of users didn't achieve processor's bacic spec:
"The worst is the Ryzen 9 3900X, which only reaches the Boost frequency in 5.6% of cases."
Has anyone done anything similar for IF frequency? No, and the reason is AMD never promised anything in this regard. Some even defended AMD with boost frequency scandal, the spec was written as "up to" in some material. So not guaranteed, apparently.
A few months after launch they locked PBO boost on 5600x to max 4850MHz from more than 5100Mhz on a single core. Silently with no explanation why.
Besides my point is the article says Zen 2 was 3800 sweet spot when this is explicitly untrue, AMD said 3733 was the sweet spot for Zen 2. AMD didn't really mention a sweet spot for Zen 3 but they did state that 4000 on Zen 3 was akin to 3800 on Zen 2 which would mean above the sweet spot since 3800 is above the sweet spot so the article is incorrect to refer to those memory speeds as the sweet spots for Zen 2 and Zen 3.
But what can be a clear defect for a reviewer can be a product hard to replace for a normal buyer, especially if we're talking about the stat that has no clear guarantee.
That's why you get launches of products that have very high rate of failure or are even DOA in high percentage - but no reviewer mentiones that. Until it is publically known, and then they candidly admit they had to RMA the product several times before they could do a review.
So I wouldn't be surprised if absolutely every reviewer's Ryzen achieved the "sweet spot" IF frequency.
I mean, no reviewer even mentioned trouble to achieve Ryzen 3000 boost frequency, and then the poll showed that it was more of a fluke if you actually got one that did that.
TPU could not get their 3700X sample above 1800 fclk when the 'sweet spot' was 1867.
None of TechSpots Zen 3's managed 2000 fclk although all could manage 1900 fclk.
2 of TechSpots Zen 2's managed 1900 fclk but 2 others topped out at 1800 fclk.
This is just from a cursory glance.
Again though, my point is that this:is false for Zen 2 and Zen 3. Zen 2 sweet spot was 3733 according to the AMD slide. And AMD never really said anything about Zen 3 sweet spot other than that DDR4-4000 would be about as likely as Zen 2 hitting DDR4 3800 1:1 so above the sweet spot.
So in the end I bought a 4x8GB Crucial 3600C16 and overclocked it to 3800C16 while lowering some timings, and it has been rock solid on all memory benchmarks I throw at it.
Just checked now and a decent DDR5 6000 kit is going for 290 USD on Amazon, ~110USD more than what I paid for the Crucial...