AMD Renoir Powers the World Record of DDR4 Memory Overclock: 6,666 MHz

[Raevenlord]

A new HWBot entry has proven what some thought impossible years ago: AMD apparently features the best memory controller in the x86 consumer space. A user going by the alias Bianbao XE achieved a 6,666 MHz frequency on a single stick of Crucial Ballistix Max. The stick's original rating is for a mere 2,666 MHz - doesn't that put things in perspective?

Another thing that puts things into perspective is that the support for such an overclocking feat was a ROG Strix B550-I Gaming motherboard (min-ITX means smaller tracing distance between CPU and memory, and thus higher signal integrity) paired with none other than AMD's Ryzen 7 4700GE 'Renoir'. The APU was underclocked and overvolted - a technique that aimed to increase stability of the memory controller whilst also reducing operating temperatures (balancing the higher voltage and lower frequency). Of course, memory timings were loosened to achieve this feat (timings of 30-27-27-58 aren't what you'd usually like to see), but then again, this wasn't meant to power the utmost memory performance - only the highest frequency. And that was definitely achieved.



View at TechPowerUp Main Site

 

[john_]

Proof that Renoir is a demonic chip

 

[Mindweaver]

That's the sign of the Beast... or of a Beast! haha

 

[londiste]

What is the limiting factor for frequency these days, memory controller or the RAM chips themselves?

 

[Vayra86]

Is this the number you get when you raise the voltage 'hella' high?

That is probably the coolest record in 2020

 

[maxitaxi96]

Renoir is monolithic Zen 2 so kinda makes sense... but what the hell has this guy done to the infinity fabric to get these clocks! A 1:4 Ratio? Is that even possible?!

 

[Vya Domus]

What is the limiting factor for frequency these days, memory controller or the RAM chips themselves?

Doesn't the memory controller have to run by definition at the same frequency as the memory ?

 

[bonehead123]

HE is here.. are YOU ready ???

 

[AnarchoPrimitiv]

Is this purely to do with the monolithic design? Or could AMD have made some architectural improvements to their memory controllers since the release of Zen2 and this might be a harbinger of increased memory capabilities with Zen3 (I know 6000Mhz+ isn't going to be the standard, I'm talking about perhaps 4000-5000Mhz being much easier to achieve at lower voltages and with increased stability).

 

[Vya Domus]

I don't think it's that hard to make a memory controller this fast, it's a matter of whether or not you need it.

 

[londiste]

Doesn't the memory controller have to run by definition at the same frequency as the memory ?

It does. But looking at the news every once in a while about new records by a few or maybe 20-30 additional MHz achieved I am not sure which one is the bottleneck here. The memory modules seem to be handpicked every time as well, often by manufacturers.

Is this purely to do with the monolithic design? Or could AMD have made some architectural improvements to their memory controllers since the release of Zen2 and this might be a harbinger of increased memory capabilities with Zen3 (I know 6000Mhz+ isn't going to be the standard, I'm talking about perhaps 4000-5000Mhz being much easier to achieve at lower voltages and with increased stability).

Lower voltages and stability you mean RAM voltages, right? That is not up to memory controller but RAM chips. I would suspect DDR5 will take over before we see another major improvement in DDR4.

 

[Vya Domus]

I am not sure which one is the bottleneck here.

I imagine it has to be the memory, it's one thing to have a tiny piece of silicon on a cutting edge node run at 6 Ghz and another to have a a bunch of huge memory chips connected together to the same.

 

[Lionheart]

Wish my RAM did that, can't even get away from 2133mhz

 

[Chrispy_]

AMD could absolutely murder the Intel ultrabook market with a 15-20W Renoir+ running Vega12 and LPDDR4X at about 5GHz.

There are rumours of a new APU that has more Vega CUs, but nothing concrete yet.

 

[aQi]

 

[KarymidoN]

AMD could absolutely murder the Intel ultrabook market with a 15-20W Renoir+ running Vega12 and LPDDR4X at about 5GHz.

There are rumours of a new APU that has more Vega CUs, but nothing concrete yet.

don't believe LPDDR4X would go that high, i mean, not enough power, thats LP = Low Power

 

[R0H1T]

don't believe LPDDR4X would go that high, i mean, not enough power, thats LP = Low Power

It does actually go pretty high & when you're saying LP you must also remember memory speeds are actually just 2500Mhz. The only issue is that there's no LPddr4x at such speeds atm so it's not gonna happen, especially since LPDDR5 is already on the market now!

 

[Frick]

AMD could absolutely murder the Intel ultrabook market with a 15-20W Renoir+ running Vega12 and LPDDR4X at about 5GHz.

There are rumours of a new APU that has more Vega CUs, but nothing concrete yet.

All you need is the OEMs being on board, which apparently is a hard thing to do.

 

[Mayclore]

Proof that Renoir is a demonic chip

I mean, AMD has been giving Intel hell lately.

 

[londiste]

I imagine it has to be the memory, it's one thing to have a tiny piece of silicon on a cutting edge node run at 6 Ghz and another to have a a bunch of huge memory chips connected together to the same.

Nothing runs at 6GHz. 6666 MT/s. Bus clock is half the transfer rate at 3333 MHz , memory speed is a quarter of that - at around 833MHz in this case.

Problem with Ryzens has not been the memory controller itself, it's the memory bus speed being tied to IF speed and the latter has been the limiting factor. Zen2 and Renoir benefit from dividers and a bunch of tweaks to get these out of the way. Don't get me wrong, Renoir's memory controller improvements are awesome. But I am seriously curious where the actual bottleneck is these days.

 

[tabascosauz]

Nothing runs at 6GHz. 6666 MT/s. Bus clock is half the transfer rate at 3333 MHz , memory speed is a quarter of that - at around 833MHz in this case.

Problem with Ryzens has not been the memory controller itself, it's the memory bus speed being tied to IF speed and the latter has been the limiting factor. Zen2 and Renoir benefit from dividers and a bunch of tweaks to get these out of the way. Don't get me wrong, Renoir's memory controller improvements are awesome. But I am seriously curious where the actual bottleneck is these days.

I think the IMC and the ICs are actually pretty close at this point, but in terms of raw frequency Rev.E probably still comes out ahead. You'll still have to push both of them quite hard. In practical terms, the voltage you're willing to run on both the IMC (VCCSA on Intel though I haven't heard of VCCSA degradation yet, VSOC on AMD) and DRAM (and degradation you're willing to risk to both) is your bottleneck. Rev.E scales on frequency like a mofo; you just have to ask yourself how long you want the memory to last.

It says so right in the article. The 4700GE is underclocked to a silly extent, and presumably VSOC is overvolted to bolster the memory controller, the latter being in and of itself is a major longevity concern.

B-die's still king for actually stable daily performance between 4000 and 5000MT/s, however. Rev.E doesn't go nearly tight enough on most timings.

Matisse (and presumably Renoir from the news as of late) IMC is getting to be superior to the Coffee Lake IMC, although Comet Lake is markedly improved. Unfortunately for us, Infinity Fabric exists, which reduces uncoupled DRAM on Ryzen to nothing more than validation stunts, 100% of the time.

 

[Chrispy_]

don't believe LPDDR4X would go that high, i mean, not enough power, thats LP = Low Power

It's already officially at 4266MT/s so 5000MT/s is only a 17% increase and LPDDR4X is pretty new, I suspect there are already modules out in the wild that will clock to 5000ish without any problem.

 

[Ruru]

That's indeed hella lot..

*ba dum tss*

Otherwise I haven't been been interesting in extreme overclocking, especially RAM OC, in years, but I still always wonder how long it will take for those to be typical consumer speeds in RAM.

 

[A.Stables]

Timings seem quite nice for that speed tbh

 

[hurakura]

Nice, but what difference does it make in real use?