G.SKILL Showcases DDR5-7000 CL40 Extreme Speed Memory

[Armenius]

While we wait for the mythical 10,000 MT/s parts from Micron, G.SKILL has introduced 7000 MT/s DDR5 memory that will be actually usable in your PC. With a CAS latency of 40 it brings latency down to 11.43 ns.

https://www.gskill.com/community/15...Showcases-DDR5-7000-CL40-Extreme-Speed-Memory

(2 November 2021) – G.SKILL International Enterprise Co., Ltd., the world’s leading manufacturer of extreme performance memory and gaming peripherals, is thrilled to announce the achievement of DDR5-7000 CL40-40-40-76 32GB (2x16GB) extreme speed, passing the Memtest stability test. 7000MT/s memory speed is an exciting milestone, as it was only seen under liquid nitrogen sub-zero temperature cooling not long ago in overclocking records. Accomplished with high-performance Samsung DDR5 components, this extreme speed memory is truly worthy of the G.SKILL flagship Trident Z5 family classification.

DDR5-7000 CL40 – Demonstrating the Ultimate Overclocking Potential of DDR5 Memory
G.SKILL has been dedicated to develop the fastest possible DDR5 memory on the latest 12th Gen Intel® Core™ desktop processors and Intel® Z690 chipset motherboards. Today, G.SKILL is proud to announce the feat of reaching DDR5-7000 extreme speed, while maintaining an ultra-low CAS latency timing of CL40-40-40-76. The memory modules that reached this monumental achievement is built with high-performance Samsung DDR5 components, and has shown to be stable under Memtest. Please refer to the screenshot below:

“We are seeing amazing overclocking potential of DDR5 memory on the latest 12th Gen Intel® Core™ desktop processors and Intel® Z690 chipset motherboards,” says Tequila Huang, Corporate Vice President of G.SKILL International. “DDR5-7000 is an incredible milestone for us, and we will continue to work with our industry partners to develop ever-faster DDR5 memory for PC enthusiasts and overclockers.”

 

[Zarathustra[H]]

I'll admit that while I generally consider myself quite knowledgeable about all things in this hobby, RAM timings are one of the areas that I have never really dug into.

Didn't we used to have CL's down in the low single digits years ago? 2 or 3? 40 seems very high.

 

[lopoetve]

I'll admit that while I generally consider myself quite knowledgeable about all things in this hobby, RAM timings are one of the areas that I have never really dug into.

Didn't we used to have CL's down in the low single digits years ago? 2 or 3? 40 seems very high.

Faster it goes, the looser those have to get.
SUPER basic details:
https://www.maketecheasier.com/what-is-ram-timing/

They're based on transfer rate.

 

[Zarathustra[H]]

Faster it goes, the looser those have to get.
SUPER basic details:
https://www.maketecheasier.com/what-is-ram-timing/

They're based on transfer rate.

Yeah, that part I understand, but it still seems high.

Take this ram - for instance.

They market it as 7000MT's, so I assume it runs at 3500Mhz. A latency of 40CL would be 40/3500 or 11.43ns (units are confusing to me here, as 40/3500 s would be 11.43ms, not ns, but everything in the RAM world expresses this in ns?)

Back in the DDR2 days a common enthusiast RAM variety was PC2-6400 or DDR2-800, running at 400mhz with a CL as low as 4, that resulted in a latency of 4/400, or 10ns.

DDR3-1600 CL8 also wound up with a 10ns latency.

...yet they are claiming this one in the above post at 11.43 is some sort of record?

 

[Armenius]

Faster it goes, the looser those have to get.
SUPER basic details:
https://www.maketecheasier.com/what-is-ram-timing/

They're based on transfer rate.

I imagine that looser timing is also needed to reduce error rate now that DDR5 is fetching two words per clock cycle instead of one with DDR4.

Yeah, that part I understand, but it still seems high.

Take this ram - for instance.

They market it as 7000MT's, so I assume it runs at 3500Mhz. A latency of 40CL would be 40/3500 or 11.43ns (units are confusing to me here, as 40/3500 s would be 11.43ms, not ns, but everything in the RAM world expresses this in ns?)

Back in the DDR2 days a common enthusiast RAM variety was PC2-6400 or DDR2-800, running at 400mhz with a CL as low as 4, that resulted in a latency of 4/400, or 10ns.

DDR3-1600 CL8 also wound up with a 10ns latency.

...yet they are claiming this one in the above post at 11.43 is some sort of record?

It's the lowest stable latency that has been shown for DDR5 so far.

 

[Zarathustra[H]]

What is unclear to me is if all of this really provides any real benefit.

I feel like the developments in system RAM since DDR3 have been mostly marginal.

Did DDR4 really make a practical difference over DDR3? Meh? Will DDR5?

RAM speed never seemed to hold back my quad channel DDR3-1866 i7-3930k.

 

[Lakados]

What is unclear to me is if all of this really provides any real benefit.

I feel like the developments in system RAM since DDR3 have been mostly marginal.

Did DDR4 really make a practical difference over DDR3? Meh? Will DDR5?

RAM speed never seemed to hold back my quad channel DDR3-1866 i7-3930k.

It depends on what you are doing, in some server and workstation applications ram speed makes a very noticeable impact, but current programming methods and memory management algorithms are designed around DDR3, they work around its limitations, and programs are "optimized" accordingly. So going to DDR4 which given its fundamentally the same as DDR3 has benefits of course as it is faster, but unless you are in situations where memory is your bottleneck you aren't really going to notice that, and even then it is pretty minute as modern CPU's and their massive cache sizes easily pick up that slack and the changes in coding methods to take advantage of those larger changes have greatly outpaced changes in programming methods to optimize for the faster ram.
From my limited understanding of the tests and demo's in scenarios that can utilize the extra channels and longer burst lengths the DDR5 shows a lower real-world latency than DDR4 despite the much looser timings.

 

[Armenius]

What is unclear to me is if all of this really provides any real benefit.

I feel like the developments in system RAM since DDR3 have been mostly marginal.

Did DDR4 really make a practical difference over DDR3? Meh? Will DDR5?

RAM speed never seemed to hold back my quad channel DDR3-1866 i7-3930k.

Bandwidth is the primary benefit. Your DDR3-1866 in quad channel theoretically offers up 59.7 GB/s, while this DDR5-7000 in quad channel would offer 224 GB/s. That is nearly four times the bandwidth, meaning you'd be getting similar bandwidth with these DDR5 modules in dual channel compared to your DDR3 modules in quad channel.

 

[SOAREVERSOR]

I'll admit that while I generally consider myself quite knowledgeable about all things in this hobby, RAM timings are one of the areas that I have never really dug into.

Didn't we used to have CL's down in the low single digits years ago? 2 or 3? 40 seems very high.

Back in the DDR 400 era of P4 and AMD 64 you could get down to CL2 or even CL 1.5 with some ICs and a fair amount of voltage. But even then this was really only with two out of four channels filled.

 

[DanNeely]

While we wait for the mythical 10,000 MT/s parts from Micron, G.SKILL has introduced 7000 MT/s DDR5 memory that will be actually usable in your PC. With a CAS latency of 40 it brings latency down to 11.43 ns.

Somewhat unfortunate timing on this comment. Despite the headline numbers that would make you think it was crap DDR5 brings enough improvements that Anandtech found that DDR5-4800 CL40 was consistently faster in multi-threaded workloads (including most games they tested not just synthetics) than DDR4-3200 CL22 on their Alder lake systems while remaining comparable in single threaded ones.

https://www.anandtech.com/show/1704...w-hybrid-performance-brings-hybrid-complexity

 

[DanNeely]

I'll admit that while I generally consider myself quite knowledgeable about all things in this hobby, RAM timings are one of the areas that I have never really dug into.

Didn't we used to have CL's down in the low single digits years ago? 2 or 3? 40 seems very high.

Every DDR level doubles the clock rate of the data bus; dram itself has hit physical limits and modern implementations are only about twice as fast than the PC-100/133 ram it replaced. The fact that it became extremely difficult to increase the clock speed of the DRAM itself was a major driving force behind first DDR and then DDR2+ increasing the memory bus speed even more independently of the actual ram speed.

All the rest of the performance gains you've gotten are from increased parallelism but while reading 8 bits at once gives 8x the throughput of reading 1 at a time the latency for the initial read doesn't get any faster. In terms of elapsed time DDR1 CL2 = DDR2 CL4 = DDR3 CL8 = DDR4 CL 16 = DDR5 CL 32 a equivalent clock speeds: So PC-133, DDR1-266, DDR2-533, DDR3-1066, DDR4-2133, and DDR5-4266 are all running the dram at the same 133mhz; and while mature memory standards eventually manage to push the dram itself a decent ways above 200mhz the initial versions always start at the slower and less technically demanding end of the range.

 

[Nenu]

Somewhat unfortunate timing on this comment. Despite the headline numbers that would make you think it was crap DDR5 brings enough improvements that Anandtech found that DDR5-4800 CL40 was consistently faster in multi-threaded workloads (including most games they tested not just synthetics) than DDR4-3200 CL22 on their Alder lake systems while remaining comparable in single threaded ones.

https://www.anandtech.com/show/1704...w-hybrid-performance-brings-hybrid-complexity

Too early, I'll wait for faster.

 

[DanNeely]

Too early, I'll wait for faster.

I'll wait for more affordable (early adopter tax always sucks); being somewhat faster in a year will be a bonus.

 

[Krenum]

I'll wait for more affordable (early adopter tax always sucks); being somewhat faster in a year will be a bonus.

I think now days its better to pay the "early adopter tax" than it is to not pay the "I can't find shit in stores tax".

 

[Nenu]

I think now days its better to pay the "early adopter tax" than it is to not pay the "I can't find shit in stores tax".

Ram availability isnt like gfx cards.
You mignt be forced to buy a bit better ram at a higher price but not get thoroughly scammed,
I'm waiting a while anyways, no need to buy now.

 

[Furious_Styles]

Ram availability isnt like gfx cards.
You mignt be forced to buy a bit better ram at a higher price but not get thoroughly scammed,
I'm waiting a while anyways, no need to buy now.

Yeah exactly, miners don't need to keep buying up all the memory they can find.

 

[Axman]

dram itself has hit physical limits and is only about twice as fast than the PC-100/133 ram it replaced.

Magnets, how do they work?

We can dope silicon in interesting new ways for decades, like we always have, but unless we switch over to some other substrate like diamond, there's going to be physical limitations.

Memory's fuzzy but IIRC ruby was the next likely contender. Cheaper to synthesize, but who knows, that was a long time ago. Maybe lab diamond tech has progressed further.

 

[DanNeely]

I think now days its better to pay the "early adopter tax" than it is to not pay the "I can't find shit in stores tax".

This forum needs an "agree, but I hate it" button to go next to "like".

 

[DanNeely]

Magnets, how do they work?

We can dope silicon in interesting new ways for decades, like we always have, but unless we switch over to some other substrate like diamond, there's going to be physical limitations.

Memory's fuzzy but IIRC ruby was the next likely contender. Cheaper to synthesize, but who knows, that was a long time ago. Maybe lab diamond tech has progressed further.

Lots of things have been hyped as the replacement for silicon over the decades I've been watching tech, I don't recall seeing ruby/sapphire mentioned but it's certainly possible, but outside of special use cases (eg power transistors, sensors, and photovoltaics) none of them have gone beyond the R&D lab stage. The biggest problem is that most of them are only good for making either P or N type transistors; and without both types for a CMOS style design idle power levels for logic circuits go sky high. Even if that's solved the initial development of industrial tooling and catching up to the density of silicon is going to be incredibly hard.