News on DDR5 ram

[NoddyPirate]

Like for like I believe it is, I can't think why it wouldn't be as the other latencies scale similarly, however there are other changes..... such as the burst length increasing to 16 bits. I believe that means it shoots 16 arrows at once, instead of 8, which may double some sort of output.

With the above in mind - I'd love your thoughts @Scott on the below....

Crucial did a white paper on it that I read a wee while back. It was interesting but fairly high level so didn't go into particularly nice detail.

I'll have to look for that paper - would help me pass the time!

Just three quick one that I have read recently - although there are many more out there too.

______

Crucial have this brief summary online - with a quote in there:

"because the latency in nanoseconds for DDR4-2400 CL17 and DDR4-2666 CL19 is roughly the same, the higher speed DDR4-2666 RAM will provide better performance"

______

Also this - (the bold highlight is the authors rather than mine):

"in a situation where the true latency of two sticks of RAM are tied, despite them having different clock speeds (because the faster-clocked choice has a worse latency). In situations where that kind of tie occurs, the higher speed of RAM takes precedence over which has superior CAS latency. Thus, when comparing a stick of DDR4-3000 RAM with a CAS of 15 and a stick of DDR4-3600 RAM with a CAS of 18 (which would both have a true latency of 10 nanoseconds), the DDR4-3600 should be preferred"

______

Ultimately, my understanding is that Bandwidth has to count for something. And since Bandwidths are measured in MB per second, then it remains an objective measure of raw performance in terms of time. Here's another bit for fun in that vein:

"DDR3-1600 CL10 as an example has a latency of 12.5 nanoseconds to initiate a read. DDR4-2666 CL17 has a latency of 12.75 nanoseconds—basically the same. But the DDR4 provides 21.3GB/s of bandwidth compared to 12.8GB/s for DDR3."

That one also goes on to point out that there was no notable difference between DDR3 and DDR4 modules with similar speeds and timings - so it's only higher speeds that provide the higher bandwidth and improved performance for the same true latency - reagardless of DDR version.

______

Throw in different BCLK's for some Overclocked RAM and the effect of FCLK speed for Ryzen and the whole picture gets even messier I expect!

I guess my simplified point is that performance per £ is likely to be on par, or maybe even less than, with DDR4 for a while.

I think you will be proven 100% correct on that for a quite a while to come. Even just the effort to recoup the R&D for DDR5 alone will see prices stay high initially, alongside some agressive marketing no doubt to make users believe their lives might change with it in the machines too!

 

[Scott]

With the above in mind - I'd love your thoughts @Scott on the below....


I'll have to look for that paper - would help me pass the time!

Just three quick one that I have read recently - although there are many more out there too.

______

Crucial have this brief summary online - with a quote in there:

"because the latency in nanoseconds for DDR4-2400 CL17 and DDR4-2666 CL19 is roughly the same, the higher speed DDR4-2666 RAM will provide better performance"

______

Also this - (the bold highlight is the authors rather than mine):

"in a situation where the true latency of two sticks of RAM are tied, despite them having different clock speeds (because the faster-clocked choice has a worse latency). In situations where that kind of tie occurs, the higher speed of RAM takes precedence over which has superior CAS latency. Thus, when comparing a stick of DDR4-3000 RAM with a CAS of 15 and a stick of DDR4-3600 RAM with a CAS of 18 (which would both have a true latency of 10 nanoseconds), the DDR4-3600 should be preferred"

______

Ultimately, my understanding is that Bandwidth has to count for something. And since Bandwidths are measured in MB per second, then it remains an objective measure of raw performance in terms of time. Here's another bit for fun in that vein:

"DDR3-1600 CL10 as an example has a latency of 12.5 nanoseconds to initiate a read. DDR4-2666 CL17 has a latency of 12.75 nanoseconds—basically the same. But the DDR4 provides 21.3GB/s of bandwidth compared to 12.8GB/s for DDR3."

That one also goes on to point out that there was no notable difference between DDR3 and DDR4 modules with similar speeds and timings - so it's only higher speeds that provide the higher bandwidth and improved performance for the same true latency - reagardless of DDR version.

______

Throw in different BCLK's for some Overclocked RAM and the effect of FCLK speed for Ryzen and the whole picture gets even messier I expect!


I think you will be proven 100% correct on that for a quite a while to come. Even just the effort to recoup the R&D for DDR5 alone will see prices stay high initially, alongside some agressive marketing no doubt to make users believe their lives might change with it in the machines too!

That was the same information that I had read..... but none of it goes on to explain why, which was a little annoying.

The bandwidth clearly makes a difference, as the higher the number the lower the total latency..... thanks to how quickly it can enact the operation.

You have the delay time plus the operation time that gives the total latency in the task. If the frequency was lower, it would take longer to complete..... I imagine. I'm thinking that the 16 vs 8 is going to be similar to the channels in how they function. But we will soon see.

I'm yet to truly understand how it all works though as I've not found any concrete explanations given. Past the understanding of what the total latency is..... we seem to just be told "trust me, it's faster"..... while benchmarks often prove otherwise.

 

[NoddyPirate]

I'm yet to truly understand how it all works though as I've not found any concrete explanations given.

I'm gald to hear you say that because I feel the very same! I read something and think "Ah yes, of course!" but then I read something else which make me go "Wait a minute, I thought that other article said the opposite??"

Past the understanding of what the total latency is..... we seem to just be told "trust me, it's faster"..... while benchmarks often prove otherwise.

Part of me can see how CAS latency is simply the time it take for an intial response to be received from RAM, but that's only for the first piece of data - the next data set wouldn't need another full CAS period to become available - conceivably it could be ready after just one more clock cycle. If this were the case, then it might possibly explain how the higher clock speed gives better performance even at the same effective latency. In other words - each new piece of data doesn't have to wait for the first bit of data to be delivered before the request for it can be made?

This article seems to suggest that possible explanation.

Anyway, I would be grateful if you would share any other info you may find - here or directly with me if you'd rather not further fatten up a thread - more info is always better!

 

[MrWilson]

I'm hoping that first generation AM5 motherboards will have backwards compatibility DDR4 RAM, otherwise however nice the Ryzen 6000 series will turn out to be it is mooted if you have to fork out a small fortune for DDR5 in your new build.
This side of things is an area where my knowledge falls shorter. what do you guys think is likely to happen?

 

[NoddyPirate]

I'm hoping that first generation AM5 motherboards will have backwards compatibility DDR4 RAM, otherwise however nice the Ryzen 6000 series will turn out to be it is mooted if you have to fork out a small fortune for DDR5 in your new build.
This side of things is an area where my knowledge falls shorter. what do you guys think is likely to happen?

The cynic in me says that the backwards compatibility is the enemy of R&D investment. Whenever DDR5 hits the road will all wheels turning then I wouldn't be surprised to see an in-built system to prevent compatability - the same way DDR2 to DDR3 had the same pin count but with a notch placed conveniently to prevent inserting one type into another slot - alongside the electrical incompatibility.

But it seems more likely that AM5 will be out before DDR5 is mainstream - in which case I would expect that DDR4 will be an option for a good while to come.

 

[NoddyPirate]

@Scott - you got me doing more reading for an hour earlier! Some old but really interesting reading from anadtech here and also here.

From as much as I can make sense of it - it seems that when it comes to bursting - if data is required from within the same page, then subsequent requests for data can be queued, effectively reducing the impact of access latencies and allowing the bandwidth to be maximised.

Whereas in the worst case if lots of small amounts of data are in different memory locations, then the bandwidth becomes somewhat neutralised and the latency becomes the limiting factor.

So effectively, for the same true latency, at worst - higher speeds make no difference but also don't hurt anything, while at best they make a big difference, with the reality sitting somewhere in between depending on the type of operation.

A neat quote from their articles (again their empahsis, not mine):

"Lower timings allow faster access to the data, while higher bandwidth allows access to more data."

and then that logically follows into:

"Applications that access large amounts of data - either sequentially or randomly - usually benefit from increased bandwidth. Bandwidth can be increased either by increasing the number of memory channels (i.e. dual-channel) or by increasing the clock speed of the memory [and can] provide a significant boost in performance. Many games and multimedia benchmarks process large amounts of data that cannot reside within the cache of the CPU, and being able to retrieve the data faster can help out."

They also say that for the same reasons above, applications that don't need large memory capacity will be more impacted by latency, and I wonder if this is where so many benchmarks fail to show any benefit from higher speeds at the same true latency.

For me though it would suggest that DDR5 could still prove to generate a signficant performance hike through the dramatically higher bandwidth, but it would likely only be seen in applications that need access to large amounts of data in the first place. For the average gamer or general user, there may be little or no change at all if the true latencies remain basically the same.....

For someone like me, with my photography workflow making use of silly amounts of RAM, I imagine I will stick with Crucials advice:

• Step 1: Identify the highest memory speed supported by both your processor and motherboard (including overclocking profiles).
• Step 2: Select the lowest latency memory that fits within your budget at that speed.

 

[Scott]

Good reading, yeah I would take exactly the same from it and the logic makes sense. It's like a subway queue in Tokyo. Latency is the time between trains....... frequency is the size of the carriage. During peak times, when the trains are being utilised to capacity.... the higher frequency makes a huge difference.... at quieter times, the latency is the limiting factor between different dimms.

Don't forget step 3, which for me is key, be prepared to have your wallet emptied.

 

[NoddyPirate]

Good reading, yeah I would take exactly the same from it and the logic makes sense. It's like a subway queue in Tokyo. Latency is the time between trains....... frequency is the size of the carriage. During peak times, when the trains are being utilised to capacity.... the higher frequency makes a huge difference.... at quieter times, the latency is the limiting factor between different dimms.

Why didn't they just write that paragraph instead, rather than page after page of barely decipherable technical deatil!! I love that analogy @Scott and I think it sums it up perfectly actually.

When you need large amounts of memory, speed has a greater impact. When you need small amounts of memory, latency has a greater impact. So, I guess that as long as the real latency is no worse, then higher speeds may be more beneficial for some users, but not all.

Don't forget step 3, which for me is key, be prepared to have your wallet emptied.

Absolutely true:

 

[Scott]

That's the sort of RAM that the DDR5 will need to be able to trade blows with. It'll be very interesting.

 

[Bigfoot]

@NoddyPirate I know this is only Wikipedia, but this explanation does show why CAS latency is not the only consideration.

CAS latency - Wikipedia

en.wikipedia.org

 

[NoddyPirate]

@NoddyPirate I know this is only Wikipedia, but this explanation does show why CAS latency is not the only consideration.

CAS latency - Wikipedia

en.wikipedia.org

Thanks @Bigfoot - Wiki can be good in fairness.

And yep I think that is kind of the point ultimately - latency is only part of the story. 👍

 

[SpyderTracks]

Some new videos on DDR5

It's confirmed to have ECC built in!!! AT LONG LAST!!!

It's going to be EXPENSIVE initially and speeds won't be much over what we're used to in the main on DDR4. But that's completely normal for early adoption anyway.