DDR5 spec finalised

[porina]

 

Quote

Marking an important milestone in computer memory development, today the JEDEC Solid State Technology Association is releasing the final specification for its next mainstream memory standard, DDR5 SDRAM. The latest iteration of the DDR standard that has been driving PCs, servers, and everything in-between since the late 90s, DDR5 once again extends the capabilities of DDR memory, doubling the peak memory speeds while greatly increasing memory sizes as well. Hardware based on the new standard is expected in 2021, with adoption starting at the server level before trickling down to client PCs and other devices later on.  

Source: https://www.anandtech.com/show/15912/ddr5-specification-released-setting-the-stage-for-ddr56400-and-beyond

 

Summary

The DDR5 specification has now been finalised, and allows actual products to be created. Up to now, the companies involved with the standard would be working with draft versions. Officially supported speeds increases from DDR4's maximum of 3200 to DDR5's starting speed of 4800, going up to an expected maximum of 6400. DDR4 modules above 3200 are available now, but these are not a standard defined speed. Similarly we would expect high speed kits of DDR5 to appear above 6400 at some point. 

 

Some changes of note going from DDR4 to DDR5:

• Nominal voltage is lowered from 1.2V to 1.1V, so potentially reducing power slightly.
• Voltage regulation will be included on module, and no longer the responsibility of the motherboard. This will apparently help with high capacity motherboards, in that the regulation circuitry it now included as needed with the module, and not loaded up front on the motherboard at design. Implicitly, mobos might be cheaper, but modules more expensive.
• Modules will have dual 32-bit channels, compared to the current single 64-bit channel. While the extra channel wont directly help with maximum bandwidth, combined with other changes internally to the ram, it may allow for better response when accessing the module.
• There is die level ECC. This isn't the same as module-wide ECC as we know it. The concern is that as ram capacities get bigger, the chances of having bad ram increase likewise. Some level of ECC will help improve yield, in a similar way employed by flash SSDs.

 

My thoughts

A combination of improving CPU clocks, more cores, and more efficient cores contributes to increasing levels of CPU core performance, with the effect the ratio of ram bandwidth has been in decline for a while. DDR5 will address that and help prevent it being as much of a bottleneck. The timescales still seem to be some way out. Maybe server implementations around this time next year. Consumer applications maybe towards end of 2021, or into 2022. Don't hold your breath.

 

It will also be interesting what this might mean to the ram overclocking crowd. With the voltage regulation on module, do we still have significant control over that? How much impact on latency performance might the two half as wide channels per module provide? How will we refer to this in future? If we were to say a future memory controller could talk to two modules simultaneously at full bandwidth, as most consumer level offerings today, would we call that quad channel? It's technically correct, although with half the width in each channel it doesn't provide a maximum bandwidth improvement over a DDR4 dual channel of otherwise same speed.

[zeusthemoose]

Question for people with more knowledge: Are future cpus going to have to specifically support this or is that only a requirement for motherboards? Any predictions on when we will get this in consumer level hardware (hopefully based off of what happened with ddr4)?

[Sakkura]

10 minutes ago, zeusthemoose said:

Question for people with more knowledge: Are future cpus going to have to specifically support this or is that only a requirement for motherboards?

Yes, the memory controller is in the CPU nowadays.

 

But you can have support for two types of memory. Intel did that with DDR3 and DDR4 for a short period.

[TVwazhere]

Me with my "fast" DDR3 1866mhz racing against DDR5 6400mhz+

[Doobeedoo]

Very good improvements. The much higher frequencies will be very interesting to see how aids the CPU especially lile initial entry kit vs faster one later on. Now to wait and see also latencies too. The tighter the better. 

[porina]

31 minutes ago, zeusthemoose said:

Question for people with more knowledge: Are future cpus going to have to specifically support this or is that only a requirement for motherboards? Any predictions on when we will get this in consumer level hardware (hopefully based off of what happened with ddr4)?

When: the linked article suggest a year or more, and even then likely in servers before consumer kit. So that would put it roughly end of next year, to early 2022. 

 

On CPU support, the memory controller will have to specifically support it. Also, the mobo would have to be laid out for it. You're not going to get a single slot that takes DDR4 and DDR5. This is more interesting in AMD's case since their memory controller is physically separate from the cores. People have speculated they could offer CPUs at the time in both DDR4 and DDR5 variations by changing the IOD, to ease the transition for those with older motherboards. This will likely be new socket time for them.

[Bananasplit_00]

57 minutes ago, TVwazhere said:

Me with my "fast" DDR3 1866mhz racing against DDR5 6400mhz+

Im sitting at a BLAZING fast DDR3 2133

[cj09beira]

1 hour ago, zeusthemoose said:

Question for people with more knowledge: Are future cpus going to have to specifically support this or is that only a requirement for motherboards? Any predictions on when we will get this in consumer level hardware (hopefully based off of what happened with ddr4)?

it will take a while, though its possible amd might make a ddr5 version of zen 3 for servers as they only need to change the io die, if things happen as normal ddr5 will be expensive and in general not really worth it for a while then prices will slowly come down.

1 hour ago, Doobeedoo said:

Very good improvements. The much higher frequencies will be very interesting to see how aids the CPU especially lile initial entry kit vs faster one later on. Now to wait and see also latencies too. The tighter the better. 

i bet those latencies will be pretty horrible.

28 minutes ago, Bananasplit_00 said:

Im sitting at a BLAZING fast DDR3 2133

i raize you 2400mhz cl13 (from a 1600mhz cl10 kit), those were the days of good overclocks.

[Vector0102]

Will we see a further increase in CAS latencies and will this be a problem?

[porina]

5 minutes ago, cj09beira said:

i bet those latencies will be pretty horrible.

Historic trend is that the standard timings go up in clock cycles, but that's offset by the increase in speeds, so the actual latency time is about constant. As always, enthusiast kits will probably come in with more aggressive settings.

[Doobeedoo]

8 minutes ago, cj09beira said:

i bet those latencies will be pretty horrible.

Much higher frequencies to offset etc.

1 minute ago, porina said:

Historic trend is that the standard timings go up in clock cycles, but that's offset by the increase in speeds, so the actual latency time is about constant. As always, enthusiast kits will probably come in with more aggressive settings.

Yeah this. So should be quite interesting to see.

[Bananasplit_00]

9 minutes ago, cj09beira said:

i raize you 2400mhz cl13 (from a 1600mhz cl10 kit), those were the days of good overclocks.

I'm 2133 cl10

[cj09beira]

 

3 minutes ago, porina said:

Historic trend is that the standard timings go up in clock cycles, but that's offset by the increase in speeds, so the actual latency time is about constant. As always, enthusiast kits will probably come in with more aggressive settings.

maybe it will be better than ddr3 to ddr4, but in that case ddr4 2400mhz had much looser timings that what i had on my ddr3 1600mhz cl10 kit that did 2400mhz cl13

[porina]

15 minutes ago, cj09beira said:

maybe it will be better than ddr3 to ddr4, but in that case ddr4 2400mhz had much looser timings that what i had on my ddr3 1600mhz cl10 kit that did 2400mhz cl13

We have the tables below to go on:

https://en.wikipedia.org/wiki/DDR3_SDRAM

https://en.wikipedia.org/wiki/DDR4_SDRAM

 

Depending on the speed grade, DDR3 did have the capability to drop to 10ns, whereas DDR4 seems to floor at 12.5ns. The DDR3 ones are listed as "optional". I don't know of any such data in the public for DDR5 yet (it's $369 if you want to buy a copy of the standard), but assuming it follows the trend for DDR4, that would put DDR5-4800C30 and DDR5-6400C40. Again, this is the standard based speed. Enthusiast modules will likely do much better than that. Standard based DDR4-3200 is only specified down to C20, but enthusiast kits typically start at C16.

[Sakkura]

2 hours ago, cj09beira said:

 

maybe it will be better than ddr3 to ddr4, but in that case ddr4 2400mhz had much looser timings that what i had on my ddr3 1600mhz cl10 kit that did 2400mhz cl13

Meh. Back in the day DDR3-1600 CL9 was the go-to at the cheaper end of the market, today that's DDR4-3000 CL16. Latencies are 11.25ns and 10.67ns respectively. DDR3 and DDR4 are so similar on latency it doesn't make any difference.

[dizmo]

I can't wait until RAM surpasses the 10,000mhz range.

[hollyh88]

5 hours ago, porina said:

 

Source: https://www.anandtech.com/show/15912/ddr5-specification-released-setting-the-stage-for-ddr56400-and-beyond

 

Summary

The DDR5 specification has now been finalised, and allows actual products to be created. Up to now, the companies involved with the standard would be working with draft versions. Officially supported speeds increases from DDR4's maximum of 3200 to DDR5's starting speed of 4800, going up to an expected maximum of 6400. DDR4 modules above 3200 are available now, but these are not a standard defined speed. Similarly we would expect high speed kits of DDR5 to appear above 6400 at some point. 

 

Some changes of note going from DDR4 to DDR5:

• Nominal voltage is lowered from 1.2V to 1.1V, so potentially reducing power slightly.
• Voltage regulation will be included on module, and no longer the responsibility of the motherboard. This will apparently help with high capacity motherboards, in that the regulation circuitry it now included as needed with the module, and not loaded up front on the motherboard at design. Implicitly, mobos might be cheaper, but modules more expensive.
• Modules will have dual 32-bit channels, compared to the current single 64-bit channel. While the extra channel wont directly help with maximum bandwidth, combined with other changes internally to the ram, it may allow for better response when accessing the module.
• There is die level ECC. This isn't the same as module-wide ECC as we know it. The concern is that as ram capacities get bigger, the chances of having bad ram increase likewise. Some level of ECC will help improve yield, in a similar way employed by flash SSDs.

 

My thoughts

A combination of improving CPU clocks, more cores, and more efficient cores contributes to increasing levels of CPU core performance, with the effect the ratio of ram bandwidth has been in decline for a while. DDR5 will address that and help prevent it being as much of a bottleneck. The timescales still seem to be some way out. Maybe server implementations around this time next year. Consumer applications maybe towards end of 2021, or into 2022. Don't hold your breath.

 

It will also be interesting what this might mean to the ram overclocking crowd. With the voltage regulation on module, do we still have significant control over that? How much impact on latency performance might the two half as wide channels per module provide? How will we refer to this in future? If we were to say a future memory controller could talk to two modules simultaneously at full bandwidth, as most consumer level offerings today, would we call that quad channel? It's technically correct, although with half the width in each channel it doesn't provide a maximum bandwidth improvement over a DDR4 dual channel of otherwise same speed.

IIIIIIIII cant wait to see ryzen all juiced up on these bad boys  

[cj09beira]

1 hour ago, Sakkura said:

Meh. Back in the day DDR3-1600 CL9 was the go-to at the cheaper end of the market, today that's DDR4-3000 CL16. Latencies are 11.25ns and 10.67ns respectively. DDR3 and DDR4 are so similar on latency it doesn't make any difference.

i am not saying it wont end up good, i am saying it might very well start pretty bad, at first ddr4 was usually around 2400mhz with bad timmings 

[poochyena]

Can't wait. Was thinking about upgrading my PC next year, but I think i'll wait an extra year for ddr5

[Aereldor]

7 hours ago, TVwazhere said:

Me with my "fast" DDR3 1866mhz racing against DDR5 6400mhz+

CAS latency of 6 has to count for something

[PianoPlayer88Key]

On-die ECC looks interesting, I'd like to see ECC RAM across the board.

 

Also having the voltage regulation on the module itself -- could that possibly mean that we'll be able to have RDIMMs and LRDIMMs on consumer boards?

The 128GB per module (512GB per 4-slot board) for UDIMMs is *NOT* going to be enough for me in the long run.  It might be okay now, but I don't want to have to replace my board this time around as fast as I had replaced it before.

My current board was bought in January 2015 (ASRock Z97 Extreme6, has 32GB RAM).  My first board was bought in February 2008 (Gigabyte GA-MA69G- S3H, had 4GB RAM limited to 3GB because of 32-bit Windows XP).

The laptop I'm on now currently has 64GB (4x16) DDR4.

Spoiler

2020-05-08:     2015-11-14:      2010-03-10: 

 

I'm planning to upgrade sometime in 2022, thinking maybe Black Friday time frame unless I find a really good deal before then.  Most likely looking at socket AM5... although I really wish AMD could come up with a way to unify AM5 and SP5.  For example have SP5 basically be 2x AM5 sockets side by side (kind of like in the picture in the spoiler below), and have mid to higher tier boards support both. 

 

Spoiler

 

I want to be able to buy a nice board and start with a cheap CPU like an Athlon (which I hope is enough faster that it absolutely destroys anything from Haswell which is my current desktop's generation), then later upgrade to a Ryzen 7/9 or Threadripper or Epyc without having to get a new board.

Also I'd like plenty of bandwidth for expansion slots, SATA (or SAS / NVME / whatever storage) ports, etc - no more of this stupid "m.2 slot shared with SATA and PCIe, if M.2 is used PCIe runs at half bandwidth or sata port disabled" crap.

 

Also from the time I build my next PC, on, I want to be able to do future upgrades ONE PART AT A TIME!  No having to rip out the motherboard and disconnect a bunch of other stuff just because I want to upgrade my CPU or RAM beyond 1 or 2 generations after the board launches (CPU) or whatever existed when the board launched (RAM). 

I probably won't be upgrading the CPU more often than every 3 or 4 years or so, except maybe if I initially get an Athlon or Ryzen 3 then a year later upgrade to like Threadripper.  If I would have to get a new board though, I'd probably be wanting to keep it for at least 12 or more years.  I'm thinking about not replacing this next board until DDR8 is out, possibly, or after I've had 2 or 3 Seasonic Prime PSUs die of old age (with each next PSU being bought after the previous one bit the dust, sorry Queen), whichever is later.

 

I could say a lot more about my gripes for upgradeability, my fears that my next CPU won't be the big upgrade I would like (one of the BIG things for me - my current CPU (i7-4790K) takes 4 days to transcode a 4-minute video to 4K H.265 in handbrake at max quality settings, and 2 minutes to transcode 2 hours of audio to 320kbps mp3 max settings in Lame, and I'd like my new same-price CPU to be as fast at video encoding (with existing software, idk if handbrake does AVX or AVX2 or if my 4790K supports it, i'm sure it doesn't support AVX512) as my current CPU is at audio encoding), but I've rambled long enough lol.

[Master Disaster]

20 minutes ago, PianoPlayer88Key said:

On-die ECC looks interesting, I'd like to see ECC RAM across the board.

Hell no, ECC memory is entirely unnecessary at home. Adding complexity also adds cost and RAM is already pricey enough thank you very much.

 

Having it as a supported option for people running workstations (as the AM4 platform does) is one thing but forcing everyone to pay more for something they simply do not need is another. Same goes for registered memory as well, add support for it to the IMC but don't make it compulsory.

[porina]

5 minutes ago, Master Disaster said:

Hell no, ECC memory is entirely unnecessary at home. Adding complexity also adds cost and RAM is already pricey enough thank you very much.

To my understanding, you're getting it in DDR5 whether you want it or not. But there are levels of implementations. The built in level is included at the die level, and is intended to catch smaller errors only. The comparison is to flash in SSDs, that without some degree of error correction, it would be far too expensive to get good data out. As ram capacities increase, the chances of a bad bit somewhere go up with it, and a basic level of correction will be implemented. It may even help prevent cost increases, in that less than perfect ram will be serviceable. This does not replace module level ECC, which can detect and correct for more error conditions.

[Master Disaster]

1 minute ago, porina said:

To my understanding, you're getting it in DDR5 whether you want it or not. But there are levels of implementations. The built in level is included at the die level, and is intended to catch smaller errors only. The comparison is to flash in SSDs, that without some degree of error correction, it would be far too expensive to get good data out. As ram capacities increase, the chances of a bad bit somewhere go up with it, and a basic level of correction will be implemented. It may even help prevent cost increases, in that less than perfect ram will be serviceable. This does not replace module level ECC, which can detect and correct for more error conditions.

TBF that does make sense. As capacity increases so does the chance of an error occurring.

[porina]

1 minute ago, Master Disaster said:

TBF that does make sense. As capacity increases so does the chance of an error occurring.

I did think after I posted that, some level of ECC may also help them attain the higher speeds. I'm not sure how comfortable I am in that sense, we expect and allow flaws to be present and correct for them, rather than the traditional approach of not having any errors in the first place. I'm sure there are far smarter people than I who have worked out what the tolerance is so that we do not suffer a degradation in data reliability.