Wednesday, April 27th 2022

AMD Ryzen 7000 "Raphael" to Ship with DDR5-5200 Native Support

AMD's upcoming Socket AM5 Ryzen 7000-series "Raphael" desktop processors will ship with native support for DDR5-5200 memory speed, according to a marketing slide by memory maker Apacer (which also owns the overclocking memory brand ZADAK). The "Zen 4" based desktop processors will feature a dual-channel DDR5 (4 sub-channel) interface, just like the 12th Gen Core "Alder Lake," but with no backwards compatibility with DDR4.

AMD already stated that Ryzen 7000 processors have a design focus on memory overclocking capabilities, including AMD EXPO, a custom memory module SPD extension standard rivaling Intel XMP 3.0, which will come with fine-grained settings specific to the AMD memory controller architecture. Until now, AMD relied on A-XMP, a motherboard vendor-enabled feature based in the UEFI firmware setup program, which translates Intel XMP SPD profiles of memory modules into AMD-approximate settings.
The Apacer slide also confirms that EPYC "Genoa" processors will come with DDR5-5200 native support. The Socket SP5 processors feature a 12-channel DDR5 (24 sub-channel) memory interface. The already-launched Ryzen 6000 "Rembrandt" mobile processors come with dual-channel (4 sub-channel) DDR5-4800 and LPDDR5-6400 interfaces.

AMD Ryzen 7000 "Raphael" processors natively supporting DDR5-5200 means that JEDEC PC5-42600 standard memory modules will run at that speed without any settings being needed in the UEFI setup program. For higher frequencies, such as DDR5-6000, AMD will rely on EXPO-certified memory modules, or possibly A-XMP like BIOS-level translators for Intel XMP 3.0 profiles.
Sources: momomo_us (Twitter), Apacer Industrial, VideoCardz
Add your own comment

50 Comments on AMD Ryzen 7000 "Raphael" to Ship with DDR5-5200 Native Support

#1
Mussels
Freshwater Moderator
So... intel people?


Is this a good start for DDR5?
I know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher

Hows this compare to the first gen intel DDR5 (stock/OC?)
Posted on Reply
#2
Crackong
They listed Alder Lake-S as the Intel Workstation platform.
No Alder Lake-X ?
Posted on Reply
#3
TheLostSwede
MusselsSo... intel people?


Is this a good start for DDR5?
I know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher

Hows this compare to the first gen intel DDR5 (stock/OC?)
Might be irrelevant, as we don't know what AMD is planning for their consumer CPUs.
Posted on Reply
#4
AnarchoPrimitiv
MusselsI know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher
Check out the memory reviews on this website, there are plenty of examples where AMD matches and even exceeds Intel's memory frequencies, I think one of the latest memory reviews demonstrated this.
Posted on Reply
#5
Chrispy_
MusselsSo... intel people?


Is this a good start for DDR5?
I know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher

Hows this compare to the first gen intel DDR5 (stock/OC?)
You can't compare clockspeeds anyway because AMD and Intel use very different memory timings which dramatically affect access latency.

The clockspeed gives you total theoretical bandwidth but neither AMD nor Intel platforms ever managed to reach those theoretical numbers with DDR4, not even with purely synthetic bandwidth tests.
Posted on Reply
#6
Tsukiyomi91
We'll wait and see if the difference is worth the wait. Also, I doubt AMD is gonna be selling them at competitive price knowing how many consumers they've pissed off with a revised R7 5800X3D and a slew of non-X CPUs that was meant to be released around the same time when the 5000 Series CPUs lineup was new.
Posted on Reply
#7
mama
Love the focus on memory overclocking. That's where the real gains will be made in the new processors.
Posted on Reply
#8
DeathtoGnomes
What was the first DD5? 4400mhz?

5200 is between that and the speculated cap of 6000mhz
Posted on Reply
#9
lexluthermiester
MusselsIs this a good start for DDR5?
IMHO, this is a very good start as long as it's stable.
DeathtoGnomesWhat was the first DD5? 4400mhz?
4800mhz
DeathtoGnomes5200 is between that and the speculated cap of 6000mhz
The official spec is 4800mhz to 7200mhz.
en.wikipedia.org/wiki/DDR5_SDRAM

Where AMD is likely going to have a good advantage with 5200mhz base is the timings. Seems logical that they would tighten them up a bit for such a speed.
Posted on Reply
#10
Jism
Now look at that Epyc platform. 12 channels up to 5200Mhz. Owned.
Posted on Reply
#11
Chrispy_
mamaLove the focus on memory overclocking. That's where the real gains will be made in the new processors.
Actually the performance gains for Zen architecture has been from FCLK (Infinity fabric) overclocking and have almost nothing to do with how fast the memory clockspeed is.

Getting more bandwidth has provided almost zero gains for Ryzen and actually dropping down to slower RAM that allows tighter timings and a 1:1 MCLK:FCLK divider is genuinely faster in every single scenario, no exceptions. DDR4-3800 is consistently the highest-performing memory speed for Zen2 and Zen3 simply because that is the highest memory speed most people use with a 1:1 MCLK:FCLK divider.

I'm not saying more bandwidth won't help, but it will only help if FLCK and memory latency do not suffer as a consequence of increasing bandwidth.

There's this graph from AMD showing a big latency penalty moving beyond DDR4-3800 because of the divider:

Confirmed by independent testing:


And then application and gaming benchmarks all over the web that prove performance takes a huge hit the minute you are forced to drop your FCLK because the RAM is too fast to run with a 1:1 FLCK-MCLK ratio.

In theory, with fast enough RAM (say 4400 or 4600) you can break even with DDR4-3800 but in reality, kits with those XMP ratings can never operate with tight enough timings to actually catch up and even ludicriously expensive DDR4-4600 rarely matches the performance of a cheapo 3600 CL17 kit. In applications that are bandwidth hungry, or in synthetic bandwidth tests, yes even DDR4-4000 helps, but those are not real-world scenarios for Zen2 or Zen3 consumer parts. They benefit far more from Xeon or EPYC platforms using quad-channel or eight-channel memory, and the applications themselves are niche/enterprise and licenses for them cost more than most people earn in a quarter.
Posted on Reply
#12
R0H1T
AnarchoPrimitivI think one of the latest memory reviews demonstrated this.
Only on their APUs & the monolithic die approach, for regular desktop chips they're still bound by IF & chiplets so generally clock lower.
Posted on Reply
#13
zlobby
Give me the PRO version already!!!
Posted on Reply
#14
aQi
MusselsSo... intel people?


Is this a good start for DDR5?
I know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher

Hows this compare to the first gen intel DDR5 (stock/OC?)
Hmm AMD you seem to be quite sure on what you started. Dont forget me when you reach 7000mhz cause by the time you cross it I would have already "zoomed" every bit of it :rockout:

Posted on Reply
#15
Denver
This is like announcing that the official support is 2666mhz in DDR4... To level with the current support, it should start at 6000mhz.
Posted on Reply
#16
Patriot
DenverThis is like announcing that the official support is 2666mhz in DDR4... To level with the current support, it should start at 6000mhz.
jedec has voltage requirements, over 1.1v is an overclock and won't be listed as stock.
Hence intel lists ddr5 4800 as stock. Trident Z5 6000 is 1.3v at its listed settings.
There however are some dimms hitting 5200 at 1.1v. So that IS the current support.
Posted on Reply
#17
docnorth
MusselsSo... intel people?


Is this a good start for DDR5?
I know Zen1 was a bit iffy even reaching 3200, but Zen2/3 settled on DDR3 3800 in the end - while Intel users can zoom a bit higher

Hows this compare to the first gen intel DDR5 (stock/OC?)
Nah, it’s just getting better, AL starting from 4800, Raphael from 5200, Raptor Lake reportedly from 5600 MB/s and so on... By the end of the year I’ll definitely choose DDR5.
Posted on Reply
#18
zlobby
Patriotjedec has voltage requirements, over 1.1v is an overclock and won't be listed as stock.
Hence intel lists ddr5 4800 as stock. Trident Z5 6000 is 1.3v at its listed settings.
There however are some dimms hitting 5200 at 1.1v. So that IS the current support.
1.1V is now OC!?!? Hot damn, I haven't delved into these specs recently!
Posted on Reply
#19
mama
Chrispy_Actually the performance gains for Zen architecture has been from FCLK (Infinity fabric) overclocking and have almost nothing to do with how fast the memory clockspeed is.

Getting more bandwidth has provided almost zero gains for Ryzen and actually dropping down to slower RAM that allows tighter timings and a 1:1 MCLK:FCLK divider is genuinely faster in every single scenario, no exceptions. DDR4-3800 is consistently the highest-performing memory speed for Zen2 and Zen3 simply because that is the highest memory speed most people use with a 1:1 MCLK:FCLK divider.

I'm not saying more bandwidth won't help, but it will only help if FLCK and memory latency do not suffer as a consequence of increasing bandwidth.

There's this graph from AMD showing a big latency penalty moving beyond DDR4-3800 because of the divider:

Confirmed by independent testing:


And then application and gaming benchmarks all over the web that prove performance takes a huge hit the minute you are forced to drop your FCLK because the RAM is too fast to run with a 1:1 FLCK-MCLK ratio.

In theory, with fast enough RAM (say 4400 or 4600) you can break even with DDR4-3800 but in reality, kits with those XMP ratings can never operate with tight enough timings to actually catch up and even ludicriously expensive DDR4-4600 rarely matches the performance of a cheapo 3600 CL17 kit. In applications that are bandwidth hungry, or in synthetic bandwidth tests, yes even DDR4-4000 helps, but those are not real-world scenarios for Zen2 or Zen3 consumer parts. They benefit far more from Xeon or EPYC platforms using quad-channel or eight-channel memory, and the applications themselves are niche/enterprise and licenses for them cost more than most people earn in a quarter.
AMD is moving away from Intel's XMP. I appreciate your analysis and agree with respect to the AM4 platform but I take the introduction of EXPO as a signal that AMD is changing up and there is clearly new scope in DDR5 when it is unshackled from XMP. The old analysis won't hold.
Posted on Reply
#20
ghazi
DeathtoGnomesWhat was the first DD5? 4400mhz?

5200 is between that and the speculated cap of 6000mhz
Who speculated 6000MHz cap? DDR5 will reach 8400MHz at a minimum. Maybe the cap for current ICs isn't far above 6000, more mature designs will fare much better.
Chrispy_They benefit far more from Xeon or EPYC platforms using quad-channel or eight-channel memory, and the applications themselves are niche/enterprise and licenses for them cost more than most people earn in a quarter.
Your post is absolutely correct. The problem for those Xeon/EPYC users is that they're all using registered memory, so the best they can manage right now is something like 3200C22. And not all of their programs would benefit either.
mamaAMD is moving away from Intel's XMP. I appreciate your analysis and agree with respect to the AM4 platform but I take the introduction of EXPO as a signal that AMD is changing up and there is clearly new scope in DDR5 when it is unshackled from XMP. The old analysis won't hold.
XMP really has nothing to do with it. I don't know what you're talking about wrt. "shackled" it's just a way to load a memory profile. My speculation is that the die-to-die Infinity Fabric is being overhauled with the new AM5 package and will support tremendously high frequencies compared to what we're used to, which will have a major impact on latency, and fast DDR5 with good OC capabilities will be necessary to unlock that potential.
Posted on Reply
#21
Aretak
mamaAMD is moving away from Intel's XMP. I appreciate your analysis and agree with respect to the AM4 platform but I take the introduction of EXPO as a signal that AMD is changing up and there is clearly new scope in DDR5 when it is unshackled from XMP. The old analysis won't hold.
"EXPO" is simply another name for memory profiles. It's a marketing term, just like "DOCP" that Asus have already been using on their AMD boards throughout AM4's lifecycle. I'm not sure why you think that renaming a concept is going to have some dramatic effect on performance. AMD just want their own branding out there.
Posted on Reply
#22
mama
Aretak"EXPO" is simply another name for memory profiles. It's a marketing term, just like "DOCP" that Asus have already been using on their AMD boards throughout AM4's lifecycle. I'm not sure why you think that renaming a concept is going to have some dramatic effect on performance. AMD just want their own branding out there.
I don't agree. It is a memory profile and more and specifically designed for the AMD memory controller architecture. I quote the report above: "AMD EXPO, a custom memory module SPD extension standard rivaling Intel XMP 3.0, which will come with fine-grained settings specific to the AMD memory controller architecture. Until now, AMD relied on A-XMP, a motherboard vendor-enabled feature based in the UEFI firmware setup program, which translates Intel XMP SPD profiles of memory modules into AMD-approximate settings."
Posted on Reply
#23
sillyconjunkie
Aretak"EXPO" is simply another name for memory profiles. It's a marketing term, just like "DOCP" that Asus have already been using on their AMD boards throughout AM4's lifecycle. I'm not sure why you think that renaming a concept is going to have some dramatic effect on performance. AMD just want their own branding out there.
mamaI don't agree. It is a memory profile and more and specifically designed for the AMD memory controller architecture. I quote the report above: "AMD EXPO, a custom memory module SPD extension standard rivaling Intel XMP 3.0, which will come with fine-grained settings specific to the AMD memory controller architecture. Until now, AMD relied on A-XMP, a motherboard vendor-enabled feature based in the UEFI firmware setup program, which translates Intel XMP SPD profiles of memory modules into AMD-approximate settings."
It's basically both. Infinity fabric will run 1:1 for the range of certified memory speeds and certified vendor offerings.

Higher mem speeds may or may not require a multiplier depending on a lot of stuff and AMD won't support non-standard configurations "officially". Same stuff..

It's more marketing than anything else.
Posted on Reply
#24
ir_cow
If this news is indeed true. I think we will see 2:1 Infinity fabric from the get-go unless AMD has massively overhauled it. I'm thinking the upper limit for Zen4 will also be DDR5-6800 before having to switch to 4:1 FLCK. Just speculation. Tell will tell soon enough.

Maybe the return of DRAM/Fan kits as well. These PMICs get toasty at 1.5v and above.
Posted on Reply
#25
Chrispy_
ghaziYour post is absolutely correct. The problem for those Xeon/EPYC users is that they're all using registered memory, so the best they can manage right now is something like 3200C22. And not all of their programs would benefit either.
Yeah, very few real-world applications are limited by memory bandwidth. Massive SQL databases *can* push bandwidth, but more commonly storage IO is the bottleneck, then. Custom applications or big-data are all potentially viable candidates but the only time I ever really run into bandwidth limitations is when the hardware is a host for VDI and multiple users are all working on large image/media applications like Premiere/After Effects

The slower ECC is definitely worse from a performance standpoint, but when it's bandwidth that's the problem slower 2133MHz ECC isn't a problem because the server has (typical Xeon Silver/Gold) two 6-channel memory controllers joined by at least 3 10GT/s QPI interconnects. It's not quite as good as having 12 memory channels but realistically there is 6x more bandwidth than a typical dual-channel consumer solution, and so running 2133 instead of 4000MT/s RAM isn't the end of the world, it's still close to 2-3x more bandwidth than the fastest dual-channel consumer platform that money can buy despite the pedestrian ECC 2133 clockspeeds.

People often cite photoshop as a bandwidth-heavy application, and they're not wrong; Photoshop filters and transforms will use all the bandwidth available. It's just that the operation takes fractions of a second, so having lower bandwidth means that the operation you perform half a dozen times an hour takes 0.5 seconds to run, instead of 0.3 seconds to run. Yes, the bandwidth makes it measurably quicker, but not in a way that impacts anyone in the real world.
Posted on Reply
Add your own comment
May 28th, 2022 02:49 EDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts