AMD AM5 Socket to Launch with DDR5-Only Memory Option, Feature Dual-Chipset Designs

[AleksandarK]

AMD is preparing to launch its highly-anticipated AM5 socket for the next generation of motherboards. And today, thanks to the sources over at Tom's Hardware, we have information regarding memory support for B650 and X670 motherboards. According to the report, both B650 and X670 chipsets will limit the user's memory option to the latest DDR5 memory standard, making it impossible for users with already existing DDR4 memory to perform a seamless upgrade to a new platform. So far, we don't have a lot of details about Zen4's integrated memory controller, and we can't be certain if it supports DDR5 only or carries legacy DDR4 support. However, it seems like B650 and X670 motherboards will have no plans to enable the DDR4 standard memory usage.

Additionally, the report confirms that the B650 chipset is connected to the AM5 socket via PCIe 4.0 x4 connection and has eight lanes of PCIe 4.0 (four of which are for M.2 SSD), four SATA, and lots of USB ports. Documents suggest that the chipset-socket connection is available using PCIe 5.0 for some AM5 processors, so we have to wait and see how it works. As far as high-end X670 is concerned, this chipset is a combination of two chipset dies, presumably a combination of two B650 modules. This doesn't work as the older north/southbridge type of a solution but rather doubled connectivity of a single B650 chipset. We have to wait for the official launch to confirm this information.


View at TechPowerUp Main Site | Source

 

[Jism]

If it's designed with only a DDR5 capable IMC then it wont be backwards compatible. AM4 is still relevant for quite some years knowing how strong a 5950X can be.

 

[DeathtoGnomes]

They should just have PCIe5 as default.

 

[john_]

This doesn't work as the older north/southbridge type of a solution but rather doubled connectivity of a single B650 chipset.

Still, I wonder if manufacturers would have the option to put two B650s on a motherboard instead of one X670. This is probably pointless stupid and anti economic, but could it maybe help in some cases with motherboard design or chipset cooling? Just a stupid thought.

 

[MaddoggMiranda]

If it's designed with only a DDR5 capable IMC then it wont be backwards compatible. AM4 is still relevant for quite some years knowing how strong a 5950X can be.

I agree reminds me of SkylakeX. When it launched way back in 2017 they are still relevant today with performance they provide. Hopefully AM4 5000 Series Processors will be just as good at providing the performance over the years.

 

[Deeveo]

Still, I wonder if manufacturers would have the option to put two B650s on a motherboard instead of one X670. This is probably pointless stupid and anti economic, but could it maybe help in some cases with motherboard design or chipset cooling? Just a stupid thought.

Takes too much space on motherboard, and would propably cost more than a x670 board. In X670 those 2 chips would be inside one package, with 2 B650 packages you would need signaling etc to be designed, have the board space for the two packages etc problems.

 

[Andyr]

If it's designed with only a DDR5 capable IMC then it wont be backwards compatible. AM4 is still relevant for quite some years knowing how strong a 5950X can be.

And the 5900x, for that matter, if you're talking content creation workflows.

 

[Ruru]

They should just have PCIe5 as default.

I doubt that it's neccessary yet as PCIe 4.0 is getting more common now.

 

[john_]

AM4 5000 will remain relevant thanks to 5800X3D and 5950X, but it could get even more interesting if AMD was introducing a 5950X3D in 6-8 months. We know that a 5950X3D is not coming, but it could offer people the perfect AM4 option. PCIe 5 and DDR5 isn't needed yet. Nvidia's 4000 series will be PCIe 4.0 anyway and who really needs 10GB+/sec speeds for their primary SSD?

 

[ModEl4]

For the 5800X replacement and up I guess it will acceptable more or less due to price segment, but it will face stiff competition from 13600K/13600KS (6P+8E) and below (probably it will be just a little bit slower in 1080p gaming but in multitasking apps like Cinebench it will be no contest with at least +40% more performance, probably +50% vs 5600X replacement)
Many people don't realize it but the real problem will be for the APU 6000G series. I will go as far to say that for people that just want an entry gaming solution and the main reason is entry gaming and they don't buy an APU for the form factor capabilities (NUC like solutions etc) 6000G is DOA.
We had 4700G Renoir with 156mm² die size at $309 OEM so let's just say +$10 for retail - with Intel the difference between tray (1 year warranty) and box with fan and 3 years warranty is less than $10 for a local official distributor in Europe.So let's just say $319 theoretical SRP.
With 5700G Cezanne (180mm²) we went at $359 SRP although the iGPU part was the same (the reason 5700G is faster in some games than 4700G despite the frequency deficit 2GHz vs 2.1GHz is Zen 3 and cache because the testing is done in 720/1080p resolutions with low/medium settings)
So in the same 7nm process AMD increased the SRP more or less in accordance with die size increasment.
With 6700G Rembrandt we have 208mm² in a more expensive 6nm process, so someone would think that AMD would have a $399 SRP or something, right?
Well no, I think they will launch it at the same SRP as 5700G just to save face and still will not be competitive with Intel Q4 offerings.
A i5 13400F will probably be a 6P+4E design and it will offer better performance in TPU 720p/1080p gaming CPU testing AND in multitasking apps like Cinebench vs 6700G, this is certain imo.
Regarding the Rembrandt APU I keep hearing people saying it will be faster than RX 6400 but it will not, even at 2.5GHz (vs 2.32GHz) it will not be faster due to the fact that RX 6400 has double ROPs (which is crucial), infinity cache + 67% more bandwidth (that it doesn't share with Zen3+ cores) and a beefier unslice (Intel sic) for ACE, geometry processor etc.
If you add CPU/Graphics/motherboard/memory costs you would easily see that AMD is not really competitive at all:
13400F vs 6700G will probably be $360-$180, so already we have a $180 difference
I think probably Intel will have also cheaper motherboard options since already we had price drops and we will see even more price drops since by Q4 we will have the new 13gen chipsets.
Plus Intel have the DDR4 based motherboard options which is a lifesaver if you check DDR5 pricing and also the launch prices of AMD DDR5 platforms won't be cheap if you check first AM4 motherboard offerings.
But let's just say for argument's sake that you will be able to find DDR5 AM5 motherboards at the same price as Intel's DDR4 motherboards, so I won't take into consideration the motherboard factor.
Right now DDR5 pricing is more than double (4800MHz vs 3200MHz) or double vs 3600MHz.
So we are talking around $140-120 vs $70-60 for the entry level 16GB kits, so another $70-60 difference per 16GB.
4GB for the AMD APU will be committed to graphics for our use case (entry level 1080p gaming) so let's say that those users that will pay $360 for 6700G + DDR5 motherboard price will be OK with 12GB (lol) so I won't take into consideration the 32GB difference...
Whatever level of performance Intel ARC have it won't matter especially for the entry level since I (reasonably) expect the pricing to be competitive with today's solutions (if not with next gen...) but let's not take Intel's potentially better in price/performance Arc solutions and just take RX 6400, and although you can find the RX 6500XT in better than SRP prices in Europe let's just say that RX6400 will be at $159 in Q4 although we would have ARC/Ada Lovelace/ RDNA3 solutions by then pressuring the market.
So for Intel we will have:
13400F $180
16GB DDR4 $70
6400RX $160
Sum $410
and for AMD we will have:
6700G $360
16GB DDR5 $140 (12GB essentially...)
Sum $500
So with $90 difference (or more with an ARC solution), why would anyone just buy a 6000G solution?
If you need entry gaming just use the $90 difference to buy a better VGA, imagine what a $250 ($160+$90) VGA you would buy at Q4 and that's it, it's no contest really.
So no, except NUCs or other niche use cases Rembrandt will not be competitive at all for entry desktop gaming!

 

[TheinsanegamerN]

I mean it makes sense, at some point you have to embrace the new standard. Supporting both would require a mroe complicated IMC, and by the end of this year the price/perf of DDR5 is going to be halfway decent. It's already improved a LOT in the last year.

Not for me though. I picked up a 5900x real cheap on sale, that'll tide me over for the next 10 years.

 

[Denver]

For the 5800X replacement and up I guess it will acceptable more or less due to price segment, but it will face stiff competition from 13600K/13600KS (6P+8E) and below (probably it will be just a little bit slower in 1080p gaming but in multitasking apps like Cinebench it will be no contest with at least +40% more performance, probably +50% vs 5600X replacement)
Many people don't realize it but the real problem will be for the APU 6000G series. I will go as far to say that for people that just want an entry gaming solution and the main reason is entry gaming and they don't buy an APU for the form factor capabilities (NUC like solutions etc) 6000G is DOA.
We had 4700G Renoir with 156mm² die size at $309 OEM so let's just say +$10 for retail - with Intel the difference between tray (1 year warranty) and box with fan and 3 years warranty is less than $10 for a local official distributor in Europe.So let's just say $319 theoretical SRP.
With 5700G Cezanne (180mm²) we went at $359 SRP although the iGPU part was the same (the reason 5700G is faster in some games than 4700G despite the frequency deficit 2GHz vs 2.1GHz is Zen 3 and cache because the testing is done in 720/1080p resolutions with low/medium settings)
So in the same 7nm process AMD increased the SRP more or less in accordance with die size increasment.
With 6700G Rembrandt we have 208mm² in a more expensive 6nm process, so someone would think that AMD would have a $399 SRP or something, right?
Well no, I think they will launch it at the same SRP as 5700G just to save face and still will not be competitive with Intel Q4 offerings.
A i5 13400F will probably be a 6P+4E design and it will offer better performance in TPU 720p/1080p gaming CPU testing AND in multitasking apps like Cinebench vs 6700G, this is certain imo.
Regarding the Rembrandt APU I keep hearing people saying it will be faster than RX 6400 but it will not, even at 2.5GHz (vs 2.32GHz) it will not be faster due to the fact that RX 6400 has double ROPs (which is crucial), infinity cache + 67% more bandwidth (that it doesn't share with Zen3+ cores) and a beefier unslice (Intel sic) for ACE, geometry processor etc.
If you add CPU/Graphics/motherboard/memory costs you would easily see that AMD is not really competitive at all:
13400F vs 6700G will probably be $360-$180, so already we have a $180 difference
I think probably Intel will have also cheaper motherboard options since already we had price drops and we will see even more price drops since by Q4 we will have the new 13gen chipsets.
Plus Intel have the DDR4 based motherboard options which is a lifesaver if you check DDR5 pricing and also the launch prices of AMD DDR5 platforms won't be cheap if you check first AM4 motherboard offerings.
But let's just say for argument's sake that you will be able to find DDR5 AM5 motherboards at the same price as Intel's DDR4 motherboards, so I won't take into consideration the motherboard factor.
Right now DDR5 pricing is more than double (4800MHz vs 3200MHz) or double vs 3600MHz.
So we are talking around $140-120 vs $70-60 for the entry level 16GB kits, so another $70-60 difference per 16GB.
4GB for the AMD APU will be committed to graphics for our use case (entry level 1080p gaming) so let's say that those users that will pay $360 for 6700G + DDR5 motherboard price will be OK with 12GB (lol) so I won't take into consideration the 32GB difference...
Whatever level of performance Intel ARC have it won't matter especially for the entry level since I (reasonably) expect the pricing to be competitive with today's solutions (if not with next gen...) but let's not take Intel's potentially better in price/performance Arc solutions and just take RX 6400, and although you can find the RX 6500XT in better than SRP prices in Europe let's just say that RX6400 will be at $159 in Q4 although we would have ARC/Ada Lovelace/ RDNA3 solutions by then pressuring the market.
So for Intel we will have:
13400F $180
16GB DDR4 $70
6400RX $160
Sum $410
and for AMD we will have:
6700G $360
16GB DDR5 $140 (12GB essentially...)
Sum $500
So with $90 difference (or more with an ARC solution), why would anyone just buy a 6000G solution?
If you need entry gaming just use the $90 difference to buy a better VGA, imagine what a $250 ($160+$90) VGA you would buy at Q4 and that's it, it's no contest really.
So no, except NUCs or other niche use cases Rembrandt will not be competitive at all for entry desktop gaming!

Sorry, but comparing unreleased products, with invented prices and performance doesn't work.

** Just a small correction, 6nm is just an improved version of 7nm, so designs produced in 7nm remain compatible, allowing you to "refresh" (RX xx50) with no additional costs involved. The cost must be the same.

 

[ModEl4]

Sorry, but comparing unreleased products, with invented prices and performance doesn't work.

** Just a small correction, 6nm is just an improved version of 7nm, so designs produced in 7nm remain compatible, allowing you to "refresh" (RX xx50) with no additional costs involved. The cost must be the same.

It works for me, lol
Just kidding, it's just a prediction, we will see if it comes true.
Regarding 6nm vs 7nm the only thing I assumed is that a 6nm wafer price is probably higher than a 7nm since it offers +18% logic density so TSMC should be selling it with a small premium, but even if AMD secured the same wafer price, it doesn't change the argument which was that Rembrandt going from 180mm² to 208mm² is costing AMD higher to produce (which in turn was an insignificant statement in the whole post that had Rembrandt's competitiveness as the main theme). Yes, N6 is design rule and IP model compatible and probably will allow AMD to refresh an existing 7nm design having in the end similar cost (taking account the small general density increase, not on a wafer level) but what are you argue about? and why you assumed I don't know the N6/N7 difference feeling the need to "correct me", especially since it doesn't play fundamentally a role in the original assumptions regarding Rembrandt competitiveness?
Anyway, if you disagree regarding Rembrandt's competitiveness just post your own analysis/forecast, pluralism is good!

 

[Valantar]

They should just have PCIe5 as default.

PCIe 5.0 makes little sense as a chipset connectivity solution. Heck, it doesn't make sense for any consumer application for the coming half decade or so - even storage. Sure, benchmarking SSDs will be faster, but given that PCIe 4.0 SSDs deliver no perceptible increase in performance over PCIe 3.0, there's no reason to expect anything else from PCIe 5.0. And x16 GPUs are nowhere near saturating 4.0.

You could always argue that a 5.0 x4 chipset link will allow for, for example, two 4.0 SSDs at full speed, but that again doesn't make much sense in reality. When real-world applications use 1/3rd or less of your SSD's peak speeds, there's no benefit to be had from this.

It would just drive up motherboard prices unnecessarily, so I'm glad they're skipping it. Heck, I hope they keep B650 PCIe 4.0-only, though sadly that's not likely given competition.

 

[Denver]

It works for me, lol
Just kidding, it's just a prediction, we will see if it comes true.
Regarding 6nm vs 7nm the only thing I assumed is that a 6nm wafer price is probably higher than a 7nm since it offers +18% logic density so TSMC should be selling it with a small premium, but even if AMD secured the same wafer price, it doesn't change the argument which was that Rembrandt going from 180mm² to 208mm² is costing AMD higher to produce (which in turn was an insignificant statement in the whole post that had Rembrandt's competitiveness as the main theme). Yes, N6 is design rule and IP model compatible and probably will allow AMD to refresh an existing 7nm design having in the end similar cost (taking account the small general density increase, not on a wafer level) but what are you argue about? and why you assumed I don't know the N6/N7 difference feeling the need to "correct me", especially since it doesn't play fundamentally a role in the original assumptions regarding Rembrandt competitiveness?
Anyway, if you disagree regarding Rembrandt's competitiveness just post your own analysis/forecast, pluralism is good!

You pointed to 6nm as the "most expensive" process. I just said that it doesn't make sense in my perception.

Also keep in mind that evaluating basic manufacturing costs is more important than evaluating the final price. The first indicates how much margin AMD can drop.

Before TSMC reached the peak of maturation of the 7nm node, the average yield of functional dies for Cezanne was 248 dies/waffer. I'm going to exaggerate and say the cost per wafer on 7nm is $10k, so $40.33 per chip, let's assume everything else costs $20 more, don't you think there's a huge margin left to play with? lol

 

[trsttte]

Nothing here is really news, DDR5 only was already confirmed by AMD and the dual chipset idea was also floated around previously as was the pcie5.0 or 8x 4.0.

PCIe 5.0 makes little sense as a chipset connectivity solution. Heck, it doesn't make sense for any consumer application for the coming half decade or so - even storage. Sure, benchmarking SSDs will be faster, but given that PCIe 4.0 SSDs deliver no perceptible increase in performance over PCIe 3.0, there's no reason to expect anything else from PCIe 5.0. And x16 GPUs are nowhere near saturating 4.0.

You could always argue that a 5.0 x4 chipset link will allow for, for example, two 4.0 SSDs at full speed, but that again doesn't make much sense in reality. When real-world applications use 1/3rd or less of your SSD's peak speeds, there's no benefit to be had from this.

It would just drive up motherboard prices unnecessarily, so I'm glad they're skipping it. Heck, I hope they keep B650 PCIe 4.0-only, though sadly that's not likely given competition.

Yes and no, it all the depends on how much you want to improve the platform overall and go for halo features that no one will use (whoever would will be buying workstation and threadripper anyway).

Like currently we have higher end boards running 2 or even 3 nvme's out of the chipset with x570, if you want to run benchs (lol, totally normal use ), not to mention all the other peripherals, that will obviously struggle.

I do agree B650 could remain 4.0 only as the article sugests, bumping up to 8x seems like a natural upgrade (since intel also upped their "DMI" to 8x), so maybe do the same as with x570/b550 where the lower end keeps the "older" version of pcie.

I do wonder what future line-ups will end up looking like, non-pro threadripper seems dead but (looking at current platforms) the gap between am4 and threadripper pro is huge. Going forward I think the gap would continue to be huge and at the same time non-pro threadripper would make even less sense with a "tr5" (variation on the sp5) that cuts down so much features (looking at how non-pro threadripper currently is)

 

[Valantar]

Nothing here is really news, DDR5 only was already confirmed by AMD and the dual chipset idea was also floated around previously as was the pcie5.0 or 8x 4.0.



Yes and no, it all the depends on how much you want to improve the platform overall and go for halo features that no one will use (whoever would will be buying workstation and threadripper anyway).

Like currently we have higher end boards running 2 or even 3 nvme's out of the chipset with x570, if you want to run benchs (lol, totally normal use ), not to mention all the other peripherals, that will obviously struggle.

I do agree B650 could remain 4.0 only as the article sugests, bumping up to 8x seems like a natural upgrade (since intel also upped their "DMI" to 8x), so maybe do the same as with x570/b550 where the lower end keeps the "older" version of pcie.

I do wonder what future line-ups will end up looking like, non-pro threadripper seems dead but (looking at current platforms) the gap between am4 and threadripper pro is huge. Going forward I think the gap would continue to be huge and at the same time non-pro threadripper would make even less sense with a "tr5" (variation on the sp5) that cuts down so much features (looking at how non-pro threadripper currently is)

16+ core MSDT CPUs have essentially killed consumer HEDT - that used to be up to 8 cores, after all. There are exceptionally few consumer applications that will meaningfully benefit from more than 16 cores, and if you have a task like that for business, well, TR Pro (or renting time on a render farm) is for you.

The same goes for connectivity, really. MGPU is dead, and most PCs have a single NVME drive in them, and might add another before the system is replaced. Enthusiasts can pack a lot more in, but we are in the clear minority. Heck, most of us still run a single AIC. IMO, >99.999% of users would never need more than a single x16 slot and three m.2 slots, whether at 3.0 or 4.0 speeds. An x4 for a capture card, NIC or USB controller could be useful for a few, but beyond that there really isn't much need even in enthusiast circles. And IMO it's perfectly fine for mainstream platforms to target those levels of IO. The high end can pack in more, but that also pushes costs higher.

 

[Max(IT)]

Are we sure DDR5 are at this point ready to be the only option available? They still are very expensive (and not that fast).

 

[Valantar]

Are we sure DDR5 are at this point ready to be the only option available? They still are very expensive (and not that fast).

Given that AM5 won't be here until at least Q3, there's a decent amount of time for prices to normalize. Also, as mentioned before, prices really aren't that bad - they just look bad because DDR4 is absolutely ridiculously cheap these days. It isn't more than a couple of years since DDR4 cost as much as DDR5 does today.

 

[ModEl4]

You pointed to 6nm as the "most expensive" process. I just said that it doesn't make sense in my perception.

Also keep in mind that evaluating basic manufacturing costs is more important than evaluating the final price. The first indicates how much margin AMD can drop.

Before TSMC reached the peak of maturation of the 7nm node, the average yield of functional dies for Cezanne was 248 dies/waffer. I'm going to exaggerate and say the cost per wafer on 7nm is $10k, so $40.33 per chip, let's assume everything else costs $20 more, don't you think there's a huge margin left to play with? lol

Again, N6 has 18% logic scaling, let's say a N7 design is 248 dies/wafer and converting it to 6nm gets you 275 dies/wafer (N7 gives -10% less dies/wafer in this hypothetical scenario) if TSMC sells the N6 wafer 10% more than N7 then you are getting similar costs per chip, I was talking per wafer. I was comparing Cezanne vs Rembrandt so indicative is the per wafer cost, let's say AMD had chosen to convert Cezanne to N6, per chip it would have the same cost but the die size would be smaller (let's say -10% so around 162mm² just for example), so the cost difference between Cezanne and Rembrandt designs isn't just what 180mm² vs 208mm² suggest only, you have to add the scaling factor of each process.
The margin is huge imo, I didn't even implied anything like that, I don't know with whom you're arguing about what.

 

[catulitechup]

 

[Denver]

Again, N6 has 18% logic scaling, let's say a N7 design is 248 dies/wafer and converting it to 6nm gets you 275 dies/wafer (N7 gives -10% less dies/wafer in this hypothetical scenario) if TSMC sells the N6 wafer 10% more than N7 then you are getting similar costs per chip, I was talking per wafer. I was comparing Cezanne vs Rembrandt so indicative is the per wafer cost, let's say AMD had chosen to convert Cezanne to N6, per chip it would have the same cost but the die size would be smaller (let's say -10% so around 162mm² just for example), so the cost difference between Cezanne and Rembrandt designs isn't just what 180mm² vs 208mm² suggest only, you have to add the scaling factor of each process.
The margin is huge imo, I didn't even implied anything like that, I don't know with whom you're arguing about what.

I just added information to take into account for your "predictions".

Also, I would like to point out that the RX 6400 has the same number of ROPs as the RX 680M according to the TPU data. Keep your crystal ball updated

 

[trsttte]

Are we sure DDR5 are at this point ready to be the only option available? They still are very expensive (and not that fast).

We need to move forward eventually, for those who can't afford the possible premium on DDR5 there's always the now cheap zen3 with even cheaper DDR4. A completely new platform is generally not for the budget conscious builder anyway

 

[Why_Me]

Thank you Intel for giving us poors the option RE DDR4 / DDR5.

 

[ModEl4]

I just added information to take into account for your "predictions".

Also, I would like to point out that the RX 6400 has the same number of ROPs as the RX 680M according to the TPU data. Keep your crystal ball updated

According to AMD, 680M has double the render back ends of vega 8 and Vega 8 has 8 RBEs, so TPU dadabase is probably wrong, you can check with them about documentation, I doubt they will have any.

https://www.notebookcheck.net/fileadmin/Notebooks/AMD/Ryzen_6000/AMD_Ryzen_680M_RDNA2.jpg