Question AMD Rembrandt/Zen 3+ APU Speculation and Discussion

[moinmoin]

Well since USB4 is heavily based on TB3, no wonder.
To be confirmed but what I understood is any PC compatible with USB4 is compatible with TB3 AND compatible with USB 3.x and USB 2. USB 1(1.1) though is unkwown (unspecced ?)
That being said lower end devices will be 20Gbps, better oines will be 40Gbps and cables will cost an arm and a leg.

Thunderbolt 4 is essentially USB 4 with all featured enabled and all specs maxed out. So 20Gbps can be USB 4, but only 40Gbps can be Thunderbolt 4 etc.

 

[DisEnchantment]

Wait, is thunderbolt confirmed for Rembrandt?

1022:162e/1022:162f are two PCIe devices providing USB4 functionality and the loaded driver is thunderbolt not xhci.

thunderbolt 274432 0 - Live 0xffffffffc026c000
63:00.5 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:162e] (prog-if 40)
Subsystem: Advanced Micro Devices, Inc. [AMD] Device [1022:162e]
Kernel driver in use: thunderbolt
Kernel modules: thunderbolt

63:00.6 USB controller [0c03]: Advanced Micro Devices, Inc. [AMD] Device [1022:162f] (prog-if 40)
Subsystem: Advanced Micro Devices, Inc. [AMD] Device [1022:162f]
Kernel driver in use: thunderbolt
Kernel modules: thunderbolt

Additionally RMB/YC can do SST/MST DP2.0 Alt mode via USB4 (2x ports)
Device compatibility is a different matter though, but if they stick to standards should be fine.

More details related to supported TB features

[ 0.292787] ACPI: USB4 _OSC: OS supports USB3+ DisplayPort+ PCIe+ XDomain+
[ 0.292789] ACPI: USB4 _OSC: OS controls USB3+ DisplayPort+ PCIe+ XDomain+

 

[zir_blazer]

They are very useful for the embedded SKUs.

Ryzen V3000 (looks RMB based) should be an enormous upgrade over the V2000 and V1000 series (Used in Tesla Model Y for instance).

https://twitter.com/i/web/status/1407044836637790212

In a Linux embedded environment, it should clobber many current high end embedded SoCs like Exynos Auto V9 or Snapdragon SA8xxx series, not to mention anybody can take a Linux distro and support comes out of the box for every peripheral. Zero binary blobs in the HLOS.

The irony is that according to ServeTheHome, which reviewed a few EPYC and Ryzen Embedded Motherboards, they are still rarely used even in those lines because vendors like to use their already developed full network stack. For the most part, the Quad 10G MACs are decoration.

I personally think than AMD would have Conroed the network market had they decided to go with the 10G MACs from the start. An alternative would have been to let them work as 1G MACs with a 1G PHY as part of the base platform instead of the standard PCIe NICs, if it was a cost issue. After all, every Motherboard has at minimum a 1G Ethernet port, and a 1G PHY may have been cheaper than a PCIe NIC...

 

[DisEnchantment]

The irony is that according to ServeTheHome, which reviewed a few EPYC and Ryzen Embedded Motherboards, they are still rarely used even in those lines because vendors like to use their already developed full network stack. For the most part, the Quad 10G MACs are decoration.

I personally think than AMD would have Conroed the network market had they decided to go with the 10G MACs from the start. An alternative would have been to let them work as 1G MACs with a 1G PHY as part of the base platform instead of the standard PCIe NICs, if it was a cost issue. After all, every Motherboard has at minimum a 1G Ethernet port, and a 1G PHY may have been cheaper than a PCIe NIC...

Not having an inbuilt MAC unit on a modern embedded SoC is unacceptable and downright ridiculous.
Almost all modern ARM SoCs (Jacinto, i.MX, Exynos Auto, Snapdragon Auto, Tegra etc) in the embedded market have inbuilt MACs, part numbers may vary in the number of MAC blocks they have but all these series I mentioned have (and I have worked on all of these at some point or the other)
Ryzen embedded is not only targeted for COTS and DIY, they are targeted also for Automotive, Industrial, Automation and the like where OEMs and T1s design their own PCBs around their use cases and not meant to be end user serviced. These folks do not buy "motherboards".
While the kind of MAC feature set may be up for discussion (i.e. supporting 802.1Q, 802.1AS, 802.1BA etc) they are standard feature in embedded domain and the OEM decides which PHY they go with depending on the target area of deployment.
Additionally, In this area regular 1000BaseT PHYs are not common, folks usually go with OABR 1000BaseT1 and sometimes optical, the SoC just need to provide an MII/RGMII interface.

 

[Tuna-Fish]

I personally think than AMD would have Conroed the network market had they decided to go with the 10G MACs from the start.

The problem with that is the licensing costs.

For all ethernet standards before 10GBASE-T, the patent holders agreed on a revenue split, formed a patent pool, and then set the licensing costs per port to maximize total revenue (that is, just low enough that every device gets the newest stuff). For 10G, this broke down, the patents were never formed into a pool, and individual patent holders set their prices to maximize their individual revenue. And since everyone else was charging such high fees that 10G will be rare, you should do that too. And so consumer networking has effectively languished in a pit for 15 years.

At this point, the hardware cost of making a 10G ethernet port is trivial if you are already using high-end logic. It costs AMD effectively nothing to stuff those ports on the die. The problem is that if they want to sell a device that does 10GBASE-T, they have to cough up tens of dollars per port to half a dozen different companies. They could only "conroe" networking if they were willing to add something like $70 to the cost of every CPU.

This mess is only going to actually get fixed on 18th of July 2023, when the last patent that covers 10GBASE-T finally expires. From that day, pretty much everything is going to come with 10G as standard. Between then and now, a few of the earlier patents will expire first, and each of them will reduce the cost of 10G ethernet incrementally.

 

[Thibsie]

Ouch, VERY intetesting read !
Explains a lot.

 

[moinmoin]

The problem with that is the licensing costs.

For all ethernet standards before 10GBASE-T, the patent holders agreed on a revenue split, formed a patent pool, and then set the licensing costs per port to maximize total revenue (that is, just low enough that every device gets the newest stuff). For 10G, this broke down, the patents were never formed into a pool, and individual patent holders set their prices to maximize their individual revenue. And since everyone else was charging such high fees that 10G will be rare, you should do that too. And so consumer networking has effectively languished in a pit for 15 years.

At this point, the hardware cost of making a 10G ethernet port is trivial if you are already using high-end logic. It costs AMD effectively nothing to stuff those ports on the die. The problem is that if they want to sell a device that does 10GBASE-T, they have to cough up tens of dollars per port to half a dozen different companies. They could only "conroe" networking if they were willing to add something like $70 to the cost of every CPU.

This mess is only going to actually get fixed on 18th of July 2023, when the last patent that covers 10GBASE-T finally expires. From that day, pretty much everything is going to come with 10G as standard. Between then and now, a few of the earlier patents will expire first, and each of them will reduce the cost of 10G ethernet incrementally.

Thanks for the read! So in these cases the tech is essentially only allowed to become mainstream after all the involved patent lapse? Perfect argument for shortening the term of patent to avoid too much tech becoming mainstream only after a delay of 20 years.

 

[Thibsie]

Thanks for the read! So in these cases the tech is essentially only allowed to become mainstream after all the involved patent lapse? Perfect argument for shortening the term of patent to avoid too much tech becoming mainstream only after a delay of 20 years.

Agreed. If patent (in this case at least) prevent technology adoption, then there's a problem.

 

[mikk]

https://twitter.com/i/web/status/1473703286750720001

Any thoughts on this?

Btw if Rembrandt supports DDR5 only I don't see how this can be a volume product before mid 2022 if not later. I guess Barcelo will be AMDs volume mobile lineup for most of 2022.

 

[Mopetar]

https://twitter.com/i/web/status/1473703286750720001

Any thoughts on this?

Btw if Rembrandt supports DDR5 only I don't see how this can be a volume product before mid 2022 if not later. I guess Barcelo will be AMDs volume mobile lineup for most of 2022.

Given all of the shortages and associated price gouging maybe Rembrandt won't be a volume part until the middle of next year. Normally it isn't beneficial to push a product out too soon because it creates complaints of shortages.

If that's going to happen anyway then why not launch as soon as you can and capitalize on it? Maybe you won't get a lot of market share from it, but if it gets some mind share at no cost then you may as well go for it.

 

[izaic3]

https://twitter.com/i/web/status/1473703286750720001

Any thoughts on this?

Btw if Rembrandt supports DDR5 only I don't see how this can be a volume product before mid 2022 if not later. I guess Barcelo will be AMDs volume mobile lineup for most of 2022.

Yeah, but remember that we're going from Vega to rdna2 not just rdna1 to rdna2. The DDR5 constraint is kind of concerning.

 

[Abwx]

Yeah, but remember that we're going from Vega to rdna2 not just rdna1 to rdna2. The DDR5 constraint is kind of concerning.

And also 50% more SPs, overall it should be something like 2x at isopower.

 

[ahimsa42]

it won't be long until we know for sure. hopefully there will be some real world gaming benchmarks at the Jan 4th announcement. worse case, RMB won't be worth upgrading to for Renior/Cezzane owners and they will skip the 1st gen DDR5 and instead wait for Zen 4 Phoenix in late 2022/early 2023.

 

[zir_blazer]

The problem with that is the licensing costs.

For all ethernet standards before 10GBASE-T, the patent holders agreed on a revenue split, formed a patent pool, and then set the licensing costs per port to maximize total revenue (that is, just low enough that every device gets the newest stuff). For 10G, this broke down, the patents were never formed into a pool, and individual patent holders set their prices to maximize their individual revenue. And since everyone else was charging such high fees that 10G will be rare, you should do that too. And so consumer networking has effectively languished in a pit for 15 years.

Do note that AMD MACs are 10GBASE-KR, and most likely they also do 1000BASE-KX. They don't do 10GBASE-T directly, you need some 10GBASE-KR -> 10GBASE-T PHY on the Motherboard itself, or go to a SFP+ Port then use a 10GBASE-T Transceiver (Or fiber, the older 1000BASE-T, or whatever else you want), so I don't know how your knowledge about 10GBASE-T patent licensing applies to the -KR version as it only makes sense that you pay the licensing on the -T products and not on the AMD SoC itself.

 

[Shivansps]

https://twitter.com/i/web/status/1473703286750720001

Any thoughts on this?

Btw if Rembrandt supports DDR5 only I don't see how this can be a volume product before mid 2022 if not later. I guess Barcelo will be AMDs volume mobile lineup for most of 2022.

I dont understand the point of that slide. Comparing RMB RDNA2 to RDNA1 makes no sence due to the fact that RDNA1 APUs do not exist.

The only thing that can bring down RMB performance is TDP and cooling, since RMB will be able to provide CZN perf at a lower TDP, it may end up doing just that. We are not going to see the true performance until it comes out in AM5 form.

 

[uzzi38]

https://twitter.com/i/web/status/1473703286750720001

Any thoughts on this?

Btw if Rembrandt supports DDR5 only I don't see how this can be a volume product before mid 2022 if not later. I guess Barcelo will be AMDs volume mobile lineup for most of 2022.

Probably just referring to overall score vs gfx scores. Doesn't matter given how small the difference is between the two.

As for the slide, I don't get the point of it. We'll have useful benchmarks soon enough.

As for volume, LPDDR5 is an option and has been in mass production for a year and a half now. DDR5 volume is slim, but you should expect that AMD's presence in thin-and-light designs is going to be more prevalent than before.

Cezanne was not a good competitor for Tiger Lake-U in thin and light designs. Rembrandt is a very good competitor for Alder Lake there.

Barcelo will play a big impact for the low and partially the mid range market - U-series laptops that previously shipped with DDR4-3200.

 

[mikk]

I dont understand the point of that slide. Comparing RMB RDNA2 to RDNA1 makes no sence due to the fact that RDNA1 APUs do not exist.

The only thing that can bring down RMB performance is TDP and cooling, since RMB will be able to provide CZN perf at a lower TDP, it may end up doing just that. We are not going to see the true performance until it comes out in AM5 form.

Maybe he refers to the Infinity Cache. A big chunk of IPC gains comes from this which is not available for Rembrandt. Only real benchmarks will tell us.

As for volume, LPDDR5 is an option and has been in mass production for a year and a half now. DDR5 volume is slim, but you should expect that AMD's presence in thin-and-light designs is going to be more prevalent than before.

Considering that even LPDDR4x is very low volume on AMD there is a big question mark about LPDDR5 as well. And no doubt they will charge a premium price over LPDDR4x. For the cheaper ones DDR4 and partially LPDDR4x is the way to go. Would be nice to have many LPDDR5 devices though, the extra bandwidth should help a lot for iGPUs.

 

[uzzi38]

Maybe he refers to the Infinity Cache. A big chunk of IPC gains comes from this which is not available for Rembrandt. Only real benchmarks will tell us.



Considering that even LPDDR4x is very low volume on AMD there is a big question mark about LPDDR5 as well. And no doubt they will charge a premium price over LPDDR4x. For the cheaper ones DDR4 and partially LPDDR4x is the way to go. Would be nice to have many LPDDR5 devices though, the extra bandwidth should help a lot for iGPUs.

Infinity Cache is not but Rembrandt comes with 4x the L2 cache of existing Vega iGPUs. Still only 4MB, but for 720p and entry 1080p that'll make a big difference. Not to mention RDNA in general introduced a whole new level of cache over Vega.

As for LPDDR4X, there's still a decent number of designs using the memory, but ultimately it's a mid-range/high-end laptop feature. Rembrandt fixes many of the last issues for OEMs contemplating using AMD such devices - I don't doubt we'll see more high end and premium thin and light AMD designs. Before it was impossible to do so as battery life was still not good enough and missing platform features.

As for DDR5 vs LPDDR5, RMB supports DDR5-5200 and LPDDR5-5500. Bandwidth differences are minimal at best.

 

[mikk]

As for LPDDR4X, there's still a decent number of designs using the memory, but ultimately it's a mid-range/high-end laptop feature. Rembrandt fixes many of the last issues for OEMs contemplating using AMD such devices - I don't doubt we'll see more high end and premium thin and light AMD designs. Before it was impossible to do so as battery life was still not good enough and missing platform features.

Cezanne-U+Lucienne-U: 88% DDR4 and 12% LPDDR4. This is a very low volume in my books. As for TGL-U the percentage is 50% DDR4 and 50% LPDDR4x.

As for DDR5 vs LPDDR5, RMB supports DDR5-5200 and LPDDR5-5500. Bandwidth differences are minimal at best.

At best, in real probably DDR5-4800 vs LPDDR5-5500. The only Rembrandt leak so far was on DDR5-4800. And because Intel only supports DDR5-4800 there won't be much OEM SO-DIMM beyond 4800 for the time being.

 

[uzzi38]

Cezanne-U+Lucienne-U: 88% DDR4 and 12% LPDDR4. This is a very low volume in my books. As for TGL-U the percentage is 50% DDR4 and 50% LPDDR4x.

Where do those numbers come from?

At best, in real probably DDR5-4800 vs LPDDR5-5500. The only Rembrandt leak so far was on DDR5-4800. And because Intel only supports DDR5-4800 there won't be much OEM SO-DIMM beyond 4800 for the time being.

https://twitter.com/i/web/status/1308484619097042945

5200.

The last point is just stupid. I've seen the official validation list from Intel specifically, you'll be seeing LPDDR5-6400 ship primarily in Intel laptops supporting the technology, but it will only be ran at LPDDR5-5200.

The memory cartel isn't going to just not up-bin and then up-sell memory whatever memory they can make.

 

[leoneazzurro]

LPDDR5 is used in various devices already, some with quite high numbers: see Apple latest and mighties, for an instance (even if in that case it's integrated on package).

 

[mikk]

Where do those numbers come from?

The numbers are from geizhals:

Notebooks mit CPU-Suffix AMD: U, Codename AMD: Cezanne/Lucienne Preisvergleich Geizhals Deutschland

Preisvergleich und Bewertungen für Notebooks mit CPU-Suffix AMD: U, Codename AMD: Cezanne/Lucienne

geizhals.de

Notebooks mit Codename Intel: Tiger Lake-UP3/Tiger Lake-UP4 Preisvergleich Geizhals Deutschland

Preisvergleich und Bewertungen für Notebooks mit Codename Intel: Tiger Lake-UP3/Tiger Lake-UP4

geizhals.de

AMD: DDR4 341 LPDDR4 41
Intel: DDR4 958 LPDDR4 437

5200.

??????

I haven't denied Rembrand doesn't support DDR5-5200. But that those not mean they will ship with DDR5-5200. Once again, the only leaked Rembrandt so far was on DDR5-4800: https://videocardz.com/newz/amd-rembrandt-8-core-16-thread-zen3-apu-for-fp7-socket-has-been-spotted

The last point is just stupid. I've seen the official validation list from Intel specifically, you'll be seeing LPDDR5-6400 ship primarily in Intel laptops supporting the technology, but it will only be ran at LPDDR5-5200.

The memory cartel isn't going to just not up-bin and then up-sell memory whatever memory they can make.

This isn't stupid, the thing is you simply cannot compare LPDDR5 with DDR5 since LPDDR5 has a minimum speed of at least 5500, there is no module with less than 5500, they are basically all downclocked. This comparison is completely invalid. Whatever LPDDR5 module OEMs will use, they can easily support the maximum of Rembrandt. This is not the case with DDR5 because DDR5-4800 is the starting point which most modules are based on. All of the validated DDR5 SODIMM modules are 4800 based in Intels validation list.

 

[xilli_fiberbit]

The problem with that is the licensing costs.

For all ethernet standards before 10GBASE-T, the patent holders agreed on a revenue split, formed a patent pool, and then set the licensing costs per port to maximize total revenue (that is, just low enough that every device gets the newest stuff). For 10G, this broke down, the patents were never formed into a pool, and individual patent holders set their prices to maximize their individual revenue. And since everyone else was charging such high fees that 10G will be rare, you should do that too. And so consumer networking has effectively languished in a pit for 15 years.

At this point, the hardware cost of making a 10G ethernet port is trivial if you are already using high-end logic. It costs AMD effectively nothing to stuff those ports on the die. The problem is that if they want to sell a device that does 10GBASE-T, they have to cough up tens of dollars per port to half a dozen different companies. They could only "conroe" networking if they were willing to add something like $70 to the cost of every CPU.

This mess is only going to actually get fixed on 18th of July 2023, when the last patent that covers 10GBASE-T finally expires. From that day, pretty much everything is going to come with 10G as standard. Between then and now, a few of the earlier patents will expire first, and each of them will reduce the cost of 10G ethernet incrementally.

It's good to see a patent expiring in 2023, i think it was less used because it's more interesting to use SFP+. An ethernet controller in 2,5 Gb/s/5Gb/s available in Ryzen desktop could be good to

 

[uzzi38]

The numbers are from geizhals:

Notebooks mit CPU-Suffix AMD: U, Codename AMD: Cezanne/Lucienne Preisvergleich Geizhals Deutschland

Preisvergleich und Bewertungen für Notebooks mit CPU-Suffix AMD: U, Codename AMD: Cezanne/Lucienne

geizhals.de

Notebooks mit Codename Intel: Tiger Lake-UP3/Tiger Lake-UP4 Preisvergleich Geizhals Deutschland

Preisvergleich und Bewertungen für Notebooks mit Codename Intel: Tiger Lake-UP3/Tiger Lake-UP4

geizhals.de

AMD: DDR4 341 LPDDR4 41
Intel: DDR4 958 LPDDR4 437

So the 50-50 split for Tiger Lake was BS. Thanks for confirming that.

That's the kind of split I'd expect with Tiger Lake getting vastly more premium and high end notebook designs. Again, I believe we'll see the gap shrink on AMD's side too.

??????

I haven't denied Rembrand doesn't support DDR5-5200. But that those not mean they will ship with DDR5-5200. Once again, the only leaked Rembrandt so far was on DDR5-4800: https://videocardz.com/newz/amd-rembrandt-8-core-16-thread-zen3-apu-for-fp7-socket-has-been-spotted

That's not a production system lol, that's just a test bench. Unless you're expecting Rembrandt to ship in a budget device that only runs single channel memory, which we both know isn't it's purpose.

It's very normal for OEMs to throw whatever memory they have lying around in such systems.

This isn't stupid, the thing is you simply cannot compare LPDDR5 with DDR5 since LPDDR5 has a minimum speed of at least 5500, there is no module with less than 5500, they are basically all downclocked. This comparison is completely invalid. Whatever LPDDR5 module OEMs will use, they can easily support the maximum of Rembrandt. This is not the case with DDR5 because DDR5-4800 is the starting point which most modules are based on. All of the validated DDR5 SODIMM modules are 4800 based in Intels validation list.

I'm sorry, did you not read the bit I wrote about LPDDR5-6400?

What you said would make sense if the list stated LPDDR5-5500 modules at some point. It didn't. They were all 6400MT/s.

Again, memory cartels aren't going to miss the chance to upsell, and shipping higher speed SODIMMs lets them do exactly that, especially in this time where DDR5 costs are so high.

By the way, why does this even matter anyway? The difference in memory bandwidth is still very low either way - smaller than the difference between DDR4 and LPDDR4X by a decent margin. Such a small one that I wouldn't be surprised if data fabrics end up being the bottleneck anyway for thin and light laptops when they have to be clocked low to maximise battery life and best use of available thermal headroom.

 

[mikk]

So the 50-50 split for Tiger Lake was BS. Thanks for confirming that.

45.616%


Nothing about the 12% percentage, this was my main point. Is this a good number in your book? AMD has a very low LPDDR4 volume, there is a big question mark about LPDDR5.

That's not a production system lol, that's just a test bench. Unless you're expecting Rembrandt to ship in a budget device that only runs single channel memory, which we both know isn't it's purpose.

Testing hardware they won't use in the shipping product, makes completely sense. You cannot know if this is just for testing, this a a random guess from your side.

It's very normal for OEMs to throw whatever memory they have lying around in such systems.

Exactly! When no SODIMM above exist they use what they have-->4800 - regardless of the supported RAM. You got it. We have seen this in the past a lot.

What you said would make sense if the list stated LPDDR5-5500 modules at some point. It didn't. They were all 6400MT/s.

This is exactly my point. All existing modules are basically rated 6400, they are just downclocked. You cannot compare DDR5 with LPDDR5.

Again, memory cartels aren't going to miss the chance to upsell, and shipping higher speed SODIMMs lets them do exactly that, especially in this time where DDR5 costs are so high.

RAM OC is not a thing for mobile and AMD is a small customer in the grand scheme. For the time being there is very little reason for a big shift above 4800 in the SODIMM market for the memory manufacturer and no big reason for notebook OEMs to buy 5200 SODIMM just for Rembrandt either.

By the way, why does this even matter anyway? The difference in memory bandwidth is still very low either way - smaller than the difference between DDR4 and LPDDR4X by a decent margin.

I'm not so sure about it. 2x8GB DDR5 DIMM is a lot slower than 2x16GB, not sure if it's the same for SODIMM. If it's the same the difference is quite big for devices running with 2x8GB DDR5.